diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/config.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d29f6f184d56bed953898eb1a3cdbaa0dbe24e53
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/config.json
@@ -0,0 +1,29 @@
+{
+    "cli_args": {
+        "unet": false,
+        "seed": 42,
+        "optimizer_mode": 10,
+        "model_parameterization": "qkvo",
+        "per_group_k": 100,
+        "muon_lr": 0.002,
+        "adam_lr": 0.002,
+        "base_dir": "logs_qa_muon/diff_modes",
+        "sgd_lr": 0.01,
+        "m_val": 15,
+        "qa_jsonl_path": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl"
+    },
+    "hyperparameters": {
+        "train_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin",
+        "val_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin",
+        "val_tokens": 491520,
+        "train_seq_len": 3072,
+        "val_seq_len": 16384,
+        "num_iterations": 10000,
+        "cooldown_frac": 0.8,
+        "vocab_size": 50257,
+        "val_loss_every": 500,
+        "save_checkpoint": false
+    },
+    "run_uuid_for_log": "9700d650-3889-4d26-873a-d01d4e4c55f3",
+    "script_code_logged_at_start": true
+}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/fixed_eval_indices.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/fixed_eval_indices.json
new file mode 100644
index 0000000000000000000000000000000000000000..a823775225c5e592eb10700e5e0319b0491b1eb6
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/fixed_eval_indices.json
@@ -0,0 +1 @@
+{"1": [1238956, 182074, 1437575, 1061037, 383150, 1176376, 926, 823011, 832520, 1266421, 512738, 144357, 848076, 890204, 213997, 95146, 261767, 467731, 832231, 217985, 913168, 107253, 1361828, 61314, 1230420, 1133619, 146690, 429587, 419151, 58695, 1579770, 503799, 1421284, 882534, 1022637, 785343, 1154604, 67783, 1325109, 243941, 1213240, 438111, 460295, 269373, 538055, 1347006, 71775, 255496, 299906, 1227973, 815402, 190082, 1304077, 1023347, 613801, 983830, 1284420, 389321, 1625224, 717538, 1172273, 992184, 1181312, 1014039, 885952, 1538489, 158933, 1667270, 1250445, 958097, 1458224, 1306495, 62945, 733843, 1360200, 540493, 762461, 501460, 1208142, 1180559, 1333588, 690481, 355756, 618511, 733586, 650301, 799437, 165533, 1238977, 323078, 1485080, 609610, 1212241, 606952, 1253407, 1420922, 327112, 701, 777907, 1626516], "0": [1390189, 1220977, 1312259, 1201125, 1235379, 1272843, 344142, 1119560, 856330, 766456, 1145928, 550624, 684681, 1309879, 54620, 1126124, 1278077, 910610, 1532044, 73645, 460165, 462723, 457959, 476877, 320873, 1408316, 989156, 1276325, 338260, 577743, 1136972, 25115, 211215, 1296818, 968705, 436781, 318224, 385325, 1516441, 533464, 1628693, 703399, 579670, 1518978, 305154, 1625960, 1400284, 713841, 1652150, 811009, 336920, 881015, 512030, 1347531, 1005706, 412167, 173312, 1136816, 343017, 537797, 1489267, 935475, 703854, 1570584, 1504269, 1458356, 1310700, 1242505, 509976, 22096, 788120, 1616850, 1591198, 1151224, 797820, 823995, 1153364, 1060220, 727674, 580729, 148912, 334290, 686098, 1633151, 1130523, 1012440, 1374480, 953410, 1381730, 1376118, 932096, 739115, 739014, 263875, 1400308, 556177, 1341771, 348626, 575350, 254846], "10": [748592, 515936, 1645691, 329828, 1104748, 1453439, 406461, 1557507, 1146644, 1398421, 163667, 228310, 407739, 585218, 785725, 1371077, 134834, 15431, 1303784, 273099, 1530332, 1320219, 1329247, 753603, 664548, 501873, 1512467, 583734, 1355643, 600440, 428998, 1426876, 121341, 1456507, 892281, 1242339, 1304508, 968336, 1593981, 634471, 1253799, 593156, 1486078, 712425, 441285, 618217, 1024661, 1395408, 1423117, 1463918, 121752, 800727, 52402, 1185143, 1651644, 1583572, 1302889, 1040220, 1234122, 333457, 1552652, 1567291, 1405501, 222708, 1469496, 875611, 894032, 783951, 934281, 109533, 198744, 643064, 197926, 224405, 1169998, 956511, 1096712, 1213224, 828780, 204659, 564225, 257079, 754287, 1598723, 821179, 1156648, 60674, 290993, 1215838, 1556010, 1025931, 687671, 252775, 771448, 574110, 578615, 279536, 1503737, 1089787, 2338], "14": [1602125, 1388977, 709154, 666138, 1446614, 572103, 99632, 25794, 1487658, 15077, 1600205, 682155, 679850, 1462049, 970283, 81336, 378196, 1627374, 832926, 1217832, 740140, 1336941, 793290, 1441127, 297494, 303372, 339524, 1363322, 525184, 593339, 61055, 1468219, 1110310, 1061855, 1046762, 410655, 449257, 530144, 1362373, 241552, 926068, 1618007, 629406, 327589, 263223, 1525290, 621372, 606355, 534601, 126046, 934024, 1181343, 1234759, 295773, 838035, 1447427, 1395837, 1568116, 486923, 465204, 98192, 1145989, 271529, 720487, 1091877, 653313, 872526, 13890, 111971, 671086, 301528, 50698, 455373, 786098, 329199, 599178, 644737, 1404606, 33829, 279256, 161938, 309352, 1537653, 21723, 309413, 478184, 340938, 1080124, 1375458, 1301528, 331423, 212589, 1220931, 112707, 454022, 1466562, 1238127, 341561, 1497264, 247245], "3": [143966, 715765, 455823, 1414629, 814361, 278667, 700211, 1286614, 1601013, 923359, 1105960, 549680, 371512, 732263, 1236795, 432942, 1379421, 488467, 679369, 1200930, 479010, 683138, 700149, 1422714, 348019, 1263383, 1044275, 820031, 969830, 75153, 1410929, 394116, 1486865, 299414, 264214, 1416493, 428688, 107076, 753671, 674417, 1456072, 307080, 1259105, 1563344, 1198484, 456551, 1061513, 1128111, 922541, 1040277, 1360208, 1411429, 259514, 883801, 716781, 1004987, 285196, 266162, 781115, 385241, 252691, 1398963, 1440582, 1300653, 756663, 231617, 1319447, 42576, 1434337, 20658, 581284, 178420, 1626533, 694079, 1531414, 1631177, 1478512, 647215, 706568, 303323, 10540, 1010611, 1150313, 159016, 1614581, 1257722, 470491, 1333309, 94699, 756222, 1456446, 296132, 1186414, 1591257, 1155050, 789439, 1472373, 1403759, 1451276, 745193], "4": [1227524, 563788, 436964, 1380176, 1648232, 267090, 273744, 1031517, 1580586, 416846, 1309866, 114170, 1649739, 1161604, 690741, 1432068, 224570, 322835, 1559902, 479718, 709757, 410142, 1668701, 1143278, 968258, 406289, 884233, 1080971, 269480, 631196, 421463, 374502, 430292, 1309826, 1422997, 1439354, 131354, 975504, 84660, 1590390, 288064, 1419944, 695652, 320713, 260376, 222416, 937739, 1104508, 1323311, 962301, 1060820, 1117857, 280792, 692359, 656651, 358878, 828406, 598701, 127434, 1511259, 1332605, 1178325, 1281187, 203589, 792701, 684176, 543034, 980077, 982509, 640126, 1252418, 1529627, 519645, 359068, 992779, 550843, 1493488, 447632, 1512599, 361760, 626672, 89042, 604248, 574827, 87735, 1432602, 473061, 668609, 925422, 311417, 71495, 682067, 221270, 600053, 673169, 1408933, 719302, 788676, 998101, 243902], "15": [1444113, 1102795, 1149896, 1181364, 349047, 1076908, 816006, 945310, 197585, 276574, 946546, 1505399, 740223, 1154813, 1312891, 1261018, 249986, 1504592, 647427, 915615, 1245136, 1288301, 802239, 410533, 805664, 1013403, 498363, 777073, 1508176, 18111, 1147340, 880289, 215773, 924760, 1301970, 318135, 1524489, 103475, 872748, 1568512, 722849, 1364157, 838956, 1451408, 970208, 1350756, 1498244, 980461, 1098677, 1141812, 463859, 95672, 1220365, 554451, 524014, 1161048, 17478, 113548, 1027741, 1657643, 1462809, 1279104, 85588, 241936, 510707, 151817, 1191227, 574942, 735662, 1046248, 1056729, 96279, 781378, 1042711, 1602339, 1503514, 299087, 898968, 553537, 293167, 897565, 572125, 425215, 916907, 1210531, 1013275, 201532, 883447, 1641610, 1185682, 425791, 831133, 1291227, 1099064, 934983, 588125, 1284591, 1163873, 524656, 1230762], "13": [1122510, 535854, 845319, 585144, 884184, 441130, 934654, 1046629, 1512962, 1190830, 1462685, 90530, 894615, 1527783, 415323, 90852, 1556905, 1002406, 366356, 1386022, 846952, 928360, 3138, 406205, 78588, 1208032, 1210206, 257177, 137047, 711380, 693628, 105893, 392053, 1301525, 677934, 1549630, 1134520, 447564, 1194061, 672969, 740755, 1045536, 633455, 1336343, 613147, 1161931, 234095, 1246643, 643373, 834221, 1351442, 1502069, 1401555, 759927, 1320337, 666063, 291222, 956633, 400583, 963313, 683474, 534882, 1101312, 1519551, 1513194, 752124, 1028953, 1383362, 169732, 785294, 160568, 1053580, 774534, 372048, 292049, 791470, 1318537, 141966, 510573, 1440830, 930544, 1352026, 463299, 1245306, 159755, 1566413, 316835, 1135200, 266962, 112892, 253603, 1328380, 703630, 21798, 268463, 7112, 1232576, 374978, 240715, 1299198], "2": [463910, 845243, 985993, 402000, 962033, 473711, 108588, 1074932, 539045, 1162793, 1086008, 778638, 751876, 53650, 465364, 803834, 820717, 9490, 473708, 26912, 970677, 1441139, 231484, 749139, 286603, 1346834, 1641988, 699767, 816768, 1220504, 978431, 1331006, 611334, 829473, 271532, 566760, 332292, 1326050, 887731, 597530, 940973, 491159, 1150471, 122010, 225465, 382413, 1207486, 1568920, 992793, 896409, 1479386, 1539347, 759024, 293232, 1345256, 1282360, 956014, 371580, 1202087, 1638038, 892501, 38073, 1651323, 1103388, 1315239, 79005, 1045045, 1230831, 1206253, 1376976, 140037, 320204, 274836, 663525, 397448, 666367, 1501713, 570453, 836424, 694423, 157979, 279074, 1235605, 536170, 301830, 778556, 1069897, 181207, 1445408, 1082031, 969552, 1083538, 1395871, 168187, 5476, 514815, 194418, 1038889, 1288748, 1535767], "8": [258910, 518286, 80108, 811775, 937854, 1623550, 1592007, 661848, 1484674, 300, 104226, 1075215, 1190104, 8885, 1062956, 88002, 1020623, 189243, 1265153, 1472857, 301412, 21371, 441401, 1078954, 1648642, 894192, 1083738, 1182805, 1347655, 316568, 1063971, 255177, 186613, 550617, 102173, 563694, 465966, 965224, 372481, 929520, 1423441, 6575, 862309, 144894, 891435, 123628, 653301, 1378790, 1247156, 1129087, 1381991, 410876, 391895, 202136, 868149, 233421, 564464, 1174972, 1385073, 238326, 483488, 507729, 1329666, 955266, 479623, 902401, 1343890, 827390, 371073, 1480060, 1241650, 1670957, 644144, 1371709, 1324243, 670126, 23715, 1566067, 70858, 522312, 558100, 301714, 359605, 1401913, 1060057, 1027855, 987700, 720084, 324937, 562646, 390910, 1398541, 1097298, 1626769, 1461630, 1430950, 650075, 1025475, 274735, 650932], "7": [181761, 629797, 231148, 365747, 1229714, 500408, 198146, 500703, 1250728, 1578820, 753672, 84476, 1379367, 355424, 1591855, 1400949, 1194404, 912060, 884197, 320928, 1096692, 714201, 331003, 1345763, 1620998, 457231, 1542074, 1108477, 1602113, 641593, 542514, 159345, 360278, 35579, 506, 470465, 1606024, 1349783, 731258, 382125, 883800, 1652746, 382071, 99241, 20216, 1398849, 245155, 1601928, 1544592, 1632383, 1597152, 1266758, 1032571, 88287, 521379, 1423306, 579207, 1152743, 271926, 418168, 8827, 1358019, 429579, 380995, 1040420, 1067138, 283335, 1150469, 280745, 1184794, 468245, 948227, 357499, 527525, 1067827, 816862, 360563, 433490, 258517, 581805, 1089641, 1537081, 473494, 1480784, 60701, 686908, 614323, 143069, 738219, 1234467, 927103, 64088, 658420, 1353941, 1517049, 753334, 198407, 434613, 850731, 1211609], "5": [1250448, 711876, 285881, 342500, 987883, 195047, 810508, 356140, 999602, 284542, 730936, 736191, 661291, 11822, 551928, 617679, 384856, 414906, 238071, 473605, 868498, 1269087, 558407, 245083, 810169, 1456559, 1657433, 70018, 865238, 416303, 1389734, 1561961, 534163, 1098802, 1494214, 1285484, 599936, 296819, 696611, 307176, 29697, 731071, 257656, 1465441, 1668863, 166995, 1668377, 1457645, 1400335, 1525120, 33323, 437285, 459069, 620176, 231704, 846424, 1391144, 364636, 63552, 1093702, 67741, 285428, 454922, 719128, 1393391, 87651, 1418561, 1312176, 652985, 1446762, 71280, 1594253, 89848, 628629, 1117344, 882598, 1014283, 1601256, 739004, 508965, 1273914, 44245, 1267971, 1430470, 611385, 930900, 1439585, 1545682, 1300966, 1331569, 364367, 27900, 107616, 1579070, 94723, 818780, 1183546, 384575, 142627, 431847], "11": [12772, 1091320, 1136723, 560663, 715094, 1097800, 1016547, 1299126, 961155, 725788, 999768, 1277062, 650665, 1323964, 1024796, 1255124, 215111, 268904, 442949, 358645, 621958, 1666210, 1154349, 375517, 614870, 1166618, 1418680, 795635, 184985, 129300, 129366, 1343289, 1151524, 193531, 634856, 1618906, 43721, 1601844, 591772, 1322095, 1522284, 431856, 574972, 1544663, 1389257, 1113893, 1515857, 614345, 809163, 132406, 95327, 997226, 423104, 910705, 261584, 765030, 1216397, 439297, 1565563, 1595785, 1622521, 975588, 1314965, 1205327, 1501612, 1642358, 656493, 1369509, 195325, 1554913, 1661675, 790468, 511387, 1227059, 1219143, 1118001, 788451, 697651, 1575073, 1345104, 167862, 1409843, 138942, 984075, 1382364, 1299538, 214601, 1541475, 459076, 353219, 1310728, 757670, 904351, 334019, 1003259, 670104, 10386, 13556, 1097931, 1526407], "6": [295849, 644897, 1207883, 1648549, 1478750, 1486750, 1156499, 621047, 1387936, 662848, 1362507, 891859, 1188350, 1298200, 1044010, 767288, 1191230, 683760, 533545, 498960, 1421819, 366846, 54389, 1382045, 1002942, 1285004, 457209, 1074763, 823805, 1508295, 719970, 837720, 19881, 1189285, 1226048, 1078656, 1433699, 14805, 254682, 614235, 805087, 900964, 343832, 561709, 243288, 465867, 497085, 1153456, 730115, 42050, 878511, 1382508, 667820, 712856, 715963, 1499782, 1516002, 444399, 726168, 1204910, 591639, 986194, 107160, 1596752, 1239812, 591854, 536489, 1347905, 1597670, 122237, 620494, 354940, 1561646, 1352115, 74785, 163277, 590349, 1429584, 1440950, 848184, 1580723, 1631128, 979479, 758861, 84655, 1079672, 673692, 1315324, 167985, 1353485, 21815, 1175739, 1360408, 1062086, 1369999, 733863, 1371964, 61279, 1152056, 220910], "9": [780278, 1156801, 827882, 1095960, 874763, 1330006, 979412, 1154745, 1399441, 515095, 1505401, 1454104, 1563117, 1316054, 57098, 43367, 1175749, 1257252, 865038, 624410, 186727, 907758, 126168, 1353431, 167806, 1021404, 1585615, 846939, 552247, 1400471, 68525, 190495, 1556857, 181528, 47677, 136468, 189446, 269049, 262467, 503841, 1367168, 382007, 486946, 621235, 622683, 469399, 1537192, 707594, 1003091, 56021, 1354302, 799290, 1618221, 696794, 836658, 1591562, 251096, 1321886, 412122, 1352217, 946767, 1267324, 213504, 747174, 1403653, 1589967, 849743, 549269, 663800, 1315189, 451416, 1330881, 1352197, 1517840, 827169, 1228725, 1097725, 955944, 1375509, 1590213, 266039, 1558465, 1498493, 604161, 608723, 1547013, 1484549, 1530909, 1325336, 1584866, 630737, 1023780, 227549, 961233, 987591, 159267, 608424, 636687, 745397, 460687], "12": [1586709, 357745, 1274743, 202991, 1250469, 837216, 757864, 185235, 493474, 621351, 22706, 1096491, 186138, 1482718, 534525, 1157131, 1350934, 1167912, 1366603, 756644, 311945, 72417, 706444, 1464146, 1236814, 74555, 983713, 1182807, 1082263, 804850, 1028788, 1405036, 932746, 742978, 1589076, 723766, 898380, 906214, 913629, 33386, 256107, 1245317, 360574, 836359, 565572, 400905, 1269191, 1042959, 886332, 1415571, 68870, 470376, 1531, 1333957, 834102, 674685, 839119, 255119, 552181, 1239034, 490552, 468715, 963242, 1612174, 21991, 1157603, 975554, 1342832, 883818, 1475260, 1666539, 1252155, 986632, 906429, 1495805, 561464, 1219097, 567704, 1463586, 1571831, 689821, 481310, 1407520, 97442, 512704, 962995, 876670, 956570, 1292534, 914981, 1655195, 870222, 196978, 617479, 1610748, 128227, 213356, 1590232, 230140, 942835]}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..0517af44e016fa642565d710ac6b25ad0d13a1d5
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d2a98c24a3c35def708b1d8aae8d57bd74e1a6edd57aa7018eb810e79479c62
+size 343042
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..9c6a10d60668ac906a3a52e306db65f174ea652b
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93d052bf5d679141026fb67a6593a10ef13b745c55458b3d2dfaa966b1a2e610
+size 419629
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..0c39fafdc3253d4d97f946d72e638ecdd00cc166
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a68dfd2a99c13640e1d495d84e9d4a2cbf4144997f74b88a22b3662dcee1af5c
+size 95948
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..dc926a2a58244bcea4b782f3ec6d63acc6c6d51b
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72857697bda43ef13819a12709769b01d76f31ce526bd5bce0f50364fc5bf602
+size 107413
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/training_log_9700d650-3889-4d26-873a-d01d4e4c55f3.txt b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/training_log_9700d650-3889-4d26-873a-d01d4e4c55f3.txt
new file mode 100644
index 0000000000000000000000000000000000000000..8a358e2214d1aee3806e0764b99117070c4880c9
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/training_log_9700d650-3889-4d26-873a-d01d4e4c55f3.txt
@@ -0,0 +1,5236 @@
+[2025-09-04 10:02:43] [Rank 0] PRINT: --- Script Start: Thu Sep  4 10:02:43 2025 ---
+[2025-09-04 10:02:43] [Rank 0] PRINT: --- Script Start: Thu Sep  4 10:02:43 2025 ---
+[2025-09-04 10:02:43] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=42, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 10:02:43] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=42, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 10:02:43] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 10:02:43] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 10:02:43] [Rank 0] PRINT: Using fixed seed: 42
+[2025-09-04 10:02:43] [Rank 0] PRINT: Using fixed seed: 42
+[2025-09-04 10:02:43] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42
+[2025-09-04 10:02:43] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42
+[2025-09-04 10:02:43] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 10:02:43] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 10:02:43] [Rank 0] PRINT: Constructing model...
+[2025-09-04 10:02:43] [Rank 0] PRINT: Constructing model...
+[2025-09-04 10:02:45] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 10:02:45] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 10:02:45] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 10:02:45] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 10:02:45] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 10:02:45] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 10:02:55] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 10:02:55] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 10:02:55] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 10:02:55] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 10:02:55] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 10:02:55] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 10:02:55] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 10:02:55] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 10:02:55] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 10:02:55] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 10:02:55] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 10:02:55] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 10:02:55] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 10:02:55] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 10:02:55] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 10:02:55] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 10:02:55] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 10:02:55] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 10:02:55] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 10:02:55] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 10:02:55] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 10:02:55] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 10:03:05] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 10:03:05] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 10:03:05] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 10:03:05] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 10:08:16] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 10:08:16] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 10:08:16] [Rank 0] PRINT: Starting training...
+[2025-09-04 10:08:16] [Rank 0] PRINT: Starting training...
+[2025-09-04 10:08:23] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/fixed_eval_indices.json
+[2025-09-04 10:08:23] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/fixed_eval_indices.json
+[2025-09-04 10:08:23] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:08:23] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:12:04] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 10:12:04] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 10:12:43] [Rank 0] step:21/10000 train_time:39493ms step_avg:1880.61ms
+[2025-09-04 10:12:43] [Rank 0] step:21/10000 train_time:39493ms step_avg:1880.61ms
+[2025-09-04 10:12:44] [Rank 0] step:41/10000 train_time:40240ms step_avg:981.47ms
+[2025-09-04 10:12:44] [Rank 0] step:41/10000 train_time:40240ms step_avg:981.47ms
+[2025-09-04 10:12:45] [Rank 0] step:61/10000 train_time:40988ms step_avg:671.93ms
+[2025-09-04 10:12:45] [Rank 0] step:61/10000 train_time:40988ms step_avg:671.93ms
+[2025-09-04 10:12:46] [Rank 0] step:81/10000 train_time:41734ms step_avg:515.24ms
+[2025-09-04 10:12:46] [Rank 0] step:81/10000 train_time:41734ms step_avg:515.24ms
+[2025-09-04 10:12:46] [Rank 0] step:101/10000 train_time:42481ms step_avg:420.61ms
+[2025-09-04 10:12:46] [Rank 0] step:101/10000 train_time:42481ms step_avg:420.61ms
+[2025-09-04 10:12:47] [Rank 0] step:121/10000 train_time:43231ms step_avg:357.28ms
+[2025-09-04 10:12:47] [Rank 0] step:121/10000 train_time:43231ms step_avg:357.28ms
+[2025-09-04 10:12:48] [Rank 0] step:141/10000 train_time:44159ms step_avg:313.19ms
+[2025-09-04 10:12:48] [Rank 0] step:141/10000 train_time:44159ms step_avg:313.19ms
+[2025-09-04 10:12:49] [Rank 0] step:161/10000 train_time:44949ms step_avg:279.18ms
+[2025-09-04 10:12:49] [Rank 0] step:161/10000 train_time:44949ms step_avg:279.18ms
+[2025-09-04 10:12:50] [Rank 0] step:181/10000 train_time:45695ms step_avg:252.46ms
+[2025-09-04 10:12:50] [Rank 0] step:181/10000 train_time:45695ms step_avg:252.46ms
+[2025-09-04 10:12:51] [Rank 0] step:201/10000 train_time:46598ms step_avg:231.83ms
+[2025-09-04 10:12:51] [Rank 0] step:201/10000 train_time:46598ms step_avg:231.83ms
+[2025-09-04 10:12:51] [Rank 0] step:221/10000 train_time:47482ms step_avg:214.85ms
+[2025-09-04 10:12:51] [Rank 0] step:221/10000 train_time:47482ms step_avg:214.85ms
+[2025-09-04 10:12:52] [Rank 0] step:241/10000 train_time:48230ms step_avg:200.12ms
+[2025-09-04 10:12:52] [Rank 0] step:241/10000 train_time:48230ms step_avg:200.12ms
+[2025-09-04 10:12:53] [Rank 0] step:261/10000 train_time:48976ms step_avg:187.65ms
+[2025-09-04 10:12:53] [Rank 0] step:261/10000 train_time:48976ms step_avg:187.65ms
+[2025-09-04 10:12:54] [Rank 0] step:281/10000 train_time:49723ms step_avg:176.95ms
+[2025-09-04 10:12:54] [Rank 0] step:281/10000 train_time:49723ms step_avg:176.95ms
+[2025-09-04 10:12:54] [Rank 0] step:301/10000 train_time:50470ms step_avg:167.68ms
+[2025-09-04 10:12:54] [Rank 0] step:301/10000 train_time:50470ms step_avg:167.68ms
+[2025-09-04 10:12:55] [Rank 0] step:321/10000 train_time:51219ms step_avg:159.56ms
+[2025-09-04 10:12:55] [Rank 0] step:321/10000 train_time:51219ms step_avg:159.56ms
+[2025-09-04 10:12:56] [Rank 0] step:341/10000 train_time:51965ms step_avg:152.39ms
+[2025-09-04 10:12:56] [Rank 0] step:341/10000 train_time:51965ms step_avg:152.39ms
+[2025-09-04 10:12:57] [Rank 0] step:361/10000 train_time:52712ms step_avg:146.02ms
+[2025-09-04 10:12:57] [Rank 0] step:361/10000 train_time:52712ms step_avg:146.02ms
+[2025-09-04 10:12:57] [Rank 0] step:381/10000 train_time:53459ms step_avg:140.31ms
+[2025-09-04 10:12:57] [Rank 0] step:381/10000 train_time:53459ms step_avg:140.31ms
+[2025-09-04 10:12:58] [Rank 0] step:401/10000 train_time:54206ms step_avg:135.18ms
+[2025-09-04 10:12:58] [Rank 0] step:401/10000 train_time:54206ms step_avg:135.18ms
+[2025-09-04 10:12:59] [Rank 0] step:421/10000 train_time:54953ms step_avg:130.53ms
+[2025-09-04 10:12:59] [Rank 0] step:421/10000 train_time:54953ms step_avg:130.53ms
+[2025-09-04 10:13:00] [Rank 0] step:441/10000 train_time:55699ms step_avg:126.30ms
+[2025-09-04 10:13:00] [Rank 0] step:441/10000 train_time:55699ms step_avg:126.30ms
+[2025-09-04 10:13:00] [Rank 0] step:461/10000 train_time:56446ms step_avg:122.44ms
+[2025-09-04 10:13:00] [Rank 0] step:461/10000 train_time:56446ms step_avg:122.44ms
+[2025-09-04 10:13:01] [Rank 0] step:481/10000 train_time:57194ms step_avg:118.91ms
+[2025-09-04 10:13:01] [Rank 0] step:481/10000 train_time:57194ms step_avg:118.91ms
+[2025-09-04 10:13:02] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:13:02] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:13:02] [Rank 0] PRINT: step:500/10000 train_loss:3.1253 val_loss:1.1188 train_time:57948ms step_avg:115.90ms
+[2025-09-04 10:13:02] [Rank 0] PRINT: step:500/10000 train_loss:3.1253 val_loss:1.1188 train_time:57948ms step_avg:115.90ms
+[2025-09-04 10:13:02] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:13:02] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:13:03] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:13:03] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:14:42] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:14:42] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:14:42] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:14:42] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:14:42] [Rank 0]   Total Loss: 3.8258
+[2025-09-04 10:14:42] [Rank 0]   Total Loss: 3.8258
+[2025-09-04 10:14:42] [Rank 0]   Total FTA (Unweighted): 0.4837
+[2025-09-04 10:14:42] [Rank 0]   Total FTA (Unweighted): 0.4837
+[2025-09-04 10:14:42] [Rank 0]   Total FTA (Weighted):   0.4838
+[2025-09-04 10:14:42] [Rank 0]   Total FTA (Weighted):   0.4838
+[2025-09-04 10:14:42] [Rank 0]   Group 0 Loss: 3.4149
+[2025-09-04 10:14:42] [Rank 0]   Group 0 Loss: 3.4149
+[2025-09-04 10:14:42] [Rank 0]   Group 1 Loss: 3.1047
+[2025-09-04 10:14:42] [Rank 0]   Group 1 Loss: 3.1047
+[2025-09-04 10:14:42] [Rank 0]   Group 2 Loss: 3.0761
+[2025-09-04 10:14:42] [Rank 0]   Group 2 Loss: 3.0761
+[2025-09-04 10:14:42] [Rank 0]   Group 3 Loss: 3.3908
+[2025-09-04 10:14:42] [Rank 0]   Group 3 Loss: 3.3908
+[2025-09-04 10:14:42] [Rank 0]   Group 4 Loss: 3.4400
+[2025-09-04 10:14:42] [Rank 0]   Group 4 Loss: 3.4400
+[2025-09-04 10:14:42] [Rank 0]   Group 5 Loss: 3.5095
+[2025-09-04 10:14:42] [Rank 0]   Group 5 Loss: 3.5095
+[2025-09-04 10:14:42] [Rank 0]   Group 6 Loss: 3.5134
+[2025-09-04 10:14:42] [Rank 0]   Group 6 Loss: 3.5134
+[2025-09-04 10:14:42] [Rank 0]   Group 7 Loss: 3.6630
+[2025-09-04 10:14:42] [Rank 0]   Group 7 Loss: 3.6630
+[2025-09-04 10:14:42] [Rank 0]   Group 8 Loss: 3.9506
+[2025-09-04 10:14:42] [Rank 0]   Group 8 Loss: 3.9506
+[2025-09-04 10:14:42] [Rank 0]   Group 9 Loss: 4.0008
+[2025-09-04 10:14:42] [Rank 0]   Group 9 Loss: 4.0008
+[2025-09-04 10:14:42] [Rank 0]   Group 10 Loss: 4.1944
+[2025-09-04 10:14:42] [Rank 0]   Group 10 Loss: 4.1944
+[2025-09-04 10:14:42] [Rank 0]   Group 11 Loss: 4.2726
+[2025-09-04 10:14:42] [Rank 0]   Group 11 Loss: 4.2726
+[2025-09-04 10:14:43] [Rank 0]   Group 12 Loss: 4.3389
+[2025-09-04 10:14:43] [Rank 0]   Group 12 Loss: 4.3389
+[2025-09-04 10:14:43] [Rank 0]   Group 13 Loss: 4.4744
+[2025-09-04 10:14:43] [Rank 0]   Group 13 Loss: 4.4744
+[2025-09-04 10:14:43] [Rank 0]   Group 14 Loss: 4.4265
+[2025-09-04 10:14:43] [Rank 0]   Group 14 Loss: 4.4265
+[2025-09-04 10:14:43] [Rank 0]   Group 15 Loss: 4.4419
+[2025-09-04 10:14:43] [Rank 0]   Group 15 Loss: 4.4419
+[2025-09-04 10:14:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:14:43] [Rank 0]   Group 4 FTA: 0.9400
+[2025-09-04 10:14:43] [Rank 0]   Group 4 FTA: 0.9400
+[2025-09-04 10:14:43] [Rank 0]   Group 5 FTA: 0.6300
+[2025-09-04 10:14:43] [Rank 0]   Group 5 FTA: 0.6300
+[2025-09-04 10:14:43] [Rank 0]   Group 6 FTA: 0.5100
+[2025-09-04 10:14:43] [Rank 0]   Group 6 FTA: 0.5100
+[2025-09-04 10:14:43] [Rank 0]   Group 7 FTA: 0.4400
+[2025-09-04 10:14:43] [Rank 0]   Group 7 FTA: 0.4400
+[2025-09-04 10:14:43] [Rank 0]   Group 8 FTA: 0.4100
+[2025-09-04 10:14:43] [Rank 0]   Group 8 FTA: 0.4100
+[2025-09-04 10:14:43] [Rank 0]   Group 9 FTA: 0.2000
+[2025-09-04 10:14:43] [Rank 0]   Group 9 FTA: 0.2000
+[2025-09-04 10:14:43] [Rank 0]   Group 10 FTA: 0.0700
+[2025-09-04 10:14:43] [Rank 0]   Group 10 FTA: 0.0700
+[2025-09-04 10:14:43] [Rank 0]   Group 11 FTA: 0.1100
+[2025-09-04 10:14:43] [Rank 0]   Group 11 FTA: 0.1100
+[2025-09-04 10:14:43] [Rank 0]   Group 12 FTA: 0.0900
+[2025-09-04 10:14:43] [Rank 0]   Group 12 FTA: 0.0900
+[2025-09-04 10:14:43] [Rank 0]   Group 13 FTA: 0.1500
+[2025-09-04 10:14:43] [Rank 0]   Group 13 FTA: 0.1500
+[2025-09-04 10:14:43] [Rank 0]   Group 14 FTA: 0.0900
+[2025-09-04 10:14:43] [Rank 0]   Group 14 FTA: 0.0900
+[2025-09-04 10:14:43] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 10:14:43] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 10:14:43] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:14:43] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:14:43] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:14:43] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:14:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:14:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:14:44] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:14:44] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:14:44] [Rank 0] step:501/10000 train_time:57963ms step_avg:115.70ms
+[2025-09-04 10:14:44] [Rank 0] step:501/10000 train_time:57963ms step_avg:115.70ms
+[2025-09-04 10:14:45] [Rank 0] step:521/10000 train_time:58730ms step_avg:112.73ms
+[2025-09-04 10:14:45] [Rank 0] step:521/10000 train_time:58730ms step_avg:112.73ms
+[2025-09-04 10:14:46] [Rank 0] step:541/10000 train_time:59478ms step_avg:109.94ms
+[2025-09-04 10:14:46] [Rank 0] step:541/10000 train_time:59478ms step_avg:109.94ms
+[2025-09-04 10:14:46] [Rank 0] step:561/10000 train_time:60225ms step_avg:107.35ms
+[2025-09-04 10:14:46] [Rank 0] step:561/10000 train_time:60225ms step_avg:107.35ms
+[2025-09-04 10:14:47] [Rank 0] step:581/10000 train_time:60972ms step_avg:104.94ms
+[2025-09-04 10:14:47] [Rank 0] step:581/10000 train_time:60972ms step_avg:104.94ms
+[2025-09-04 10:14:48] [Rank 0] step:601/10000 train_time:61719ms step_avg:102.69ms
+[2025-09-04 10:14:48] [Rank 0] step:601/10000 train_time:61719ms step_avg:102.69ms
+[2025-09-04 10:14:49] [Rank 0] step:621/10000 train_time:62466ms step_avg:100.59ms
+[2025-09-04 10:14:49] [Rank 0] step:621/10000 train_time:62466ms step_avg:100.59ms
+[2025-09-04 10:14:49] [Rank 0] step:641/10000 train_time:63213ms step_avg:98.62ms
+[2025-09-04 10:14:49] [Rank 0] step:641/10000 train_time:63213ms step_avg:98.62ms
+[2025-09-04 10:14:50] [Rank 0] step:661/10000 train_time:63961ms step_avg:96.76ms
+[2025-09-04 10:14:50] [Rank 0] step:661/10000 train_time:63961ms step_avg:96.76ms
+[2025-09-04 10:14:51] [Rank 0] step:681/10000 train_time:64708ms step_avg:95.02ms
+[2025-09-04 10:14:51] [Rank 0] step:681/10000 train_time:64708ms step_avg:95.02ms
+[2025-09-04 10:14:52] [Rank 0] step:701/10000 train_time:65456ms step_avg:93.38ms
+[2025-09-04 10:14:52] [Rank 0] step:701/10000 train_time:65456ms step_avg:93.38ms
+[2025-09-04 10:14:52] [Rank 0] step:721/10000 train_time:66203ms step_avg:91.82ms
+[2025-09-04 10:14:52] [Rank 0] step:721/10000 train_time:66203ms step_avg:91.82ms
+[2025-09-04 10:14:53] [Rank 0] step:741/10000 train_time:66949ms step_avg:90.35ms
+[2025-09-04 10:14:53] [Rank 0] step:741/10000 train_time:66949ms step_avg:90.35ms
+[2025-09-04 10:14:54] [Rank 0] step:761/10000 train_time:67706ms step_avg:88.97ms
+[2025-09-04 10:14:54] [Rank 0] step:761/10000 train_time:67706ms step_avg:88.97ms
+[2025-09-04 10:14:55] [Rank 0] step:781/10000 train_time:68761ms step_avg:88.04ms
+[2025-09-04 10:14:55] [Rank 0] step:781/10000 train_time:68761ms step_avg:88.04ms
+[2025-09-04 10:14:56] [Rank 0] step:801/10000 train_time:69513ms step_avg:86.78ms
+[2025-09-04 10:14:56] [Rank 0] step:801/10000 train_time:69513ms step_avg:86.78ms
+[2025-09-04 10:14:57] [Rank 0] step:821/10000 train_time:70535ms step_avg:85.91ms
+[2025-09-04 10:14:57] [Rank 0] step:821/10000 train_time:70535ms step_avg:85.91ms
+[2025-09-04 10:14:58] [Rank 0] step:841/10000 train_time:71596ms step_avg:85.13ms
+[2025-09-04 10:14:58] [Rank 0] step:841/10000 train_time:71596ms step_avg:85.13ms
+[2025-09-04 10:14:58] [Rank 0] step:861/10000 train_time:72346ms step_avg:84.03ms
+[2025-09-04 10:14:58] [Rank 0] step:861/10000 train_time:72346ms step_avg:84.03ms
+[2025-09-04 10:14:59] [Rank 0] step:881/10000 train_time:73097ms step_avg:82.97ms
+[2025-09-04 10:14:59] [Rank 0] step:881/10000 train_time:73097ms step_avg:82.97ms
+[2025-09-04 10:15:00] [Rank 0] step:901/10000 train_time:73848ms step_avg:81.96ms
+[2025-09-04 10:15:00] [Rank 0] step:901/10000 train_time:73848ms step_avg:81.96ms
+[2025-09-04 10:15:01] [Rank 0] step:921/10000 train_time:74599ms step_avg:81.00ms
+[2025-09-04 10:15:01] [Rank 0] step:921/10000 train_time:74599ms step_avg:81.00ms
+[2025-09-04 10:15:01] [Rank 0] step:941/10000 train_time:75397ms step_avg:80.12ms
+[2025-09-04 10:15:01] [Rank 0] step:941/10000 train_time:75397ms step_avg:80.12ms
+[2025-09-04 10:15:02] [Rank 0] step:961/10000 train_time:76203ms step_avg:79.30ms
+[2025-09-04 10:15:02] [Rank 0] step:961/10000 train_time:76203ms step_avg:79.30ms
+[2025-09-04 10:15:03] [Rank 0] step:981/10000 train_time:76954ms step_avg:78.44ms
+[2025-09-04 10:15:03] [Rank 0] step:981/10000 train_time:76954ms step_avg:78.44ms
+[2025-09-04 10:15:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:15:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:15:04] [Rank 0] PRINT: step:1000/10000 train_loss:0.9698 val_loss:0.8631 train_time:77711ms step_avg:77.71ms
+[2025-09-04 10:15:04] [Rank 0] PRINT: step:1000/10000 train_loss:0.9698 val_loss:0.8631 train_time:77711ms step_avg:77.71ms
+[2025-09-04 10:15:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:15:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:15:04] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:15:04] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:16:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:16:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:16:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:16:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:16:43] [Rank 0]   Total Loss: 4.2587
+[2025-09-04 10:16:43] [Rank 0]   Total Loss: 4.2587
+[2025-09-04 10:16:43] [Rank 0]   Total FTA (Unweighted): 0.6788
+[2025-09-04 10:16:43] [Rank 0]   Total FTA (Unweighted): 0.6788
+[2025-09-04 10:16:43] [Rank 0]   Total FTA (Weighted):   0.6787
+[2025-09-04 10:16:43] [Rank 0]   Total FTA (Weighted):   0.6787
+[2025-09-04 10:16:43] [Rank 0]   Group 0 Loss: 4.0637
+[2025-09-04 10:16:43] [Rank 0]   Group 0 Loss: 4.0637
+[2025-09-04 10:16:43] [Rank 0]   Group 1 Loss: 3.7140
+[2025-09-04 10:16:43] [Rank 0]   Group 1 Loss: 3.7140
+[2025-09-04 10:16:43] [Rank 0]   Group 2 Loss: 3.6738
+[2025-09-04 10:16:43] [Rank 0]   Group 2 Loss: 3.6738
+[2025-09-04 10:16:43] [Rank 0]   Group 3 Loss: 3.9391
+[2025-09-04 10:16:43] [Rank 0]   Group 3 Loss: 3.9391
+[2025-09-04 10:16:43] [Rank 0]   Group 4 Loss: 4.0430
+[2025-09-04 10:16:43] [Rank 0]   Group 4 Loss: 4.0430
+[2025-09-04 10:16:43] [Rank 0]   Group 5 Loss: 3.9945
+[2025-09-04 10:16:43] [Rank 0]   Group 5 Loss: 3.9945
+[2025-09-04 10:16:44] [Rank 0]   Group 6 Loss: 3.9691
+[2025-09-04 10:16:44] [Rank 0]   Group 6 Loss: 3.9691
+[2025-09-04 10:16:44] [Rank 0]   Group 7 Loss: 4.0251
+[2025-09-04 10:16:44] [Rank 0]   Group 7 Loss: 4.0251
+[2025-09-04 10:16:44] [Rank 0]   Group 8 Loss: 4.2706
+[2025-09-04 10:16:44] [Rank 0]   Group 8 Loss: 4.2706
+[2025-09-04 10:16:44] [Rank 0]   Group 9 Loss: 4.2406
+[2025-09-04 10:16:44] [Rank 0]   Group 9 Loss: 4.2406
+[2025-09-04 10:16:44] [Rank 0]   Group 10 Loss: 4.5018
+[2025-09-04 10:16:44] [Rank 0]   Group 10 Loss: 4.5018
+[2025-09-04 10:16:44] [Rank 0]   Group 11 Loss: 4.6086
+[2025-09-04 10:16:44] [Rank 0]   Group 11 Loss: 4.6086
+[2025-09-04 10:16:44] [Rank 0]   Group 12 Loss: 4.6144
+[2025-09-04 10:16:44] [Rank 0]   Group 12 Loss: 4.6144
+[2025-09-04 10:16:44] [Rank 0]   Group 13 Loss: 4.7808
+[2025-09-04 10:16:44] [Rank 0]   Group 13 Loss: 4.7808
+[2025-09-04 10:16:44] [Rank 0]   Group 14 Loss: 4.8129
+[2025-09-04 10:16:44] [Rank 0]   Group 14 Loss: 4.8129
+[2025-09-04 10:16:44] [Rank 0]   Group 15 Loss: 4.8867
+[2025-09-04 10:16:44] [Rank 0]   Group 15 Loss: 4.8867
+[2025-09-04 10:16:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:16:44] [Rank 0]   Group 7 FTA: 0.8800
+[2025-09-04 10:16:44] [Rank 0]   Group 7 FTA: 0.8800
+[2025-09-04 10:16:44] [Rank 0]   Group 8 FTA: 0.7600
+[2025-09-04 10:16:44] [Rank 0]   Group 8 FTA: 0.7600
+[2025-09-04 10:16:44] [Rank 0]   Group 9 FTA: 0.6300
+[2025-09-04 10:16:44] [Rank 0]   Group 9 FTA: 0.6300
+[2025-09-04 10:16:44] [Rank 0]   Group 10 FTA: 0.6100
+[2025-09-04 10:16:44] [Rank 0]   Group 10 FTA: 0.6100
+[2025-09-04 10:16:44] [Rank 0]   Group 11 FTA: 0.4100
+[2025-09-04 10:16:44] [Rank 0]   Group 11 FTA: 0.4100
+[2025-09-04 10:16:44] [Rank 0]   Group 12 FTA: 0.2000
+[2025-09-04 10:16:44] [Rank 0]   Group 12 FTA: 0.2000
+[2025-09-04 10:16:44] [Rank 0]   Group 13 FTA: 0.1400
+[2025-09-04 10:16:44] [Rank 0]   Group 13 FTA: 0.1400
+[2025-09-04 10:16:44] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 10:16:44] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 10:16:44] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 10:16:44] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 10:16:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:16:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:16:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:16:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:16:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:16:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:16:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:16:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:16:45] [Rank 0] step:1001/10000 train_time:77725ms step_avg:77.65ms
+[2025-09-04 10:16:45] [Rank 0] step:1001/10000 train_time:77725ms step_avg:77.65ms
+[2025-09-04 10:16:46] [Rank 0] step:1021/10000 train_time:78502ms step_avg:76.89ms
+[2025-09-04 10:16:46] [Rank 0] step:1021/10000 train_time:78502ms step_avg:76.89ms
+[2025-09-04 10:16:47] [Rank 0] step:1041/10000 train_time:79253ms step_avg:76.13ms
+[2025-09-04 10:16:47] [Rank 0] step:1041/10000 train_time:79253ms step_avg:76.13ms
+[2025-09-04 10:16:47] [Rank 0] step:1061/10000 train_time:80004ms step_avg:75.40ms
+[2025-09-04 10:16:47] [Rank 0] step:1061/10000 train_time:80004ms step_avg:75.40ms
+[2025-09-04 10:16:48] [Rank 0] step:1081/10000 train_time:80755ms step_avg:74.70ms
+[2025-09-04 10:16:48] [Rank 0] step:1081/10000 train_time:80755ms step_avg:74.70ms
+[2025-09-04 10:16:49] [Rank 0] step:1101/10000 train_time:81506ms step_avg:74.03ms
+[2025-09-04 10:16:49] [Rank 0] step:1101/10000 train_time:81506ms step_avg:74.03ms
+[2025-09-04 10:16:50] [Rank 0] step:1121/10000 train_time:82257ms step_avg:73.38ms
+[2025-09-04 10:16:50] [Rank 0] step:1121/10000 train_time:82257ms step_avg:73.38ms
+[2025-09-04 10:16:50] [Rank 0] step:1141/10000 train_time:83008ms step_avg:72.75ms
+[2025-09-04 10:16:50] [Rank 0] step:1141/10000 train_time:83008ms step_avg:72.75ms
+[2025-09-04 10:16:51] [Rank 0] step:1161/10000 train_time:83759ms step_avg:72.14ms
+[2025-09-04 10:16:51] [Rank 0] step:1161/10000 train_time:83759ms step_avg:72.14ms
+[2025-09-04 10:16:52] [Rank 0] step:1181/10000 train_time:84510ms step_avg:71.56ms
+[2025-09-04 10:16:52] [Rank 0] step:1181/10000 train_time:84510ms step_avg:71.56ms
+[2025-09-04 10:16:53] [Rank 0] step:1201/10000 train_time:85262ms step_avg:70.99ms
+[2025-09-04 10:16:53] [Rank 0] step:1201/10000 train_time:85262ms step_avg:70.99ms
+[2025-09-04 10:16:53] [Rank 0] step:1221/10000 train_time:86014ms step_avg:70.45ms
+[2025-09-04 10:16:53] [Rank 0] step:1221/10000 train_time:86014ms step_avg:70.45ms
+[2025-09-04 10:16:54] [Rank 0] step:1241/10000 train_time:86765ms step_avg:69.92ms
+[2025-09-04 10:16:54] [Rank 0] step:1241/10000 train_time:86765ms step_avg:69.92ms
+[2025-09-04 10:16:55] [Rank 0] step:1261/10000 train_time:87518ms step_avg:69.40ms
+[2025-09-04 10:16:55] [Rank 0] step:1261/10000 train_time:87518ms step_avg:69.40ms
+[2025-09-04 10:16:56] [Rank 0] step:1281/10000 train_time:88270ms step_avg:68.91ms
+[2025-09-04 10:16:56] [Rank 0] step:1281/10000 train_time:88270ms step_avg:68.91ms
+[2025-09-04 10:16:56] [Rank 0] step:1301/10000 train_time:89022ms step_avg:68.43ms
+[2025-09-04 10:16:56] [Rank 0] step:1301/10000 train_time:89022ms step_avg:68.43ms
+[2025-09-04 10:16:57] [Rank 0] step:1321/10000 train_time:89774ms step_avg:67.96ms
+[2025-09-04 10:16:57] [Rank 0] step:1321/10000 train_time:89774ms step_avg:67.96ms
+[2025-09-04 10:16:58] [Rank 0] step:1341/10000 train_time:90527ms step_avg:67.51ms
+[2025-09-04 10:16:58] [Rank 0] step:1341/10000 train_time:90527ms step_avg:67.51ms
+[2025-09-04 10:16:59] [Rank 0] step:1361/10000 train_time:91278ms step_avg:67.07ms
+[2025-09-04 10:16:59] [Rank 0] step:1361/10000 train_time:91278ms step_avg:67.07ms
+[2025-09-04 10:16:59] [Rank 0] step:1381/10000 train_time:92039ms step_avg:66.65ms
+[2025-09-04 10:16:59] [Rank 0] step:1381/10000 train_time:92039ms step_avg:66.65ms
+[2025-09-04 10:17:00] [Rank 0] step:1401/10000 train_time:92791ms step_avg:66.23ms
+[2025-09-04 10:17:00] [Rank 0] step:1401/10000 train_time:92791ms step_avg:66.23ms
+[2025-09-04 10:17:01] [Rank 0] step:1421/10000 train_time:93797ms step_avg:66.01ms
+[2025-09-04 10:17:01] [Rank 0] step:1421/10000 train_time:93797ms step_avg:66.01ms
+[2025-09-04 10:17:02] [Rank 0] step:1441/10000 train_time:94549ms step_avg:65.61ms
+[2025-09-04 10:17:02] [Rank 0] step:1441/10000 train_time:94549ms step_avg:65.61ms
+[2025-09-04 10:17:03] [Rank 0] step:1461/10000 train_time:95302ms step_avg:65.23ms
+[2025-09-04 10:17:03] [Rank 0] step:1461/10000 train_time:95302ms step_avg:65.23ms
+[2025-09-04 10:17:04] [Rank 0] step:1481/10000 train_time:96333ms step_avg:65.05ms
+[2025-09-04 10:17:04] [Rank 0] step:1481/10000 train_time:96333ms step_avg:65.05ms
+[2025-09-04 10:17:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:17:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:17:05] [Rank 0] PRINT: step:1500/10000 train_loss:0.8269 val_loss:0.7809 train_time:97091ms step_avg:64.73ms
+[2025-09-04 10:17:05] [Rank 0] PRINT: step:1500/10000 train_loss:0.8269 val_loss:0.7809 train_time:97091ms step_avg:64.73ms
+[2025-09-04 10:17:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:17:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:17:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:17:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:18:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:18:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:18:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:18:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:18:43] [Rank 0]   Total Loss: 4.3542
+[2025-09-04 10:18:43] [Rank 0]   Total Loss: 4.3542
+[2025-09-04 10:18:43] [Rank 0]   Total FTA (Unweighted): 0.7619
+[2025-09-04 10:18:43] [Rank 0]   Total FTA (Unweighted): 0.7619
+[2025-09-04 10:18:43] [Rank 0]   Total FTA (Weighted):   0.7619
+[2025-09-04 10:18:43] [Rank 0]   Total FTA (Weighted):   0.7619
+[2025-09-04 10:18:43] [Rank 0]   Group 0 Loss: 4.3814
+[2025-09-04 10:18:43] [Rank 0]   Group 0 Loss: 4.3814
+[2025-09-04 10:18:43] [Rank 0]   Group 1 Loss: 3.9575
+[2025-09-04 10:18:43] [Rank 0]   Group 1 Loss: 3.9575
+[2025-09-04 10:18:43] [Rank 0]   Group 2 Loss: 3.8489
+[2025-09-04 10:18:43] [Rank 0]   Group 2 Loss: 3.8489
+[2025-09-04 10:18:43] [Rank 0]   Group 3 Loss: 4.1024
+[2025-09-04 10:18:43] [Rank 0]   Group 3 Loss: 4.1024
+[2025-09-04 10:18:43] [Rank 0]   Group 4 Loss: 4.1703
+[2025-09-04 10:18:43] [Rank 0]   Group 4 Loss: 4.1703
+[2025-09-04 10:18:43] [Rank 0]   Group 5 Loss: 4.1865
+[2025-09-04 10:18:43] [Rank 0]   Group 5 Loss: 4.1865
+[2025-09-04 10:18:43] [Rank 0]   Group 6 Loss: 4.0668
+[2025-09-04 10:18:43] [Rank 0]   Group 6 Loss: 4.0668
+[2025-09-04 10:18:43] [Rank 0]   Group 7 Loss: 4.1641
+[2025-09-04 10:18:43] [Rank 0]   Group 7 Loss: 4.1641
+[2025-09-04 10:18:43] [Rank 0]   Group 8 Loss: 4.3402
+[2025-09-04 10:18:43] [Rank 0]   Group 8 Loss: 4.3402
+[2025-09-04 10:18:43] [Rank 0]   Group 9 Loss: 4.3358
+[2025-09-04 10:18:43] [Rank 0]   Group 9 Loss: 4.3358
+[2025-09-04 10:18:43] [Rank 0]   Group 10 Loss: 4.4748
+[2025-09-04 10:18:43] [Rank 0]   Group 10 Loss: 4.4748
+[2025-09-04 10:18:43] [Rank 0]   Group 11 Loss: 4.5495
+[2025-09-04 10:18:43] [Rank 0]   Group 11 Loss: 4.5495
+[2025-09-04 10:18:43] [Rank 0]   Group 12 Loss: 4.5569
+[2025-09-04 10:18:43] [Rank 0]   Group 12 Loss: 4.5569
+[2025-09-04 10:18:43] [Rank 0]   Group 13 Loss: 4.7689
+[2025-09-04 10:18:43] [Rank 0]   Group 13 Loss: 4.7689
+[2025-09-04 10:18:43] [Rank 0]   Group 14 Loss: 4.8238
+[2025-09-04 10:18:43] [Rank 0]   Group 14 Loss: 4.8238
+[2025-09-04 10:18:43] [Rank 0]   Group 15 Loss: 4.9400
+[2025-09-04 10:18:43] [Rank 0]   Group 15 Loss: 4.9400
+[2025-09-04 10:18:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:18:43] [Rank 0]   Group 8 FTA: 0.9600
+[2025-09-04 10:18:43] [Rank 0]   Group 8 FTA: 0.9600
+[2025-09-04 10:18:43] [Rank 0]   Group 9 FTA: 0.8100
+[2025-09-04 10:18:43] [Rank 0]   Group 9 FTA: 0.8100
+[2025-09-04 10:18:43] [Rank 0]   Group 10 FTA: 0.8400
+[2025-09-04 10:18:43] [Rank 0]   Group 10 FTA: 0.8400
+[2025-09-04 10:18:43] [Rank 0]   Group 11 FTA: 0.6700
+[2025-09-04 10:18:43] [Rank 0]   Group 11 FTA: 0.6700
+[2025-09-04 10:18:43] [Rank 0]   Group 12 FTA: 0.4500
+[2025-09-04 10:18:43] [Rank 0]   Group 12 FTA: 0.4500
+[2025-09-04 10:18:43] [Rank 0]   Group 13 FTA: 0.2000
+[2025-09-04 10:18:43] [Rank 0]   Group 13 FTA: 0.2000
+[2025-09-04 10:18:43] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 10:18:43] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 10:18:43] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 10:18:43] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 10:18:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:18:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:18:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:18:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:18:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:18:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:18:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:18:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:18:45] [Rank 0] step:1501/10000 train_time:97107ms step_avg:64.69ms
+[2025-09-04 10:18:45] [Rank 0] step:1501/10000 train_time:97107ms step_avg:64.69ms
+[2025-09-04 10:18:45] [Rank 0] step:1521/10000 train_time:97874ms step_avg:64.35ms
+[2025-09-04 10:18:45] [Rank 0] step:1521/10000 train_time:97874ms step_avg:64.35ms
+[2025-09-04 10:18:46] [Rank 0] step:1541/10000 train_time:98626ms step_avg:64.00ms
+[2025-09-04 10:18:46] [Rank 0] step:1541/10000 train_time:98626ms step_avg:64.00ms
+[2025-09-04 10:18:47] [Rank 0] step:1561/10000 train_time:99377ms step_avg:63.66ms
+[2025-09-04 10:18:47] [Rank 0] step:1561/10000 train_time:99377ms step_avg:63.66ms
+[2025-09-04 10:18:48] [Rank 0] step:1581/10000 train_time:100129ms step_avg:63.33ms
+[2025-09-04 10:18:48] [Rank 0] step:1581/10000 train_time:100129ms step_avg:63.33ms
+[2025-09-04 10:18:49] [Rank 0] step:1601/10000 train_time:100881ms step_avg:63.01ms
+[2025-09-04 10:18:49] [Rank 0] step:1601/10000 train_time:100881ms step_avg:63.01ms
+[2025-09-04 10:18:49] [Rank 0] step:1621/10000 train_time:101632ms step_avg:62.70ms
+[2025-09-04 10:18:49] [Rank 0] step:1621/10000 train_time:101632ms step_avg:62.70ms
+[2025-09-04 10:18:50] [Rank 0] step:1641/10000 train_time:102658ms step_avg:62.56ms
+[2025-09-04 10:18:50] [Rank 0] step:1641/10000 train_time:102658ms step_avg:62.56ms
+[2025-09-04 10:18:51] [Rank 0] step:1661/10000 train_time:103409ms step_avg:62.26ms
+[2025-09-04 10:18:51] [Rank 0] step:1661/10000 train_time:103409ms step_avg:62.26ms
+[2025-09-04 10:18:52] [Rank 0] step:1681/10000 train_time:104160ms step_avg:61.96ms
+[2025-09-04 10:18:52] [Rank 0] step:1681/10000 train_time:104160ms step_avg:61.96ms
+[2025-09-04 10:18:53] [Rank 0] step:1701/10000 train_time:104911ms step_avg:61.68ms
+[2025-09-04 10:18:53] [Rank 0] step:1701/10000 train_time:104911ms step_avg:61.68ms
+[2025-09-04 10:18:53] [Rank 0] step:1721/10000 train_time:105662ms step_avg:61.40ms
+[2025-09-04 10:18:53] [Rank 0] step:1721/10000 train_time:105662ms step_avg:61.40ms
+[2025-09-04 10:18:54] [Rank 0] step:1741/10000 train_time:106414ms step_avg:61.12ms
+[2025-09-04 10:18:54] [Rank 0] step:1741/10000 train_time:106414ms step_avg:61.12ms
+[2025-09-04 10:18:55] [Rank 0] step:1761/10000 train_time:107166ms step_avg:60.86ms
+[2025-09-04 10:18:55] [Rank 0] step:1761/10000 train_time:107166ms step_avg:60.86ms
+[2025-09-04 10:18:56] [Rank 0] step:1781/10000 train_time:107918ms step_avg:60.59ms
+[2025-09-04 10:18:56] [Rank 0] step:1781/10000 train_time:107918ms step_avg:60.59ms
+[2025-09-04 10:18:56] [Rank 0] step:1801/10000 train_time:108670ms step_avg:60.34ms
+[2025-09-04 10:18:56] [Rank 0] step:1801/10000 train_time:108670ms step_avg:60.34ms
+[2025-09-04 10:18:57] [Rank 0] step:1821/10000 train_time:109423ms step_avg:60.09ms
+[2025-09-04 10:18:57] [Rank 0] step:1821/10000 train_time:109423ms step_avg:60.09ms
+[2025-09-04 10:18:58] [Rank 0] step:1841/10000 train_time:110175ms step_avg:59.84ms
+[2025-09-04 10:18:58] [Rank 0] step:1841/10000 train_time:110175ms step_avg:59.84ms
+[2025-09-04 10:18:59] [Rank 0] step:1861/10000 train_time:110927ms step_avg:59.61ms
+[2025-09-04 10:18:59] [Rank 0] step:1861/10000 train_time:110927ms step_avg:59.61ms
+[2025-09-04 10:18:59] [Rank 0] step:1881/10000 train_time:111679ms step_avg:59.37ms
+[2025-09-04 10:18:59] [Rank 0] step:1881/10000 train_time:111679ms step_avg:59.37ms
+[2025-09-04 10:19:00] [Rank 0] step:1901/10000 train_time:112431ms step_avg:59.14ms
+[2025-09-04 10:19:00] [Rank 0] step:1901/10000 train_time:112431ms step_avg:59.14ms
+[2025-09-04 10:19:01] [Rank 0] step:1921/10000 train_time:113183ms step_avg:58.92ms
+[2025-09-04 10:19:01] [Rank 0] step:1921/10000 train_time:113183ms step_avg:58.92ms
+[2025-09-04 10:19:02] [Rank 0] step:1941/10000 train_time:113935ms step_avg:58.70ms
+[2025-09-04 10:19:02] [Rank 0] step:1941/10000 train_time:113935ms step_avg:58.70ms
+[2025-09-04 10:19:02] [Rank 0] step:1961/10000 train_time:114688ms step_avg:58.48ms
+[2025-09-04 10:19:02] [Rank 0] step:1961/10000 train_time:114688ms step_avg:58.48ms
+[2025-09-04 10:19:03] [Rank 0] step:1981/10000 train_time:115439ms step_avg:58.27ms
+[2025-09-04 10:19:03] [Rank 0] step:1981/10000 train_time:115439ms step_avg:58.27ms
+[2025-09-04 10:19:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:19:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:19:04] [Rank 0] PRINT: step:2000/10000 train_loss:0.7682 val_loss:0.7356 train_time:116196ms step_avg:58.10ms
+[2025-09-04 10:19:04] [Rank 0] PRINT: step:2000/10000 train_loss:0.7682 val_loss:0.7356 train_time:116196ms step_avg:58.10ms
+[2025-09-04 10:19:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:19:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:19:04] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:19:04] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:20:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:20:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:20:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:20:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:20:43] [Rank 0]   Total Loss: 4.5559
+[2025-09-04 10:20:43] [Rank 0]   Total Loss: 4.5559
+[2025-09-04 10:20:43] [Rank 0]   Total FTA (Unweighted): 0.8056
+[2025-09-04 10:20:43] [Rank 0]   Total FTA (Unweighted): 0.8056
+[2025-09-04 10:20:43] [Rank 0]   Total FTA (Weighted):   0.8056
+[2025-09-04 10:20:43] [Rank 0]   Total FTA (Weighted):   0.8056
+[2025-09-04 10:20:43] [Rank 0]   Group 0 Loss: 4.5276
+[2025-09-04 10:20:43] [Rank 0]   Group 0 Loss: 4.5276
+[2025-09-04 10:20:43] [Rank 0]   Group 1 Loss: 4.1393
+[2025-09-04 10:20:43] [Rank 0]   Group 1 Loss: 4.1393
+[2025-09-04 10:20:43] [Rank 0]   Group 2 Loss: 4.0222
+[2025-09-04 10:20:43] [Rank 0]   Group 2 Loss: 4.0222
+[2025-09-04 10:20:43] [Rank 0]   Group 3 Loss: 4.4135
+[2025-09-04 10:20:43] [Rank 0]   Group 3 Loss: 4.4135
+[2025-09-04 10:20:43] [Rank 0]   Group 4 Loss: 4.3583
+[2025-09-04 10:20:43] [Rank 0]   Group 4 Loss: 4.3583
+[2025-09-04 10:20:43] [Rank 0]   Group 5 Loss: 4.4570
+[2025-09-04 10:20:43] [Rank 0]   Group 5 Loss: 4.4570
+[2025-09-04 10:20:43] [Rank 0]   Group 6 Loss: 4.3417
+[2025-09-04 10:20:43] [Rank 0]   Group 6 Loss: 4.3417
+[2025-09-04 10:20:43] [Rank 0]   Group 7 Loss: 4.4242
+[2025-09-04 10:20:43] [Rank 0]   Group 7 Loss: 4.4242
+[2025-09-04 10:20:43] [Rank 0]   Group 8 Loss: 4.6096
+[2025-09-04 10:20:43] [Rank 0]   Group 8 Loss: 4.6096
+[2025-09-04 10:20:43] [Rank 0]   Group 9 Loss: 4.5872
+[2025-09-04 10:20:43] [Rank 0]   Group 9 Loss: 4.5872
+[2025-09-04 10:20:43] [Rank 0]   Group 10 Loss: 4.6598
+[2025-09-04 10:20:43] [Rank 0]   Group 10 Loss: 4.6598
+[2025-09-04 10:20:43] [Rank 0]   Group 11 Loss: 4.7817
+[2025-09-04 10:20:43] [Rank 0]   Group 11 Loss: 4.7817
+[2025-09-04 10:20:43] [Rank 0]   Group 12 Loss: 4.7103
+[2025-09-04 10:20:43] [Rank 0]   Group 12 Loss: 4.7103
+[2025-09-04 10:20:43] [Rank 0]   Group 13 Loss: 4.9096
+[2025-09-04 10:20:43] [Rank 0]   Group 13 Loss: 4.9096
+[2025-09-04 10:20:43] [Rank 0]   Group 14 Loss: 4.9042
+[2025-09-04 10:20:43] [Rank 0]   Group 14 Loss: 4.9042
+[2025-09-04 10:20:43] [Rank 0]   Group 15 Loss: 5.0481
+[2025-09-04 10:20:43] [Rank 0]   Group 15 Loss: 5.0481
+[2025-09-04 10:20:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:20:43] [Rank 0]   Group 9 FTA: 0.9500
+[2025-09-04 10:20:43] [Rank 0]   Group 9 FTA: 0.9500
+[2025-09-04 10:20:43] [Rank 0]   Group 10 FTA: 0.9500
+[2025-09-04 10:20:43] [Rank 0]   Group 10 FTA: 0.9500
+[2025-09-04 10:20:43] [Rank 0]   Group 11 FTA: 0.8200
+[2025-09-04 10:20:43] [Rank 0]   Group 11 FTA: 0.8200
+[2025-09-04 10:20:43] [Rank 0]   Group 12 FTA: 0.6300
+[2025-09-04 10:20:43] [Rank 0]   Group 12 FTA: 0.6300
+[2025-09-04 10:20:43] [Rank 0]   Group 13 FTA: 0.2900
+[2025-09-04 10:20:43] [Rank 0]   Group 13 FTA: 0.2900
+[2025-09-04 10:20:43] [Rank 0]   Group 14 FTA: 0.1400
+[2025-09-04 10:20:43] [Rank 0]   Group 14 FTA: 0.1400
+[2025-09-04 10:20:43] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 10:20:43] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 10:20:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:20:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:20:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:20:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:20:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:20:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:20:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:20:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:20:45] [Rank 0] step:2001/10000 train_time:116213ms step_avg:58.08ms
+[2025-09-04 10:20:45] [Rank 0] step:2001/10000 train_time:116213ms step_avg:58.08ms
+[2025-09-04 10:20:46] [Rank 0] step:2021/10000 train_time:117234ms step_avg:58.01ms
+[2025-09-04 10:20:46] [Rank 0] step:2021/10000 train_time:117234ms step_avg:58.01ms
+[2025-09-04 10:20:46] [Rank 0] step:2041/10000 train_time:117985ms step_avg:57.81ms
+[2025-09-04 10:20:46] [Rank 0] step:2041/10000 train_time:117985ms step_avg:57.81ms
+[2025-09-04 10:20:47] [Rank 0] step:2061/10000 train_time:118736ms step_avg:57.61ms
+[2025-09-04 10:20:47] [Rank 0] step:2061/10000 train_time:118736ms step_avg:57.61ms
+[2025-09-04 10:20:48] [Rank 0] step:2081/10000 train_time:119487ms step_avg:57.42ms
+[2025-09-04 10:20:48] [Rank 0] step:2081/10000 train_time:119487ms step_avg:57.42ms
+[2025-09-04 10:20:49] [Rank 0] step:2101/10000 train_time:120239ms step_avg:57.23ms
+[2025-09-04 10:20:49] [Rank 0] step:2101/10000 train_time:120239ms step_avg:57.23ms
+[2025-09-04 10:20:49] [Rank 0] step:2121/10000 train_time:120990ms step_avg:57.04ms
+[2025-09-04 10:20:49] [Rank 0] step:2121/10000 train_time:120990ms step_avg:57.04ms
+[2025-09-04 10:20:50] [Rank 0] step:2141/10000 train_time:121743ms step_avg:56.86ms
+[2025-09-04 10:20:50] [Rank 0] step:2141/10000 train_time:121743ms step_avg:56.86ms
+[2025-09-04 10:20:51] [Rank 0] step:2161/10000 train_time:122496ms step_avg:56.68ms
+[2025-09-04 10:20:51] [Rank 0] step:2161/10000 train_time:122496ms step_avg:56.68ms
+[2025-09-04 10:20:52] [Rank 0] step:2181/10000 train_time:123248ms step_avg:56.51ms
+[2025-09-04 10:20:52] [Rank 0] step:2181/10000 train_time:123248ms step_avg:56.51ms
+[2025-09-04 10:20:52] [Rank 0] step:2201/10000 train_time:124001ms step_avg:56.34ms
+[2025-09-04 10:20:52] [Rank 0] step:2201/10000 train_time:124001ms step_avg:56.34ms
+[2025-09-04 10:20:53] [Rank 0] step:2221/10000 train_time:124760ms step_avg:56.17ms
+[2025-09-04 10:20:53] [Rank 0] step:2221/10000 train_time:124760ms step_avg:56.17ms
+[2025-09-04 10:20:54] [Rank 0] step:2241/10000 train_time:125521ms step_avg:56.01ms
+[2025-09-04 10:20:54] [Rank 0] step:2241/10000 train_time:125521ms step_avg:56.01ms
+[2025-09-04 10:20:55] [Rank 0] step:2261/10000 train_time:126284ms step_avg:55.85ms
+[2025-09-04 10:20:55] [Rank 0] step:2261/10000 train_time:126284ms step_avg:55.85ms
+[2025-09-04 10:20:56] [Rank 0] step:2281/10000 train_time:127046ms step_avg:55.70ms
+[2025-09-04 10:20:56] [Rank 0] step:2281/10000 train_time:127046ms step_avg:55.70ms
+[2025-09-04 10:20:56] [Rank 0] step:2301/10000 train_time:127808ms step_avg:55.54ms
+[2025-09-04 10:20:56] [Rank 0] step:2301/10000 train_time:127808ms step_avg:55.54ms
+[2025-09-04 10:20:57] [Rank 0] step:2321/10000 train_time:128571ms step_avg:55.39ms
+[2025-09-04 10:20:57] [Rank 0] step:2321/10000 train_time:128571ms step_avg:55.39ms
+[2025-09-04 10:20:58] [Rank 0] step:2341/10000 train_time:129333ms step_avg:55.25ms
+[2025-09-04 10:20:58] [Rank 0] step:2341/10000 train_time:129333ms step_avg:55.25ms
+[2025-09-04 10:20:59] [Rank 0] step:2361/10000 train_time:130095ms step_avg:55.10ms
+[2025-09-04 10:20:59] [Rank 0] step:2361/10000 train_time:130095ms step_avg:55.10ms
+[2025-09-04 10:20:59] [Rank 0] step:2381/10000 train_time:130858ms step_avg:54.96ms
+[2025-09-04 10:20:59] [Rank 0] step:2381/10000 train_time:130858ms step_avg:54.96ms
+[2025-09-04 10:21:00] [Rank 0] step:2401/10000 train_time:131620ms step_avg:54.82ms
+[2025-09-04 10:21:00] [Rank 0] step:2401/10000 train_time:131620ms step_avg:54.82ms
+[2025-09-04 10:21:01] [Rank 0] step:2421/10000 train_time:132384ms step_avg:54.68ms
+[2025-09-04 10:21:01] [Rank 0] step:2421/10000 train_time:132384ms step_avg:54.68ms
+[2025-09-04 10:21:02] [Rank 0] step:2441/10000 train_time:133147ms step_avg:54.55ms
+[2025-09-04 10:21:02] [Rank 0] step:2441/10000 train_time:133147ms step_avg:54.55ms
+[2025-09-04 10:21:02] [Rank 0] step:2461/10000 train_time:133910ms step_avg:54.41ms
+[2025-09-04 10:21:02] [Rank 0] step:2461/10000 train_time:133910ms step_avg:54.41ms
+[2025-09-04 10:21:03] [Rank 0] step:2481/10000 train_time:134674ms step_avg:54.28ms
+[2025-09-04 10:21:03] [Rank 0] step:2481/10000 train_time:134674ms step_avg:54.28ms
+[2025-09-04 10:21:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:21:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:21:04] [Rank 0] PRINT: step:2500/10000 train_loss:0.7311 val_loss:0.7034 train_time:135443ms step_avg:54.18ms
+[2025-09-04 10:21:04] [Rank 0] PRINT: step:2500/10000 train_loss:0.7311 val_loss:0.7034 train_time:135443ms step_avg:54.18ms
+[2025-09-04 10:21:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:21:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:21:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:21:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:22:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:22:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:22:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:22:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:22:44] [Rank 0]   Total Loss: 4.5857
+[2025-09-04 10:22:44] [Rank 0]   Total Loss: 4.5857
+[2025-09-04 10:22:44] [Rank 0]   Total FTA (Unweighted): 0.8356
+[2025-09-04 10:22:44] [Rank 0]   Total FTA (Unweighted): 0.8356
+[2025-09-04 10:22:44] [Rank 0]   Total FTA (Weighted):   0.8356
+[2025-09-04 10:22:44] [Rank 0]   Total FTA (Weighted):   0.8356
+[2025-09-04 10:22:44] [Rank 0]   Group 0 Loss: 4.5381
+[2025-09-04 10:22:44] [Rank 0]   Group 0 Loss: 4.5381
+[2025-09-04 10:22:44] [Rank 0]   Group 1 Loss: 4.2413
+[2025-09-04 10:22:44] [Rank 0]   Group 1 Loss: 4.2413
+[2025-09-04 10:22:44] [Rank 0]   Group 2 Loss: 4.0608
+[2025-09-04 10:22:44] [Rank 0]   Group 2 Loss: 4.0608
+[2025-09-04 10:22:44] [Rank 0]   Group 3 Loss: 4.5341
+[2025-09-04 10:22:44] [Rank 0]   Group 3 Loss: 4.5341
+[2025-09-04 10:22:44] [Rank 0]   Group 4 Loss: 4.4299
+[2025-09-04 10:22:44] [Rank 0]   Group 4 Loss: 4.4299
+[2025-09-04 10:22:44] [Rank 0]   Group 5 Loss: 4.5168
+[2025-09-04 10:22:44] [Rank 0]   Group 5 Loss: 4.5168
+[2025-09-04 10:22:44] [Rank 0]   Group 6 Loss: 4.4381
+[2025-09-04 10:22:44] [Rank 0]   Group 6 Loss: 4.4381
+[2025-09-04 10:22:44] [Rank 0]   Group 7 Loss: 4.4418
+[2025-09-04 10:22:44] [Rank 0]   Group 7 Loss: 4.4418
+[2025-09-04 10:22:44] [Rank 0]   Group 8 Loss: 4.6170
+[2025-09-04 10:22:44] [Rank 0]   Group 8 Loss: 4.6170
+[2025-09-04 10:22:44] [Rank 0]   Group 9 Loss: 4.6605
+[2025-09-04 10:22:44] [Rank 0]   Group 9 Loss: 4.6605
+[2025-09-04 10:22:44] [Rank 0]   Group 10 Loss: 4.7762
+[2025-09-04 10:22:44] [Rank 0]   Group 10 Loss: 4.7762
+[2025-09-04 10:22:44] [Rank 0]   Group 11 Loss: 4.7546
+[2025-09-04 10:22:44] [Rank 0]   Group 11 Loss: 4.7546
+[2025-09-04 10:22:44] [Rank 0]   Group 12 Loss: 4.7002
+[2025-09-04 10:22:44] [Rank 0]   Group 12 Loss: 4.7002
+[2025-09-04 10:22:44] [Rank 0]   Group 13 Loss: 4.8506
+[2025-09-04 10:22:44] [Rank 0]   Group 13 Loss: 4.8506
+[2025-09-04 10:22:44] [Rank 0]   Group 14 Loss: 4.8338
+[2025-09-04 10:22:44] [Rank 0]   Group 14 Loss: 4.8338
+[2025-09-04 10:22:44] [Rank 0]   Group 15 Loss: 4.9772
+[2025-09-04 10:22:44] [Rank 0]   Group 15 Loss: 4.9772
+[2025-09-04 10:22:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:22:44] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 10:22:44] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 10:22:44] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 10:22:44] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 10:22:44] [Rank 0]   Group 11 FTA: 0.9500
+[2025-09-04 10:22:44] [Rank 0]   Group 11 FTA: 0.9500
+[2025-09-04 10:22:44] [Rank 0]   Group 12 FTA: 0.7800
+[2025-09-04 10:22:44] [Rank 0]   Group 12 FTA: 0.7800
+[2025-09-04 10:22:44] [Rank 0]   Group 13 FTA: 0.4400
+[2025-09-04 10:22:44] [Rank 0]   Group 13 FTA: 0.4400
+[2025-09-04 10:22:44] [Rank 0]   Group 14 FTA: 0.1400
+[2025-09-04 10:22:44] [Rank 0]   Group 14 FTA: 0.1400
+[2025-09-04 10:22:44] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 10:22:44] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 10:22:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:22:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:22:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:22:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:22:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:22:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:22:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:22:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:22:45] [Rank 0] step:2501/10000 train_time:135459ms step_avg:54.16ms
+[2025-09-04 10:22:45] [Rank 0] step:2501/10000 train_time:135459ms step_avg:54.16ms
+[2025-09-04 10:22:46] [Rank 0] step:2521/10000 train_time:136232ms step_avg:54.04ms
+[2025-09-04 10:22:46] [Rank 0] step:2521/10000 train_time:136232ms step_avg:54.04ms
+[2025-09-04 10:22:47] [Rank 0] step:2541/10000 train_time:136993ms step_avg:53.91ms
+[2025-09-04 10:22:47] [Rank 0] step:2541/10000 train_time:136993ms step_avg:53.91ms
+[2025-09-04 10:22:48] [Rank 0] step:2561/10000 train_time:137755ms step_avg:53.79ms
+[2025-09-04 10:22:48] [Rank 0] step:2561/10000 train_time:137755ms step_avg:53.79ms
+[2025-09-04 10:22:48] [Rank 0] step:2581/10000 train_time:138516ms step_avg:53.67ms
+[2025-09-04 10:22:48] [Rank 0] step:2581/10000 train_time:138516ms step_avg:53.67ms
+[2025-09-04 10:22:49] [Rank 0] step:2601/10000 train_time:139277ms step_avg:53.55ms
+[2025-09-04 10:22:49] [Rank 0] step:2601/10000 train_time:139277ms step_avg:53.55ms
+[2025-09-04 10:22:50] [Rank 0] step:2621/10000 train_time:140039ms step_avg:53.43ms
+[2025-09-04 10:22:50] [Rank 0] step:2621/10000 train_time:140039ms step_avg:53.43ms
+[2025-09-04 10:22:51] [Rank 0] step:2641/10000 train_time:140800ms step_avg:53.31ms
+[2025-09-04 10:22:51] [Rank 0] step:2641/10000 train_time:140800ms step_avg:53.31ms
+[2025-09-04 10:22:51] [Rank 0] step:2661/10000 train_time:141561ms step_avg:53.20ms
+[2025-09-04 10:22:51] [Rank 0] step:2661/10000 train_time:141561ms step_avg:53.20ms
+[2025-09-04 10:22:52] [Rank 0] step:2681/10000 train_time:142322ms step_avg:53.09ms
+[2025-09-04 10:22:52] [Rank 0] step:2681/10000 train_time:142322ms step_avg:53.09ms
+[2025-09-04 10:22:53] [Rank 0] step:2701/10000 train_time:143084ms step_avg:52.97ms
+[2025-09-04 10:22:53] [Rank 0] step:2701/10000 train_time:143084ms step_avg:52.97ms
+[2025-09-04 10:22:54] [Rank 0] step:2721/10000 train_time:143845ms step_avg:52.86ms
+[2025-09-04 10:22:54] [Rank 0] step:2721/10000 train_time:143845ms step_avg:52.86ms
+[2025-09-04 10:22:55] [Rank 0] step:2741/10000 train_time:144606ms step_avg:52.76ms
+[2025-09-04 10:22:55] [Rank 0] step:2741/10000 train_time:144606ms step_avg:52.76ms
+[2025-09-04 10:22:55] [Rank 0] step:2761/10000 train_time:145367ms step_avg:52.65ms
+[2025-09-04 10:22:55] [Rank 0] step:2761/10000 train_time:145367ms step_avg:52.65ms
+[2025-09-04 10:22:56] [Rank 0] step:2781/10000 train_time:146129ms step_avg:52.55ms
+[2025-09-04 10:22:56] [Rank 0] step:2781/10000 train_time:146129ms step_avg:52.55ms
+[2025-09-04 10:22:57] [Rank 0] step:2801/10000 train_time:146891ms step_avg:52.44ms
+[2025-09-04 10:22:57] [Rank 0] step:2801/10000 train_time:146891ms step_avg:52.44ms
+[2025-09-04 10:22:58] [Rank 0] step:2821/10000 train_time:147925ms step_avg:52.44ms
+[2025-09-04 10:22:58] [Rank 0] step:2821/10000 train_time:147925ms step_avg:52.44ms
+[2025-09-04 10:22:59] [Rank 0] step:2841/10000 train_time:148687ms step_avg:52.34ms
+[2025-09-04 10:22:59] [Rank 0] step:2841/10000 train_time:148687ms step_avg:52.34ms
+[2025-09-04 10:22:59] [Rank 0] step:2861/10000 train_time:149449ms step_avg:52.24ms
+[2025-09-04 10:22:59] [Rank 0] step:2861/10000 train_time:149449ms step_avg:52.24ms
+[2025-09-04 10:23:00] [Rank 0] step:2881/10000 train_time:150212ms step_avg:52.14ms
+[2025-09-04 10:23:00] [Rank 0] step:2881/10000 train_time:150212ms step_avg:52.14ms
+[2025-09-04 10:23:01] [Rank 0] step:2901/10000 train_time:150975ms step_avg:52.04ms
+[2025-09-04 10:23:01] [Rank 0] step:2901/10000 train_time:150975ms step_avg:52.04ms
+[2025-09-04 10:23:02] [Rank 0] step:2921/10000 train_time:151841ms step_avg:51.98ms
+[2025-09-04 10:23:02] [Rank 0] step:2921/10000 train_time:151841ms step_avg:51.98ms
+[2025-09-04 10:23:03] [Rank 0] step:2941/10000 train_time:152603ms step_avg:51.89ms
+[2025-09-04 10:23:03] [Rank 0] step:2941/10000 train_time:152603ms step_avg:51.89ms
+[2025-09-04 10:23:03] [Rank 0] step:2961/10000 train_time:153365ms step_avg:51.80ms
+[2025-09-04 10:23:03] [Rank 0] step:2961/10000 train_time:153365ms step_avg:51.80ms
+[2025-09-04 10:23:04] [Rank 0] step:2981/10000 train_time:154128ms step_avg:51.70ms
+[2025-09-04 10:23:04] [Rank 0] step:2981/10000 train_time:154128ms step_avg:51.70ms
+[2025-09-04 10:23:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:23:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:23:05] [Rank 0] PRINT: step:3000/10000 train_loss:0.7048 val_loss:0.6841 train_time:154896ms step_avg:51.63ms
+[2025-09-04 10:23:05] [Rank 0] PRINT: step:3000/10000 train_loss:0.7048 val_loss:0.6841 train_time:154896ms step_avg:51.63ms
+[2025-09-04 10:23:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:23:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:23:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:23:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:24:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:24:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:24:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:24:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:24:44] [Rank 0]   Total Loss: 4.7578
+[2025-09-04 10:24:44] [Rank 0]   Total Loss: 4.7578
+[2025-09-04 10:24:44] [Rank 0]   Total FTA (Unweighted): 0.8650
+[2025-09-04 10:24:44] [Rank 0]   Total FTA (Unweighted): 0.8650
+[2025-09-04 10:24:44] [Rank 0]   Total FTA (Weighted):   0.8650
+[2025-09-04 10:24:44] [Rank 0]   Total FTA (Weighted):   0.8650
+[2025-09-04 10:24:44] [Rank 0]   Group 0 Loss: 4.7961
+[2025-09-04 10:24:44] [Rank 0]   Group 0 Loss: 4.7961
+[2025-09-04 10:24:44] [Rank 0]   Group 1 Loss: 4.2583
+[2025-09-04 10:24:44] [Rank 0]   Group 1 Loss: 4.2583
+[2025-09-04 10:24:44] [Rank 0]   Group 2 Loss: 4.2470
+[2025-09-04 10:24:44] [Rank 0]   Group 2 Loss: 4.2470
+[2025-09-04 10:24:44] [Rank 0]   Group 3 Loss: 4.6657
+[2025-09-04 10:24:44] [Rank 0]   Group 3 Loss: 4.6657
+[2025-09-04 10:24:44] [Rank 0]   Group 4 Loss: 4.6788
+[2025-09-04 10:24:44] [Rank 0]   Group 4 Loss: 4.6788
+[2025-09-04 10:24:44] [Rank 0]   Group 5 Loss: 4.7178
+[2025-09-04 10:24:44] [Rank 0]   Group 5 Loss: 4.7178
+[2025-09-04 10:24:44] [Rank 0]   Group 6 Loss: 4.6162
+[2025-09-04 10:24:44] [Rank 0]   Group 6 Loss: 4.6162
+[2025-09-04 10:24:44] [Rank 0]   Group 7 Loss: 4.6559
+[2025-09-04 10:24:44] [Rank 0]   Group 7 Loss: 4.6559
+[2025-09-04 10:24:44] [Rank 0]   Group 8 Loss: 4.8171
+[2025-09-04 10:24:44] [Rank 0]   Group 8 Loss: 4.8171
+[2025-09-04 10:24:44] [Rank 0]   Group 9 Loss: 4.8639
+[2025-09-04 10:24:44] [Rank 0]   Group 9 Loss: 4.8639
+[2025-09-04 10:24:44] [Rank 0]   Group 10 Loss: 4.9400
+[2025-09-04 10:24:44] [Rank 0]   Group 10 Loss: 4.9400
+[2025-09-04 10:24:44] [Rank 0]   Group 11 Loss: 4.9622
+[2025-09-04 10:24:44] [Rank 0]   Group 11 Loss: 4.9622
+[2025-09-04 10:24:44] [Rank 0]   Group 12 Loss: 4.8809
+[2025-09-04 10:24:44] [Rank 0]   Group 12 Loss: 4.8809
+[2025-09-04 10:24:44] [Rank 0]   Group 13 Loss: 5.0139
+[2025-09-04 10:24:44] [Rank 0]   Group 13 Loss: 5.0139
+[2025-09-04 10:24:44] [Rank 0]   Group 14 Loss: 4.9639
+[2025-09-04 10:24:44] [Rank 0]   Group 14 Loss: 4.9639
+[2025-09-04 10:24:44] [Rank 0]   Group 15 Loss: 5.0470
+[2025-09-04 10:24:44] [Rank 0]   Group 15 Loss: 5.0470
+[2025-09-04 10:24:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:24:44] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 10:24:44] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 10:24:44] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 10:24:44] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 10:24:44] [Rank 0]   Group 12 FTA: 0.9400
+[2025-09-04 10:24:44] [Rank 0]   Group 12 FTA: 0.9400
+[2025-09-04 10:24:44] [Rank 0]   Group 13 FTA: 0.5700
+[2025-09-04 10:24:44] [Rank 0]   Group 13 FTA: 0.5700
+[2025-09-04 10:24:44] [Rank 0]   Group 14 FTA: 0.2200
+[2025-09-04 10:24:44] [Rank 0]   Group 14 FTA: 0.2200
+[2025-09-04 10:24:44] [Rank 0]   Group 15 FTA: 0.1400
+[2025-09-04 10:24:44] [Rank 0]   Group 15 FTA: 0.1400
+[2025-09-04 10:24:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:24:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:24:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:24:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:24:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:24:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:24:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:24:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:24:45] [Rank 0] step:3001/10000 train_time:154911ms step_avg:51.62ms
+[2025-09-04 10:24:45] [Rank 0] step:3001/10000 train_time:154911ms step_avg:51.62ms
+[2025-09-04 10:24:46] [Rank 0] step:3021/10000 train_time:155700ms step_avg:51.54ms
+[2025-09-04 10:24:46] [Rank 0] step:3021/10000 train_time:155700ms step_avg:51.54ms
+[2025-09-04 10:24:47] [Rank 0] step:3041/10000 train_time:156461ms step_avg:51.45ms
+[2025-09-04 10:24:47] [Rank 0] step:3041/10000 train_time:156461ms step_avg:51.45ms
+[2025-09-04 10:24:48] [Rank 0] step:3061/10000 train_time:157223ms step_avg:51.36ms
+[2025-09-04 10:24:48] [Rank 0] step:3061/10000 train_time:157223ms step_avg:51.36ms
+[2025-09-04 10:24:48] [Rank 0] step:3081/10000 train_time:157984ms step_avg:51.28ms
+[2025-09-04 10:24:48] [Rank 0] step:3081/10000 train_time:157984ms step_avg:51.28ms
+[2025-09-04 10:24:49] [Rank 0] step:3101/10000 train_time:158746ms step_avg:51.19ms
+[2025-09-04 10:24:49] [Rank 0] step:3101/10000 train_time:158746ms step_avg:51.19ms
+[2025-09-04 10:24:50] [Rank 0] step:3121/10000 train_time:159508ms step_avg:51.11ms
+[2025-09-04 10:24:50] [Rank 0] step:3121/10000 train_time:159508ms step_avg:51.11ms
+[2025-09-04 10:24:51] [Rank 0] step:3141/10000 train_time:160271ms step_avg:51.03ms
+[2025-09-04 10:24:51] [Rank 0] step:3141/10000 train_time:160271ms step_avg:51.03ms
+[2025-09-04 10:24:51] [Rank 0] step:3161/10000 train_time:161033ms step_avg:50.94ms
+[2025-09-04 10:24:51] [Rank 0] step:3161/10000 train_time:161033ms step_avg:50.94ms
+[2025-09-04 10:24:52] [Rank 0] step:3181/10000 train_time:161795ms step_avg:50.86ms
+[2025-09-04 10:24:52] [Rank 0] step:3181/10000 train_time:161795ms step_avg:50.86ms
+[2025-09-04 10:24:53] [Rank 0] step:3201/10000 train_time:162569ms step_avg:50.79ms
+[2025-09-04 10:24:53] [Rank 0] step:3201/10000 train_time:162569ms step_avg:50.79ms
+[2025-09-04 10:24:54] [Rank 0] step:3221/10000 train_time:163332ms step_avg:50.71ms
+[2025-09-04 10:24:54] [Rank 0] step:3221/10000 train_time:163332ms step_avg:50.71ms
+[2025-09-04 10:24:54] [Rank 0] step:3241/10000 train_time:164095ms step_avg:50.63ms
+[2025-09-04 10:24:54] [Rank 0] step:3241/10000 train_time:164095ms step_avg:50.63ms
+[2025-09-04 10:24:55] [Rank 0] step:3261/10000 train_time:164858ms step_avg:50.55ms
+[2025-09-04 10:24:55] [Rank 0] step:3261/10000 train_time:164858ms step_avg:50.55ms
+[2025-09-04 10:24:56] [Rank 0] step:3281/10000 train_time:165620ms step_avg:50.48ms
+[2025-09-04 10:24:56] [Rank 0] step:3281/10000 train_time:165620ms step_avg:50.48ms
+[2025-09-04 10:24:57] [Rank 0] step:3301/10000 train_time:166382ms step_avg:50.40ms
+[2025-09-04 10:24:57] [Rank 0] step:3301/10000 train_time:166382ms step_avg:50.40ms
+[2025-09-04 10:24:57] [Rank 0] step:3321/10000 train_time:167145ms step_avg:50.33ms
+[2025-09-04 10:24:57] [Rank 0] step:3321/10000 train_time:167145ms step_avg:50.33ms
+[2025-09-04 10:24:58] [Rank 0] step:3341/10000 train_time:167907ms step_avg:50.26ms
+[2025-09-04 10:24:58] [Rank 0] step:3341/10000 train_time:167907ms step_avg:50.26ms
+[2025-09-04 10:24:59] [Rank 0] step:3361/10000 train_time:168669ms step_avg:50.18ms
+[2025-09-04 10:24:59] [Rank 0] step:3361/10000 train_time:168669ms step_avg:50.18ms
+[2025-09-04 10:25:00] [Rank 0] step:3381/10000 train_time:169432ms step_avg:50.11ms
+[2025-09-04 10:25:00] [Rank 0] step:3381/10000 train_time:169432ms step_avg:50.11ms
+[2025-09-04 10:25:01] [Rank 0] step:3401/10000 train_time:170195ms step_avg:50.04ms
+[2025-09-04 10:25:01] [Rank 0] step:3401/10000 train_time:170195ms step_avg:50.04ms
+[2025-09-04 10:25:01] [Rank 0] step:3421/10000 train_time:170988ms step_avg:49.98ms
+[2025-09-04 10:25:01] [Rank 0] step:3421/10000 train_time:170988ms step_avg:49.98ms
+[2025-09-04 10:25:02] [Rank 0] step:3441/10000 train_time:171792ms step_avg:49.92ms
+[2025-09-04 10:25:02] [Rank 0] step:3441/10000 train_time:171792ms step_avg:49.92ms
+[2025-09-04 10:25:03] [Rank 0] step:3461/10000 train_time:172555ms step_avg:49.86ms
+[2025-09-04 10:25:03] [Rank 0] step:3461/10000 train_time:172555ms step_avg:49.86ms
+[2025-09-04 10:25:04] [Rank 0] step:3481/10000 train_time:173317ms step_avg:49.79ms
+[2025-09-04 10:25:04] [Rank 0] step:3481/10000 train_time:173317ms step_avg:49.79ms
+[2025-09-04 10:25:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:25:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:25:05] [Rank 0] PRINT: step:3500/10000 train_loss:0.6883 val_loss:0.6711 train_time:174085ms step_avg:49.74ms
+[2025-09-04 10:25:05] [Rank 0] PRINT: step:3500/10000 train_loss:0.6883 val_loss:0.6711 train_time:174085ms step_avg:49.74ms
+[2025-09-04 10:25:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:25:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:25:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:25:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:26:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:26:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:26:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:26:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:26:43] [Rank 0]   Total Loss: 4.7852
+[2025-09-04 10:26:43] [Rank 0]   Total Loss: 4.7852
+[2025-09-04 10:26:43] [Rank 0]   Total FTA (Unweighted): 0.8781
+[2025-09-04 10:26:43] [Rank 0]   Total FTA (Unweighted): 0.8781
+[2025-09-04 10:26:43] [Rank 0]   Total FTA (Weighted):   0.8781
+[2025-09-04 10:26:43] [Rank 0]   Total FTA (Weighted):   0.8781
+[2025-09-04 10:26:43] [Rank 0]   Group 0 Loss: 4.7889
+[2025-09-04 10:26:43] [Rank 0]   Group 0 Loss: 4.7889
+[2025-09-04 10:26:43] [Rank 0]   Group 1 Loss: 4.3274
+[2025-09-04 10:26:43] [Rank 0]   Group 1 Loss: 4.3274
+[2025-09-04 10:26:43] [Rank 0]   Group 2 Loss: 4.2933
+[2025-09-04 10:26:43] [Rank 0]   Group 2 Loss: 4.2933
+[2025-09-04 10:26:43] [Rank 0]   Group 3 Loss: 4.7572
+[2025-09-04 10:26:43] [Rank 0]   Group 3 Loss: 4.7572
+[2025-09-04 10:26:43] [Rank 0]   Group 4 Loss: 4.6749
+[2025-09-04 10:26:43] [Rank 0]   Group 4 Loss: 4.6749
+[2025-09-04 10:26:43] [Rank 0]   Group 5 Loss: 4.7294
+[2025-09-04 10:26:43] [Rank 0]   Group 5 Loss: 4.7294
+[2025-09-04 10:26:43] [Rank 0]   Group 6 Loss: 4.6169
+[2025-09-04 10:26:43] [Rank 0]   Group 6 Loss: 4.6169
+[2025-09-04 10:26:43] [Rank 0]   Group 7 Loss: 4.7002
+[2025-09-04 10:26:43] [Rank 0]   Group 7 Loss: 4.7002
+[2025-09-04 10:26:43] [Rank 0]   Group 8 Loss: 4.8535
+[2025-09-04 10:26:43] [Rank 0]   Group 8 Loss: 4.8535
+[2025-09-04 10:26:43] [Rank 0]   Group 9 Loss: 4.8998
+[2025-09-04 10:26:43] [Rank 0]   Group 9 Loss: 4.8998
+[2025-09-04 10:26:43] [Rank 0]   Group 10 Loss: 4.9680
+[2025-09-04 10:26:43] [Rank 0]   Group 10 Loss: 4.9680
+[2025-09-04 10:26:43] [Rank 0]   Group 11 Loss: 5.0129
+[2025-09-04 10:26:43] [Rank 0]   Group 11 Loss: 5.0129
+[2025-09-04 10:26:43] [Rank 0]   Group 12 Loss: 4.9183
+[2025-09-04 10:26:43] [Rank 0]   Group 12 Loss: 4.9183
+[2025-09-04 10:26:43] [Rank 0]   Group 13 Loss: 5.0265
+[2025-09-04 10:26:43] [Rank 0]   Group 13 Loss: 5.0265
+[2025-09-04 10:26:43] [Rank 0]   Group 14 Loss: 4.9666
+[2025-09-04 10:26:43] [Rank 0]   Group 14 Loss: 4.9666
+[2025-09-04 10:26:43] [Rank 0]   Group 15 Loss: 5.0297
+[2025-09-04 10:26:43] [Rank 0]   Group 15 Loss: 5.0297
+[2025-09-04 10:26:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:26:43] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 10:26:43] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 10:26:43] [Rank 0]   Group 13 FTA: 0.7000
+[2025-09-04 10:26:43] [Rank 0]   Group 13 FTA: 0.7000
+[2025-09-04 10:26:43] [Rank 0]   Group 14 FTA: 0.2400
+[2025-09-04 10:26:43] [Rank 0]   Group 14 FTA: 0.2400
+[2025-09-04 10:26:43] [Rank 0]   Group 15 FTA: 0.1300
+[2025-09-04 10:26:43] [Rank 0]   Group 15 FTA: 0.1300
+[2025-09-04 10:26:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:26:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:26:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:26:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:26:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:26:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:26:44] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:26:44] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:26:45] [Rank 0] step:3501/10000 train_time:174101ms step_avg:49.73ms
+[2025-09-04 10:26:45] [Rank 0] step:3501/10000 train_time:174101ms step_avg:49.73ms
+[2025-09-04 10:26:45] [Rank 0] step:3521/10000 train_time:174872ms step_avg:49.67ms
+[2025-09-04 10:26:45] [Rank 0] step:3521/10000 train_time:174872ms step_avg:49.67ms
+[2025-09-04 10:26:46] [Rank 0] step:3541/10000 train_time:175638ms step_avg:49.60ms
+[2025-09-04 10:26:46] [Rank 0] step:3541/10000 train_time:175638ms step_avg:49.60ms
+[2025-09-04 10:26:47] [Rank 0] step:3561/10000 train_time:176399ms step_avg:49.54ms
+[2025-09-04 10:26:47] [Rank 0] step:3561/10000 train_time:176399ms step_avg:49.54ms
+[2025-09-04 10:26:48] [Rank 0] step:3581/10000 train_time:177161ms step_avg:49.47ms
+[2025-09-04 10:26:48] [Rank 0] step:3581/10000 train_time:177161ms step_avg:49.47ms
+[2025-09-04 10:26:48] [Rank 0] step:3601/10000 train_time:177923ms step_avg:49.41ms
+[2025-09-04 10:26:48] [Rank 0] step:3601/10000 train_time:177923ms step_avg:49.41ms
+[2025-09-04 10:26:49] [Rank 0] step:3621/10000 train_time:178684ms step_avg:49.35ms
+[2025-09-04 10:26:49] [Rank 0] step:3621/10000 train_time:178684ms step_avg:49.35ms
+[2025-09-04 10:26:50] [Rank 0] step:3641/10000 train_time:179714ms step_avg:49.36ms
+[2025-09-04 10:26:50] [Rank 0] step:3641/10000 train_time:179714ms step_avg:49.36ms
+[2025-09-04 10:26:51] [Rank 0] step:3661/10000 train_time:180476ms step_avg:49.30ms
+[2025-09-04 10:26:51] [Rank 0] step:3661/10000 train_time:180476ms step_avg:49.30ms
+[2025-09-04 10:26:52] [Rank 0] step:3681/10000 train_time:181237ms step_avg:49.24ms
+[2025-09-04 10:26:52] [Rank 0] step:3681/10000 train_time:181237ms step_avg:49.24ms
+[2025-09-04 10:26:52] [Rank 0] step:3701/10000 train_time:181999ms step_avg:49.18ms
+[2025-09-04 10:26:52] [Rank 0] step:3701/10000 train_time:181999ms step_avg:49.18ms
+[2025-09-04 10:26:53] [Rank 0] step:3721/10000 train_time:182760ms step_avg:49.12ms
+[2025-09-04 10:26:53] [Rank 0] step:3721/10000 train_time:182760ms step_avg:49.12ms
+[2025-09-04 10:26:54] [Rank 0] step:3741/10000 train_time:183522ms step_avg:49.06ms
+[2025-09-04 10:26:54] [Rank 0] step:3741/10000 train_time:183522ms step_avg:49.06ms
+[2025-09-04 10:26:55] [Rank 0] step:3761/10000 train_time:184283ms step_avg:49.00ms
+[2025-09-04 10:26:55] [Rank 0] step:3761/10000 train_time:184283ms step_avg:49.00ms
+[2025-09-04 10:26:55] [Rank 0] step:3781/10000 train_time:185045ms step_avg:48.94ms
+[2025-09-04 10:26:55] [Rank 0] step:3781/10000 train_time:185045ms step_avg:48.94ms
+[2025-09-04 10:26:56] [Rank 0] step:3801/10000 train_time:185807ms step_avg:48.88ms
+[2025-09-04 10:26:56] [Rank 0] step:3801/10000 train_time:185807ms step_avg:48.88ms
+[2025-09-04 10:26:57] [Rank 0] step:3821/10000 train_time:186568ms step_avg:48.83ms
+[2025-09-04 10:26:57] [Rank 0] step:3821/10000 train_time:186568ms step_avg:48.83ms
+[2025-09-04 10:26:58] [Rank 0] step:3841/10000 train_time:187330ms step_avg:48.77ms
+[2025-09-04 10:26:58] [Rank 0] step:3841/10000 train_time:187330ms step_avg:48.77ms
+[2025-09-04 10:26:59] [Rank 0] step:3861/10000 train_time:188092ms step_avg:48.72ms
+[2025-09-04 10:26:59] [Rank 0] step:3861/10000 train_time:188092ms step_avg:48.72ms
+[2025-09-04 10:26:59] [Rank 0] step:3881/10000 train_time:188854ms step_avg:48.66ms
+[2025-09-04 10:26:59] [Rank 0] step:3881/10000 train_time:188854ms step_avg:48.66ms
+[2025-09-04 10:27:00] [Rank 0] step:3901/10000 train_time:189620ms step_avg:48.61ms
+[2025-09-04 10:27:00] [Rank 0] step:3901/10000 train_time:189620ms step_avg:48.61ms
+[2025-09-04 10:27:01] [Rank 0] step:3921/10000 train_time:190382ms step_avg:48.55ms
+[2025-09-04 10:27:01] [Rank 0] step:3921/10000 train_time:190382ms step_avg:48.55ms
+[2025-09-04 10:27:02] [Rank 0] step:3941/10000 train_time:191145ms step_avg:48.50ms
+[2025-09-04 10:27:02] [Rank 0] step:3941/10000 train_time:191145ms step_avg:48.50ms
+[2025-09-04 10:27:02] [Rank 0] step:3961/10000 train_time:191907ms step_avg:48.45ms
+[2025-09-04 10:27:02] [Rank 0] step:3961/10000 train_time:191907ms step_avg:48.45ms
+[2025-09-04 10:27:03] [Rank 0] step:3981/10000 train_time:192671ms step_avg:48.40ms
+[2025-09-04 10:27:03] [Rank 0] step:3981/10000 train_time:192671ms step_avg:48.40ms
+[2025-09-04 10:27:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:27:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:27:04] [Rank 0] PRINT: step:4000/10000 train_loss:0.6763 val_loss:0.6599 train_time:193436ms step_avg:48.36ms
+[2025-09-04 10:27:04] [Rank 0] PRINT: step:4000/10000 train_loss:0.6763 val_loss:0.6599 train_time:193436ms step_avg:48.36ms
+[2025-09-04 10:27:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:27:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:27:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:27:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:28:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:28:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:28:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:28:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:28:43] [Rank 0]   Total Loss: 4.8364
+[2025-09-04 10:28:43] [Rank 0]   Total Loss: 4.8364
+[2025-09-04 10:28:43] [Rank 0]   Total FTA (Unweighted): 0.8956
+[2025-09-04 10:28:43] [Rank 0]   Total FTA (Unweighted): 0.8956
+[2025-09-04 10:28:43] [Rank 0]   Total FTA (Weighted):   0.8956
+[2025-09-04 10:28:43] [Rank 0]   Total FTA (Weighted):   0.8956
+[2025-09-04 10:28:43] [Rank 0]   Group 0 Loss: 4.9035
+[2025-09-04 10:28:43] [Rank 0]   Group 0 Loss: 4.9035
+[2025-09-04 10:28:43] [Rank 0]   Group 1 Loss: 4.3696
+[2025-09-04 10:28:43] [Rank 0]   Group 1 Loss: 4.3696
+[2025-09-04 10:28:43] [Rank 0]   Group 2 Loss: 4.3380
+[2025-09-04 10:28:43] [Rank 0]   Group 2 Loss: 4.3380
+[2025-09-04 10:28:43] [Rank 0]   Group 3 Loss: 4.7999
+[2025-09-04 10:28:43] [Rank 0]   Group 3 Loss: 4.7999
+[2025-09-04 10:28:43] [Rank 0]   Group 4 Loss: 4.7195
+[2025-09-04 10:28:43] [Rank 0]   Group 4 Loss: 4.7195
+[2025-09-04 10:28:43] [Rank 0]   Group 5 Loss: 4.8064
+[2025-09-04 10:28:43] [Rank 0]   Group 5 Loss: 4.8064
+[2025-09-04 10:28:43] [Rank 0]   Group 6 Loss: 4.6719
+[2025-09-04 10:28:43] [Rank 0]   Group 6 Loss: 4.6719
+[2025-09-04 10:28:43] [Rank 0]   Group 7 Loss: 4.7718
+[2025-09-04 10:28:43] [Rank 0]   Group 7 Loss: 4.7718
+[2025-09-04 10:28:43] [Rank 0]   Group 8 Loss: 4.9463
+[2025-09-04 10:28:43] [Rank 0]   Group 8 Loss: 4.9463
+[2025-09-04 10:28:43] [Rank 0]   Group 9 Loss: 4.9292
+[2025-09-04 10:28:43] [Rank 0]   Group 9 Loss: 4.9292
+[2025-09-04 10:28:43] [Rank 0]   Group 10 Loss: 4.9999
+[2025-09-04 10:28:43] [Rank 0]   Group 10 Loss: 4.9999
+[2025-09-04 10:28:43] [Rank 0]   Group 11 Loss: 5.0481
+[2025-09-04 10:28:43] [Rank 0]   Group 11 Loss: 5.0481
+[2025-09-04 10:28:43] [Rank 0]   Group 12 Loss: 4.9577
+[2025-09-04 10:28:43] [Rank 0]   Group 12 Loss: 4.9577
+[2025-09-04 10:28:43] [Rank 0]   Group 13 Loss: 5.0701
+[2025-09-04 10:28:43] [Rank 0]   Group 13 Loss: 5.0701
+[2025-09-04 10:28:43] [Rank 0]   Group 14 Loss: 5.0060
+[2025-09-04 10:28:43] [Rank 0]   Group 14 Loss: 5.0060
+[2025-09-04 10:28:43] [Rank 0]   Group 15 Loss: 5.0437
+[2025-09-04 10:28:43] [Rank 0]   Group 15 Loss: 5.0437
+[2025-09-04 10:28:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 10:28:43] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 10:28:43] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:28:43] [Rank 0]   Group 12 FTA: 0.9600
+[2025-09-04 10:28:43] [Rank 0]   Group 12 FTA: 0.9600
+[2025-09-04 10:28:43] [Rank 0]   Group 13 FTA: 0.7800
+[2025-09-04 10:28:43] [Rank 0]   Group 13 FTA: 0.7800
+[2025-09-04 10:28:43] [Rank 0]   Group 14 FTA: 0.4300
+[2025-09-04 10:28:43] [Rank 0]   Group 14 FTA: 0.4300
+[2025-09-04 10:28:43] [Rank 0]   Group 15 FTA: 0.1700
+[2025-09-04 10:28:43] [Rank 0]   Group 15 FTA: 0.1700
+[2025-09-04 10:28:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:28:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:28:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:28:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:28:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:28:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:28:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:28:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:28:45] [Rank 0] step:4001/10000 train_time:193454ms step_avg:48.35ms
+[2025-09-04 10:28:45] [Rank 0] step:4001/10000 train_time:193454ms step_avg:48.35ms
+[2025-09-04 10:28:46] [Rank 0] step:4021/10000 train_time:194492ms step_avg:48.37ms
+[2025-09-04 10:28:46] [Rank 0] step:4021/10000 train_time:194492ms step_avg:48.37ms
+[2025-09-04 10:28:46] [Rank 0] step:4041/10000 train_time:195253ms step_avg:48.32ms
+[2025-09-04 10:28:46] [Rank 0] step:4041/10000 train_time:195253ms step_avg:48.32ms
+[2025-09-04 10:28:47] [Rank 0] step:4061/10000 train_time:196016ms step_avg:48.27ms
+[2025-09-04 10:28:47] [Rank 0] step:4061/10000 train_time:196016ms step_avg:48.27ms
+[2025-09-04 10:28:48] [Rank 0] step:4081/10000 train_time:196781ms step_avg:48.22ms
+[2025-09-04 10:28:48] [Rank 0] step:4081/10000 train_time:196781ms step_avg:48.22ms
+[2025-09-04 10:28:49] [Rank 0] step:4101/10000 train_time:197545ms step_avg:48.17ms
+[2025-09-04 10:28:49] [Rank 0] step:4101/10000 train_time:197545ms step_avg:48.17ms
+[2025-09-04 10:28:49] [Rank 0] step:4121/10000 train_time:198310ms step_avg:48.12ms
+[2025-09-04 10:28:49] [Rank 0] step:4121/10000 train_time:198310ms step_avg:48.12ms
+[2025-09-04 10:28:50] [Rank 0] step:4141/10000 train_time:199072ms step_avg:48.07ms
+[2025-09-04 10:28:50] [Rank 0] step:4141/10000 train_time:199072ms step_avg:48.07ms
+[2025-09-04 10:28:51] [Rank 0] step:4161/10000 train_time:199834ms step_avg:48.03ms
+[2025-09-04 10:28:51] [Rank 0] step:4161/10000 train_time:199834ms step_avg:48.03ms
+[2025-09-04 10:28:52] [Rank 0] step:4181/10000 train_time:200597ms step_avg:47.98ms
+[2025-09-04 10:28:52] [Rank 0] step:4181/10000 train_time:200597ms step_avg:47.98ms
+[2025-09-04 10:28:52] [Rank 0] step:4201/10000 train_time:201357ms step_avg:47.93ms
+[2025-09-04 10:28:52] [Rank 0] step:4201/10000 train_time:201357ms step_avg:47.93ms
+[2025-09-04 10:28:53] [Rank 0] step:4221/10000 train_time:202118ms step_avg:47.88ms
+[2025-09-04 10:28:53] [Rank 0] step:4221/10000 train_time:202118ms step_avg:47.88ms
+[2025-09-04 10:28:54] [Rank 0] step:4241/10000 train_time:202881ms step_avg:47.84ms
+[2025-09-04 10:28:54] [Rank 0] step:4241/10000 train_time:202881ms step_avg:47.84ms
+[2025-09-04 10:28:55] [Rank 0] step:4261/10000 train_time:203644ms step_avg:47.79ms
+[2025-09-04 10:28:55] [Rank 0] step:4261/10000 train_time:203644ms step_avg:47.79ms
+[2025-09-04 10:28:56] [Rank 0] step:4281/10000 train_time:204407ms step_avg:47.75ms
+[2025-09-04 10:28:56] [Rank 0] step:4281/10000 train_time:204407ms step_avg:47.75ms
+[2025-09-04 10:28:56] [Rank 0] step:4301/10000 train_time:205169ms step_avg:47.70ms
+[2025-09-04 10:28:56] [Rank 0] step:4301/10000 train_time:205169ms step_avg:47.70ms
+[2025-09-04 10:28:57] [Rank 0] step:4321/10000 train_time:205933ms step_avg:47.66ms
+[2025-09-04 10:28:57] [Rank 0] step:4321/10000 train_time:205933ms step_avg:47.66ms
+[2025-09-04 10:28:58] [Rank 0] step:4341/10000 train_time:206696ms step_avg:47.61ms
+[2025-09-04 10:28:58] [Rank 0] step:4341/10000 train_time:206696ms step_avg:47.61ms
+[2025-09-04 10:28:59] [Rank 0] step:4361/10000 train_time:207457ms step_avg:47.57ms
+[2025-09-04 10:28:59] [Rank 0] step:4361/10000 train_time:207457ms step_avg:47.57ms
+[2025-09-04 10:28:59] [Rank 0] step:4381/10000 train_time:208219ms step_avg:47.53ms
+[2025-09-04 10:28:59] [Rank 0] step:4381/10000 train_time:208219ms step_avg:47.53ms
+[2025-09-04 10:29:00] [Rank 0] step:4401/10000 train_time:208981ms step_avg:47.48ms
+[2025-09-04 10:29:00] [Rank 0] step:4401/10000 train_time:208981ms step_avg:47.48ms
+[2025-09-04 10:29:01] [Rank 0] step:4421/10000 train_time:209744ms step_avg:47.44ms
+[2025-09-04 10:29:01] [Rank 0] step:4421/10000 train_time:209744ms step_avg:47.44ms
+[2025-09-04 10:29:02] [Rank 0] step:4441/10000 train_time:210506ms step_avg:47.40ms
+[2025-09-04 10:29:02] [Rank 0] step:4441/10000 train_time:210506ms step_avg:47.40ms
+[2025-09-04 10:29:02] [Rank 0] step:4461/10000 train_time:211269ms step_avg:47.36ms
+[2025-09-04 10:29:02] [Rank 0] step:4461/10000 train_time:211269ms step_avg:47.36ms
+[2025-09-04 10:29:03] [Rank 0] step:4481/10000 train_time:212037ms step_avg:47.32ms
+[2025-09-04 10:29:03] [Rank 0] step:4481/10000 train_time:212037ms step_avg:47.32ms
+[2025-09-04 10:29:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:29:04] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:29:04] [Rank 0] PRINT: step:4500/10000 train_loss:0.6663 val_loss:0.6497 train_time:212805ms step_avg:47.29ms
+[2025-09-04 10:29:04] [Rank 0] PRINT: step:4500/10000 train_loss:0.6663 val_loss:0.6497 train_time:212805ms step_avg:47.29ms
+[2025-09-04 10:29:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:29:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:29:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:29:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:30:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:30:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:30:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:30:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:30:44] [Rank 0]   Total Loss: 4.9791
+[2025-09-04 10:30:44] [Rank 0]   Total Loss: 4.9791
+[2025-09-04 10:30:44] [Rank 0]   Total FTA (Unweighted): 0.9119
+[2025-09-04 10:30:44] [Rank 0]   Total FTA (Unweighted): 0.9119
+[2025-09-04 10:30:44] [Rank 0]   Total FTA (Weighted):   0.9119
+[2025-09-04 10:30:44] [Rank 0]   Total FTA (Weighted):   0.9119
+[2025-09-04 10:30:44] [Rank 0]   Group 0 Loss: 5.1075
+[2025-09-04 10:30:44] [Rank 0]   Group 0 Loss: 5.1075
+[2025-09-04 10:30:44] [Rank 0]   Group 1 Loss: 4.5186
+[2025-09-04 10:30:44] [Rank 0]   Group 1 Loss: 4.5186
+[2025-09-04 10:30:44] [Rank 0]   Group 2 Loss: 4.4467
+[2025-09-04 10:30:44] [Rank 0]   Group 2 Loss: 4.4467
+[2025-09-04 10:30:44] [Rank 0]   Group 3 Loss: 4.8867
+[2025-09-04 10:30:44] [Rank 0]   Group 3 Loss: 4.8867
+[2025-09-04 10:30:44] [Rank 0]   Group 4 Loss: 4.8231
+[2025-09-04 10:30:44] [Rank 0]   Group 4 Loss: 4.8231
+[2025-09-04 10:30:44] [Rank 0]   Group 5 Loss: 4.9503
+[2025-09-04 10:30:44] [Rank 0]   Group 5 Loss: 4.9503
+[2025-09-04 10:30:44] [Rank 0]   Group 6 Loss: 4.8320
+[2025-09-04 10:30:44] [Rank 0]   Group 6 Loss: 4.8320
+[2025-09-04 10:30:44] [Rank 0]   Group 7 Loss: 4.9262
+[2025-09-04 10:30:44] [Rank 0]   Group 7 Loss: 4.9262
+[2025-09-04 10:30:44] [Rank 0]   Group 8 Loss: 5.0941
+[2025-09-04 10:30:44] [Rank 0]   Group 8 Loss: 5.0941
+[2025-09-04 10:30:44] [Rank 0]   Group 9 Loss: 5.0633
+[2025-09-04 10:30:44] [Rank 0]   Group 9 Loss: 5.0633
+[2025-09-04 10:30:44] [Rank 0]   Group 10 Loss: 5.1659
+[2025-09-04 10:30:44] [Rank 0]   Group 10 Loss: 5.1659
+[2025-09-04 10:30:44] [Rank 0]   Group 11 Loss: 5.1770
+[2025-09-04 10:30:44] [Rank 0]   Group 11 Loss: 5.1770
+[2025-09-04 10:30:44] [Rank 0]   Group 12 Loss: 5.0614
+[2025-09-04 10:30:44] [Rank 0]   Group 12 Loss: 5.0614
+[2025-09-04 10:30:44] [Rank 0]   Group 13 Loss: 5.2142
+[2025-09-04 10:30:44] [Rank 0]   Group 13 Loss: 5.2142
+[2025-09-04 10:30:44] [Rank 0]   Group 14 Loss: 5.1850
+[2025-09-04 10:30:44] [Rank 0]   Group 14 Loss: 5.1850
+[2025-09-04 10:30:44] [Rank 0]   Group 15 Loss: 5.2132
+[2025-09-04 10:30:44] [Rank 0]   Group 15 Loss: 5.2132
+[2025-09-04 10:30:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 7 FTA: 0.9900
+[2025-09-04 10:30:44] [Rank 0]   Group 7 FTA: 0.9900
+[2025-09-04 10:30:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:30:44] [Rank 0]   Group 13 FTA: 0.9300
+[2025-09-04 10:30:44] [Rank 0]   Group 13 FTA: 0.9300
+[2025-09-04 10:30:44] [Rank 0]   Group 14 FTA: 0.4300
+[2025-09-04 10:30:44] [Rank 0]   Group 14 FTA: 0.4300
+[2025-09-04 10:30:44] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 10:30:44] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 10:30:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:30:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:30:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:30:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:30:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:30:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:30:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:30:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:30:46] [Rank 0] step:4501/10000 train_time:212823ms step_avg:47.28ms
+[2025-09-04 10:30:46] [Rank 0] step:4501/10000 train_time:212823ms step_avg:47.28ms
+[2025-09-04 10:30:46] [Rank 0] step:4521/10000 train_time:213672ms step_avg:47.26ms
+[2025-09-04 10:30:46] [Rank 0] step:4521/10000 train_time:213672ms step_avg:47.26ms
+[2025-09-04 10:30:47] [Rank 0] step:4541/10000 train_time:214434ms step_avg:47.22ms
+[2025-09-04 10:30:47] [Rank 0] step:4541/10000 train_time:214434ms step_avg:47.22ms
+[2025-09-04 10:30:48] [Rank 0] step:4561/10000 train_time:215196ms step_avg:47.18ms
+[2025-09-04 10:30:48] [Rank 0] step:4561/10000 train_time:215196ms step_avg:47.18ms
+[2025-09-04 10:30:49] [Rank 0] step:4581/10000 train_time:215958ms step_avg:47.14ms
+[2025-09-04 10:30:49] [Rank 0] step:4581/10000 train_time:215958ms step_avg:47.14ms
+[2025-09-04 10:30:49] [Rank 0] step:4601/10000 train_time:216720ms step_avg:47.10ms
+[2025-09-04 10:30:49] [Rank 0] step:4601/10000 train_time:216720ms step_avg:47.10ms
+[2025-09-04 10:30:50] [Rank 0] step:4621/10000 train_time:217482ms step_avg:47.06ms
+[2025-09-04 10:30:50] [Rank 0] step:4621/10000 train_time:217482ms step_avg:47.06ms
+[2025-09-04 10:30:51] [Rank 0] step:4641/10000 train_time:218244ms step_avg:47.03ms
+[2025-09-04 10:30:51] [Rank 0] step:4641/10000 train_time:218244ms step_avg:47.03ms
+[2025-09-04 10:30:52] [Rank 0] step:4661/10000 train_time:219005ms step_avg:46.99ms
+[2025-09-04 10:30:52] [Rank 0] step:4661/10000 train_time:219005ms step_avg:46.99ms
+[2025-09-04 10:30:53] [Rank 0] step:4681/10000 train_time:219767ms step_avg:46.95ms
+[2025-09-04 10:30:53] [Rank 0] step:4681/10000 train_time:219767ms step_avg:46.95ms
+[2025-09-04 10:30:53] [Rank 0] step:4701/10000 train_time:220528ms step_avg:46.91ms
+[2025-09-04 10:30:53] [Rank 0] step:4701/10000 train_time:220528ms step_avg:46.91ms
+[2025-09-04 10:30:54] [Rank 0] step:4721/10000 train_time:221291ms step_avg:46.87ms
+[2025-09-04 10:30:54] [Rank 0] step:4721/10000 train_time:221291ms step_avg:46.87ms
+[2025-09-04 10:30:55] [Rank 0] step:4741/10000 train_time:222052ms step_avg:46.84ms
+[2025-09-04 10:30:55] [Rank 0] step:4741/10000 train_time:222052ms step_avg:46.84ms
+[2025-09-04 10:30:56] [Rank 0] step:4761/10000 train_time:222814ms step_avg:46.80ms
+[2025-09-04 10:30:56] [Rank 0] step:4761/10000 train_time:222814ms step_avg:46.80ms
+[2025-09-04 10:30:56] [Rank 0] step:4781/10000 train_time:223575ms step_avg:46.76ms
+[2025-09-04 10:30:56] [Rank 0] step:4781/10000 train_time:223575ms step_avg:46.76ms
+[2025-09-04 10:30:57] [Rank 0] step:4801/10000 train_time:224337ms step_avg:46.73ms
+[2025-09-04 10:30:57] [Rank 0] step:4801/10000 train_time:224337ms step_avg:46.73ms
+[2025-09-04 10:30:58] [Rank 0] step:4821/10000 train_time:225099ms step_avg:46.69ms
+[2025-09-04 10:30:58] [Rank 0] step:4821/10000 train_time:225099ms step_avg:46.69ms
+[2025-09-04 10:30:59] [Rank 0] step:4841/10000 train_time:226170ms step_avg:46.72ms
+[2025-09-04 10:30:59] [Rank 0] step:4841/10000 train_time:226170ms step_avg:46.72ms
+[2025-09-04 10:31:00] [Rank 0] step:4861/10000 train_time:226930ms step_avg:46.68ms
+[2025-09-04 10:31:00] [Rank 0] step:4861/10000 train_time:226930ms step_avg:46.68ms
+[2025-09-04 10:31:00] [Rank 0] step:4881/10000 train_time:227691ms step_avg:46.65ms
+[2025-09-04 10:31:00] [Rank 0] step:4881/10000 train_time:227691ms step_avg:46.65ms
+[2025-09-04 10:31:01] [Rank 0] step:4901/10000 train_time:228453ms step_avg:46.61ms
+[2025-09-04 10:31:01] [Rank 0] step:4901/10000 train_time:228453ms step_avg:46.61ms
+[2025-09-04 10:31:02] [Rank 0] step:4921/10000 train_time:229216ms step_avg:46.58ms
+[2025-09-04 10:31:02] [Rank 0] step:4921/10000 train_time:229216ms step_avg:46.58ms
+[2025-09-04 10:31:03] [Rank 0] step:4941/10000 train_time:229978ms step_avg:46.54ms
+[2025-09-04 10:31:03] [Rank 0] step:4941/10000 train_time:229978ms step_avg:46.54ms
+[2025-09-04 10:31:04] [Rank 0] step:4961/10000 train_time:230741ms step_avg:46.51ms
+[2025-09-04 10:31:04] [Rank 0] step:4961/10000 train_time:230741ms step_avg:46.51ms
+[2025-09-04 10:31:04] [Rank 0] step:4981/10000 train_time:231503ms step_avg:46.48ms
+[2025-09-04 10:31:04] [Rank 0] step:4981/10000 train_time:231503ms step_avg:46.48ms
+[2025-09-04 10:31:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:31:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:31:05] [Rank 0] PRINT: step:5000/10000 train_loss:0.6575 val_loss:0.6425 train_time:232270ms step_avg:46.45ms
+[2025-09-04 10:31:05] [Rank 0] PRINT: step:5000/10000 train_loss:0.6575 val_loss:0.6425 train_time:232270ms step_avg:46.45ms
+[2025-09-04 10:31:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:31:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:31:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:31:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:32:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:32:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:32:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:32:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:32:44] [Rank 0]   Total Loss: 4.9835
+[2025-09-04 10:32:44] [Rank 0]   Total Loss: 4.9835
+[2025-09-04 10:32:44] [Rank 0]   Total FTA (Unweighted): 0.9219
+[2025-09-04 10:32:44] [Rank 0]   Total FTA (Unweighted): 0.9219
+[2025-09-04 10:32:44] [Rank 0]   Total FTA (Weighted):   0.9219
+[2025-09-04 10:32:44] [Rank 0]   Total FTA (Weighted):   0.9219
+[2025-09-04 10:32:44] [Rank 0]   Group 0 Loss: 5.0943
+[2025-09-04 10:32:44] [Rank 0]   Group 0 Loss: 5.0943
+[2025-09-04 10:32:44] [Rank 0]   Group 1 Loss: 4.4923
+[2025-09-04 10:32:44] [Rank 0]   Group 1 Loss: 4.4923
+[2025-09-04 10:32:44] [Rank 0]   Group 2 Loss: 4.4506
+[2025-09-04 10:32:44] [Rank 0]   Group 2 Loss: 4.4506
+[2025-09-04 10:32:44] [Rank 0]   Group 3 Loss: 4.8754
+[2025-09-04 10:32:44] [Rank 0]   Group 3 Loss: 4.8754
+[2025-09-04 10:32:44] [Rank 0]   Group 4 Loss: 4.8692
+[2025-09-04 10:32:44] [Rank 0]   Group 4 Loss: 4.8692
+[2025-09-04 10:32:44] [Rank 0]   Group 5 Loss: 4.9800
+[2025-09-04 10:32:44] [Rank 0]   Group 5 Loss: 4.9800
+[2025-09-04 10:32:44] [Rank 0]   Group 6 Loss: 4.8160
+[2025-09-04 10:32:44] [Rank 0]   Group 6 Loss: 4.8160
+[2025-09-04 10:32:44] [Rank 0]   Group 7 Loss: 4.9182
+[2025-09-04 10:32:44] [Rank 0]   Group 7 Loss: 4.9182
+[2025-09-04 10:32:44] [Rank 0]   Group 8 Loss: 5.0900
+[2025-09-04 10:32:44] [Rank 0]   Group 8 Loss: 5.0900
+[2025-09-04 10:32:44] [Rank 0]   Group 9 Loss: 5.0898
+[2025-09-04 10:32:44] [Rank 0]   Group 9 Loss: 5.0898
+[2025-09-04 10:32:44] [Rank 0]   Group 10 Loss: 5.1483
+[2025-09-04 10:32:44] [Rank 0]   Group 10 Loss: 5.1483
+[2025-09-04 10:32:44] [Rank 0]   Group 11 Loss: 5.1683
+[2025-09-04 10:32:44] [Rank 0]   Group 11 Loss: 5.1683
+[2025-09-04 10:32:44] [Rank 0]   Group 12 Loss: 5.1318
+[2025-09-04 10:32:44] [Rank 0]   Group 12 Loss: 5.1318
+[2025-09-04 10:32:44] [Rank 0]   Group 13 Loss: 5.2410
+[2025-09-04 10:32:44] [Rank 0]   Group 13 Loss: 5.2410
+[2025-09-04 10:32:44] [Rank 0]   Group 14 Loss: 5.1860
+[2025-09-04 10:32:44] [Rank 0]   Group 14 Loss: 5.1860
+[2025-09-04 10:32:44] [Rank 0]   Group 15 Loss: 5.1845
+[2025-09-04 10:32:44] [Rank 0]   Group 15 Loss: 5.1845
+[2025-09-04 10:32:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:32:44] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 10:32:44] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 10:32:44] [Rank 0]   Group 13 FTA: 0.9400
+[2025-09-04 10:32:44] [Rank 0]   Group 13 FTA: 0.9400
+[2025-09-04 10:32:44] [Rank 0]   Group 14 FTA: 0.5700
+[2025-09-04 10:32:44] [Rank 0]   Group 14 FTA: 0.5700
+[2025-09-04 10:32:44] [Rank 0]   Group 15 FTA: 0.2500
+[2025-09-04 10:32:44] [Rank 0]   Group 15 FTA: 0.2500
+[2025-09-04 10:32:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:32:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:32:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:32:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:32:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:32:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:32:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:32:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:32:45] [Rank 0] step:5001/10000 train_time:232285ms step_avg:46.45ms
+[2025-09-04 10:32:45] [Rank 0] step:5001/10000 train_time:232285ms step_avg:46.45ms
+[2025-09-04 10:32:46] [Rank 0] step:5021/10000 train_time:233070ms step_avg:46.42ms
+[2025-09-04 10:32:46] [Rank 0] step:5021/10000 train_time:233070ms step_avg:46.42ms
+[2025-09-04 10:32:47] [Rank 0] step:5041/10000 train_time:233831ms step_avg:46.39ms
+[2025-09-04 10:32:47] [Rank 0] step:5041/10000 train_time:233831ms step_avg:46.39ms
+[2025-09-04 10:32:48] [Rank 0] step:5061/10000 train_time:234592ms step_avg:46.35ms
+[2025-09-04 10:32:48] [Rank 0] step:5061/10000 train_time:234592ms step_avg:46.35ms
+[2025-09-04 10:32:48] [Rank 0] step:5081/10000 train_time:235352ms step_avg:46.32ms
+[2025-09-04 10:32:48] [Rank 0] step:5081/10000 train_time:235352ms step_avg:46.32ms
+[2025-09-04 10:32:49] [Rank 0] step:5101/10000 train_time:236114ms step_avg:46.29ms
+[2025-09-04 10:32:49] [Rank 0] step:5101/10000 train_time:236114ms step_avg:46.29ms
+[2025-09-04 10:32:50] [Rank 0] step:5121/10000 train_time:237138ms step_avg:46.31ms
+[2025-09-04 10:32:50] [Rank 0] step:5121/10000 train_time:237138ms step_avg:46.31ms
+[2025-09-04 10:32:51] [Rank 0] step:5141/10000 train_time:237899ms step_avg:46.27ms
+[2025-09-04 10:32:51] [Rank 0] step:5141/10000 train_time:237899ms step_avg:46.27ms
+[2025-09-04 10:32:52] [Rank 0] step:5161/10000 train_time:238661ms step_avg:46.24ms
+[2025-09-04 10:32:52] [Rank 0] step:5161/10000 train_time:238661ms step_avg:46.24ms
+[2025-09-04 10:32:53] [Rank 0] step:5181/10000 train_time:239675ms step_avg:46.26ms
+[2025-09-04 10:32:53] [Rank 0] step:5181/10000 train_time:239675ms step_avg:46.26ms
+[2025-09-04 10:32:54] [Rank 0] step:5201/10000 train_time:240436ms step_avg:46.23ms
+[2025-09-04 10:32:54] [Rank 0] step:5201/10000 train_time:240436ms step_avg:46.23ms
+[2025-09-04 10:32:54] [Rank 0] step:5221/10000 train_time:241198ms step_avg:46.20ms
+[2025-09-04 10:32:54] [Rank 0] step:5221/10000 train_time:241198ms step_avg:46.20ms
+[2025-09-04 10:32:55] [Rank 0] step:5241/10000 train_time:241960ms step_avg:46.17ms
+[2025-09-04 10:32:55] [Rank 0] step:5241/10000 train_time:241960ms step_avg:46.17ms
+[2025-09-04 10:32:56] [Rank 0] step:5261/10000 train_time:242721ms step_avg:46.14ms
+[2025-09-04 10:32:56] [Rank 0] step:5261/10000 train_time:242721ms step_avg:46.14ms
+[2025-09-04 10:32:57] [Rank 0] step:5281/10000 train_time:243483ms step_avg:46.11ms
+[2025-09-04 10:32:57] [Rank 0] step:5281/10000 train_time:243483ms step_avg:46.11ms
+[2025-09-04 10:32:57] [Rank 0] step:5301/10000 train_time:244244ms step_avg:46.08ms
+[2025-09-04 10:32:57] [Rank 0] step:5301/10000 train_time:244244ms step_avg:46.08ms
+[2025-09-04 10:32:58] [Rank 0] step:5321/10000 train_time:245005ms step_avg:46.04ms
+[2025-09-04 10:32:58] [Rank 0] step:5321/10000 train_time:245005ms step_avg:46.04ms
+[2025-09-04 10:32:59] [Rank 0] step:5341/10000 train_time:245767ms step_avg:46.02ms
+[2025-09-04 10:32:59] [Rank 0] step:5341/10000 train_time:245767ms step_avg:46.02ms
+[2025-09-04 10:33:00] [Rank 0] step:5361/10000 train_time:246528ms step_avg:45.99ms
+[2025-09-04 10:33:00] [Rank 0] step:5361/10000 train_time:246528ms step_avg:45.99ms
+[2025-09-04 10:33:00] [Rank 0] step:5381/10000 train_time:247289ms step_avg:45.96ms
+[2025-09-04 10:33:00] [Rank 0] step:5381/10000 train_time:247289ms step_avg:45.96ms
+[2025-09-04 10:33:01] [Rank 0] step:5401/10000 train_time:248054ms step_avg:45.93ms
+[2025-09-04 10:33:01] [Rank 0] step:5401/10000 train_time:248054ms step_avg:45.93ms
+[2025-09-04 10:33:02] [Rank 0] step:5421/10000 train_time:248815ms step_avg:45.90ms
+[2025-09-04 10:33:02] [Rank 0] step:5421/10000 train_time:248815ms step_avg:45.90ms
+[2025-09-04 10:33:03] [Rank 0] step:5441/10000 train_time:249577ms step_avg:45.87ms
+[2025-09-04 10:33:03] [Rank 0] step:5441/10000 train_time:249577ms step_avg:45.87ms
+[2025-09-04 10:33:03] [Rank 0] step:5461/10000 train_time:250338ms step_avg:45.84ms
+[2025-09-04 10:33:03] [Rank 0] step:5461/10000 train_time:250338ms step_avg:45.84ms
+[2025-09-04 10:33:04] [Rank 0] step:5481/10000 train_time:251100ms step_avg:45.81ms
+[2025-09-04 10:33:04] [Rank 0] step:5481/10000 train_time:251100ms step_avg:45.81ms
+[2025-09-04 10:33:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:33:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:33:05] [Rank 0] PRINT: step:5500/10000 train_loss:0.6501 val_loss:0.6354 train_time:251867ms step_avg:45.79ms
+[2025-09-04 10:33:05] [Rank 0] PRINT: step:5500/10000 train_loss:0.6501 val_loss:0.6354 train_time:251867ms step_avg:45.79ms
+[2025-09-04 10:33:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:33:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:33:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:33:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:34:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:34:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:34:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:34:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:34:44] [Rank 0]   Total Loss: 4.9939
+[2025-09-04 10:34:44] [Rank 0]   Total Loss: 4.9939
+[2025-09-04 10:34:44] [Rank 0]   Total FTA (Unweighted): 0.9350
+[2025-09-04 10:34:44] [Rank 0]   Total FTA (Unweighted): 0.9350
+[2025-09-04 10:34:44] [Rank 0]   Total FTA (Weighted):   0.9350
+[2025-09-04 10:34:44] [Rank 0]   Total FTA (Weighted):   0.9350
+[2025-09-04 10:34:44] [Rank 0]   Group 0 Loss: 5.0532
+[2025-09-04 10:34:44] [Rank 0]   Group 0 Loss: 5.0532
+[2025-09-04 10:34:44] [Rank 0]   Group 1 Loss: 4.6316
+[2025-09-04 10:34:44] [Rank 0]   Group 1 Loss: 4.6316
+[2025-09-04 10:34:44] [Rank 0]   Group 2 Loss: 4.4765
+[2025-09-04 10:34:44] [Rank 0]   Group 2 Loss: 4.4765
+[2025-09-04 10:34:44] [Rank 0]   Group 3 Loss: 4.9092
+[2025-09-04 10:34:44] [Rank 0]   Group 3 Loss: 4.9092
+[2025-09-04 10:34:44] [Rank 0]   Group 4 Loss: 4.8351
+[2025-09-04 10:34:44] [Rank 0]   Group 4 Loss: 4.8351
+[2025-09-04 10:34:44] [Rank 0]   Group 5 Loss: 4.9830
+[2025-09-04 10:34:44] [Rank 0]   Group 5 Loss: 4.9830
+[2025-09-04 10:34:44] [Rank 0]   Group 6 Loss: 4.8158
+[2025-09-04 10:34:44] [Rank 0]   Group 6 Loss: 4.8158
+[2025-09-04 10:34:44] [Rank 0]   Group 7 Loss: 4.9304
+[2025-09-04 10:34:44] [Rank 0]   Group 7 Loss: 4.9304
+[2025-09-04 10:34:44] [Rank 0]   Group 8 Loss: 5.0834
+[2025-09-04 10:34:44] [Rank 0]   Group 8 Loss: 5.0834
+[2025-09-04 10:34:44] [Rank 0]   Group 9 Loss: 5.0729
+[2025-09-04 10:34:44] [Rank 0]   Group 9 Loss: 5.0729
+[2025-09-04 10:34:44] [Rank 0]   Group 10 Loss: 5.2120
+[2025-09-04 10:34:44] [Rank 0]   Group 10 Loss: 5.2120
+[2025-09-04 10:34:44] [Rank 0]   Group 11 Loss: 5.1731
+[2025-09-04 10:34:44] [Rank 0]   Group 11 Loss: 5.1731
+[2025-09-04 10:34:44] [Rank 0]   Group 12 Loss: 5.1229
+[2025-09-04 10:34:44] [Rank 0]   Group 12 Loss: 5.1229
+[2025-09-04 10:34:44] [Rank 0]   Group 13 Loss: 5.2459
+[2025-09-04 10:34:44] [Rank 0]   Group 13 Loss: 5.2459
+[2025-09-04 10:34:44] [Rank 0]   Group 14 Loss: 5.1786
+[2025-09-04 10:34:44] [Rank 0]   Group 14 Loss: 5.1786
+[2025-09-04 10:34:44] [Rank 0]   Group 15 Loss: 5.1790
+[2025-09-04 10:34:44] [Rank 0]   Group 15 Loss: 5.1790
+[2025-09-04 10:34:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:34:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:34:45] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 10:34:45] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 10:34:45] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 10:34:45] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 10:34:45] [Rank 0]   Group 14 FTA: 0.6900
+[2025-09-04 10:34:45] [Rank 0]   Group 14 FTA: 0.6900
+[2025-09-04 10:34:45] [Rank 0]   Group 15 FTA: 0.3100
+[2025-09-04 10:34:45] [Rank 0]   Group 15 FTA: 0.3100
+[2025-09-04 10:34:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:34:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:34:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:34:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:34:46] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:34:46] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:34:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:34:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:34:46] [Rank 0] step:5501/10000 train_time:251883ms step_avg:45.79ms
+[2025-09-04 10:34:46] [Rank 0] step:5501/10000 train_time:251883ms step_avg:45.79ms
+[2025-09-04 10:34:47] [Rank 0] step:5521/10000 train_time:252649ms step_avg:45.76ms
+[2025-09-04 10:34:47] [Rank 0] step:5521/10000 train_time:252649ms step_avg:45.76ms
+[2025-09-04 10:34:48] [Rank 0] step:5541/10000 train_time:253412ms step_avg:45.73ms
+[2025-09-04 10:34:48] [Rank 0] step:5541/10000 train_time:253412ms step_avg:45.73ms
+[2025-09-04 10:34:48] [Rank 0] step:5561/10000 train_time:254174ms step_avg:45.71ms
+[2025-09-04 10:34:48] [Rank 0] step:5561/10000 train_time:254174ms step_avg:45.71ms
+[2025-09-04 10:34:49] [Rank 0] step:5581/10000 train_time:254936ms step_avg:45.68ms
+[2025-09-04 10:34:49] [Rank 0] step:5581/10000 train_time:254936ms step_avg:45.68ms
+[2025-09-04 10:34:50] [Rank 0] step:5601/10000 train_time:255702ms step_avg:45.65ms
+[2025-09-04 10:34:50] [Rank 0] step:5601/10000 train_time:255702ms step_avg:45.65ms
+[2025-09-04 10:34:51] [Rank 0] step:5621/10000 train_time:256468ms step_avg:45.63ms
+[2025-09-04 10:34:51] [Rank 0] step:5621/10000 train_time:256468ms step_avg:45.63ms
+[2025-09-04 10:34:52] [Rank 0] step:5641/10000 train_time:257504ms step_avg:45.65ms
+[2025-09-04 10:34:52] [Rank 0] step:5641/10000 train_time:257504ms step_avg:45.65ms
+[2025-09-04 10:34:52] [Rank 0] step:5661/10000 train_time:258268ms step_avg:45.62ms
+[2025-09-04 10:34:52] [Rank 0] step:5661/10000 train_time:258268ms step_avg:45.62ms
+[2025-09-04 10:34:53] [Rank 0] step:5681/10000 train_time:259030ms step_avg:45.60ms
+[2025-09-04 10:34:53] [Rank 0] step:5681/10000 train_time:259030ms step_avg:45.60ms
+[2025-09-04 10:34:54] [Rank 0] step:5701/10000 train_time:259792ms step_avg:45.57ms
+[2025-09-04 10:34:54] [Rank 0] step:5701/10000 train_time:259792ms step_avg:45.57ms
+[2025-09-04 10:34:55] [Rank 0] step:5721/10000 train_time:260554ms step_avg:45.54ms
+[2025-09-04 10:34:55] [Rank 0] step:5721/10000 train_time:260554ms step_avg:45.54ms
+[2025-09-04 10:34:55] [Rank 0] step:5741/10000 train_time:261318ms step_avg:45.52ms
+[2025-09-04 10:34:55] [Rank 0] step:5741/10000 train_time:261318ms step_avg:45.52ms
+[2025-09-04 10:34:56] [Rank 0] step:5761/10000 train_time:262351ms step_avg:45.54ms
+[2025-09-04 10:34:56] [Rank 0] step:5761/10000 train_time:262351ms step_avg:45.54ms
+[2025-09-04 10:34:57] [Rank 0] step:5781/10000 train_time:263114ms step_avg:45.51ms
+[2025-09-04 10:34:57] [Rank 0] step:5781/10000 train_time:263114ms step_avg:45.51ms
+[2025-09-04 10:34:58] [Rank 0] step:5801/10000 train_time:263876ms step_avg:45.49ms
+[2025-09-04 10:34:58] [Rank 0] step:5801/10000 train_time:263876ms step_avg:45.49ms
+[2025-09-04 10:34:59] [Rank 0] step:5821/10000 train_time:264921ms step_avg:45.51ms
+[2025-09-04 10:34:59] [Rank 0] step:5821/10000 train_time:264921ms step_avg:45.51ms
+[2025-09-04 10:35:00] [Rank 0] step:5841/10000 train_time:265683ms step_avg:45.49ms
+[2025-09-04 10:35:00] [Rank 0] step:5841/10000 train_time:265683ms step_avg:45.49ms
+[2025-09-04 10:35:01] [Rank 0] step:5861/10000 train_time:266448ms step_avg:45.46ms
+[2025-09-04 10:35:01] [Rank 0] step:5861/10000 train_time:266448ms step_avg:45.46ms
+[2025-09-04 10:35:01] [Rank 0] step:5881/10000 train_time:267240ms step_avg:45.44ms
+[2025-09-04 10:35:01] [Rank 0] step:5881/10000 train_time:267240ms step_avg:45.44ms
+[2025-09-04 10:35:02] [Rank 0] step:5901/10000 train_time:268032ms step_avg:45.42ms
+[2025-09-04 10:35:02] [Rank 0] step:5901/10000 train_time:268032ms step_avg:45.42ms
+[2025-09-04 10:35:03] [Rank 0] step:5921/10000 train_time:268794ms step_avg:45.40ms
+[2025-09-04 10:35:03] [Rank 0] step:5921/10000 train_time:268794ms step_avg:45.40ms
+[2025-09-04 10:35:04] [Rank 0] step:5941/10000 train_time:269556ms step_avg:45.37ms
+[2025-09-04 10:35:04] [Rank 0] step:5941/10000 train_time:269556ms step_avg:45.37ms
+[2025-09-04 10:35:04] [Rank 0] step:5961/10000 train_time:270318ms step_avg:45.35ms
+[2025-09-04 10:35:04] [Rank 0] step:5961/10000 train_time:270318ms step_avg:45.35ms
+[2025-09-04 10:35:05] [Rank 0] step:5981/10000 train_time:271079ms step_avg:45.32ms
+[2025-09-04 10:35:05] [Rank 0] step:5981/10000 train_time:271079ms step_avg:45.32ms
+[2025-09-04 10:35:06] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:35:06] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:35:06] [Rank 0] PRINT: step:6000/10000 train_loss:0.6431 val_loss:0.6292 train_time:271847ms step_avg:45.31ms
+[2025-09-04 10:35:06] [Rank 0] PRINT: step:6000/10000 train_loss:0.6431 val_loss:0.6292 train_time:271847ms step_avg:45.31ms
+[2025-09-04 10:35:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:35:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:35:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:35:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:36:45] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:36:45] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:36:45] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:36:45] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:36:45] [Rank 0]   Total Loss: 4.9600
+[2025-09-04 10:36:45] [Rank 0]   Total Loss: 4.9600
+[2025-09-04 10:36:45] [Rank 0]   Total FTA (Unweighted): 0.9469
+[2025-09-04 10:36:45] [Rank 0]   Total FTA (Unweighted): 0.9469
+[2025-09-04 10:36:45] [Rank 0]   Total FTA (Weighted):   0.9469
+[2025-09-04 10:36:45] [Rank 0]   Total FTA (Weighted):   0.9469
+[2025-09-04 10:36:45] [Rank 0]   Group 0 Loss: 5.0226
+[2025-09-04 10:36:45] [Rank 0]   Group 0 Loss: 5.0226
+[2025-09-04 10:36:45] [Rank 0]   Group 1 Loss: 4.4677
+[2025-09-04 10:36:45] [Rank 0]   Group 1 Loss: 4.4677
+[2025-09-04 10:36:45] [Rank 0]   Group 2 Loss: 4.4235
+[2025-09-04 10:36:45] [Rank 0]   Group 2 Loss: 4.4235
+[2025-09-04 10:36:45] [Rank 0]   Group 3 Loss: 4.9067
+[2025-09-04 10:36:45] [Rank 0]   Group 3 Loss: 4.9067
+[2025-09-04 10:36:45] [Rank 0]   Group 4 Loss: 4.7773
+[2025-09-04 10:36:45] [Rank 0]   Group 4 Loss: 4.7773
+[2025-09-04 10:36:45] [Rank 0]   Group 5 Loss: 4.9617
+[2025-09-04 10:36:45] [Rank 0]   Group 5 Loss: 4.9617
+[2025-09-04 10:36:45] [Rank 0]   Group 6 Loss: 4.8046
+[2025-09-04 10:36:45] [Rank 0]   Group 6 Loss: 4.8046
+[2025-09-04 10:36:45] [Rank 0]   Group 7 Loss: 4.8862
+[2025-09-04 10:36:45] [Rank 0]   Group 7 Loss: 4.8862
+[2025-09-04 10:36:45] [Rank 0]   Group 8 Loss: 5.0820
+[2025-09-04 10:36:45] [Rank 0]   Group 8 Loss: 5.0820
+[2025-09-04 10:36:45] [Rank 0]   Group 9 Loss: 5.0511
+[2025-09-04 10:36:45] [Rank 0]   Group 9 Loss: 5.0511
+[2025-09-04 10:36:45] [Rank 0]   Group 10 Loss: 5.1828
+[2025-09-04 10:36:45] [Rank 0]   Group 10 Loss: 5.1828
+[2025-09-04 10:36:45] [Rank 0]   Group 11 Loss: 5.1616
+[2025-09-04 10:36:45] [Rank 0]   Group 11 Loss: 5.1616
+[2025-09-04 10:36:45] [Rank 0]   Group 12 Loss: 5.0862
+[2025-09-04 10:36:45] [Rank 0]   Group 12 Loss: 5.0862
+[2025-09-04 10:36:45] [Rank 0]   Group 13 Loss: 5.2277
+[2025-09-04 10:36:45] [Rank 0]   Group 13 Loss: 5.2277
+[2025-09-04 10:36:45] [Rank 0]   Group 14 Loss: 5.1616
+[2025-09-04 10:36:45] [Rank 0]   Group 14 Loss: 5.1616
+[2025-09-04 10:36:45] [Rank 0]   Group 15 Loss: 5.1570
+[2025-09-04 10:36:45] [Rank 0]   Group 15 Loss: 5.1570
+[2025-09-04 10:36:45] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:36:45] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 10:36:46] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 10:36:46] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:36:46] [Rank 0]   Group 13 FTA: 0.9700
+[2025-09-04 10:36:46] [Rank 0]   Group 13 FTA: 0.9700
+[2025-09-04 10:36:46] [Rank 0]   Group 14 FTA: 0.7200
+[2025-09-04 10:36:46] [Rank 0]   Group 14 FTA: 0.7200
+[2025-09-04 10:36:46] [Rank 0]   Group 15 FTA: 0.4700
+[2025-09-04 10:36:46] [Rank 0]   Group 15 FTA: 0.4700
+[2025-09-04 10:36:46] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:36:46] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:36:46] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:36:46] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:36:47] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:36:47] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:36:47] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:36:47] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:36:47] [Rank 0] step:6001/10000 train_time:271864ms step_avg:45.30ms
+[2025-09-04 10:36:47] [Rank 0] step:6001/10000 train_time:271864ms step_avg:45.30ms
+[2025-09-04 10:36:48] [Rank 0] step:6021/10000 train_time:272908ms step_avg:45.33ms
+[2025-09-04 10:36:48] [Rank 0] step:6021/10000 train_time:272908ms step_avg:45.33ms
+[2025-09-04 10:36:49] [Rank 0] step:6041/10000 train_time:273670ms step_avg:45.30ms
+[2025-09-04 10:36:49] [Rank 0] step:6041/10000 train_time:273670ms step_avg:45.30ms
+[2025-09-04 10:36:50] [Rank 0] step:6061/10000 train_time:274431ms step_avg:45.28ms
+[2025-09-04 10:36:50] [Rank 0] step:6061/10000 train_time:274431ms step_avg:45.28ms
+[2025-09-04 10:36:50] [Rank 0] step:6081/10000 train_time:275192ms step_avg:45.25ms
+[2025-09-04 10:36:50] [Rank 0] step:6081/10000 train_time:275192ms step_avg:45.25ms
+[2025-09-04 10:36:51] [Rank 0] step:6101/10000 train_time:275954ms step_avg:45.23ms
+[2025-09-04 10:36:51] [Rank 0] step:6101/10000 train_time:275954ms step_avg:45.23ms
+[2025-09-04 10:36:52] [Rank 0] step:6121/10000 train_time:276715ms step_avg:45.21ms
+[2025-09-04 10:36:52] [Rank 0] step:6121/10000 train_time:276715ms step_avg:45.21ms
+[2025-09-04 10:36:53] [Rank 0] step:6141/10000 train_time:277476ms step_avg:45.18ms
+[2025-09-04 10:36:53] [Rank 0] step:6141/10000 train_time:277476ms step_avg:45.18ms
+[2025-09-04 10:36:53] [Rank 0] step:6161/10000 train_time:278238ms step_avg:45.16ms
+[2025-09-04 10:36:53] [Rank 0] step:6161/10000 train_time:278238ms step_avg:45.16ms
+[2025-09-04 10:36:54] [Rank 0] step:6181/10000 train_time:279000ms step_avg:45.14ms
+[2025-09-04 10:36:54] [Rank 0] step:6181/10000 train_time:279000ms step_avg:45.14ms
+[2025-09-04 10:36:55] [Rank 0] step:6201/10000 train_time:279761ms step_avg:45.12ms
+[2025-09-04 10:36:55] [Rank 0] step:6201/10000 train_time:279761ms step_avg:45.12ms
+[2025-09-04 10:36:56] [Rank 0] step:6221/10000 train_time:280525ms step_avg:45.09ms
+[2025-09-04 10:36:56] [Rank 0] step:6221/10000 train_time:280525ms step_avg:45.09ms
+[2025-09-04 10:36:56] [Rank 0] step:6241/10000 train_time:281284ms step_avg:45.07ms
+[2025-09-04 10:36:56] [Rank 0] step:6241/10000 train_time:281284ms step_avg:45.07ms
+[2025-09-04 10:36:57] [Rank 0] step:6261/10000 train_time:282046ms step_avg:45.05ms
+[2025-09-04 10:36:57] [Rank 0] step:6261/10000 train_time:282046ms step_avg:45.05ms
+[2025-09-04 10:36:58] [Rank 0] step:6281/10000 train_time:282807ms step_avg:45.03ms
+[2025-09-04 10:36:58] [Rank 0] step:6281/10000 train_time:282807ms step_avg:45.03ms
+[2025-09-04 10:36:59] [Rank 0] step:6301/10000 train_time:283568ms step_avg:45.00ms
+[2025-09-04 10:36:59] [Rank 0] step:6301/10000 train_time:283568ms step_avg:45.00ms
+[2025-09-04 10:37:00] [Rank 0] step:6321/10000 train_time:284330ms step_avg:44.98ms
+[2025-09-04 10:37:00] [Rank 0] step:6321/10000 train_time:284330ms step_avg:44.98ms
+[2025-09-04 10:37:00] [Rank 0] step:6341/10000 train_time:285091ms step_avg:44.96ms
+[2025-09-04 10:37:00] [Rank 0] step:6341/10000 train_time:285091ms step_avg:44.96ms
+[2025-09-04 10:37:01] [Rank 0] step:6361/10000 train_time:285852ms step_avg:44.94ms
+[2025-09-04 10:37:01] [Rank 0] step:6361/10000 train_time:285852ms step_avg:44.94ms
+[2025-09-04 10:37:02] [Rank 0] step:6381/10000 train_time:286614ms step_avg:44.92ms
+[2025-09-04 10:37:02] [Rank 0] step:6381/10000 train_time:286614ms step_avg:44.92ms
+[2025-09-04 10:37:03] [Rank 0] step:6401/10000 train_time:287376ms step_avg:44.90ms
+[2025-09-04 10:37:03] [Rank 0] step:6401/10000 train_time:287376ms step_avg:44.90ms
+[2025-09-04 10:37:04] [Rank 0] step:6421/10000 train_time:288330ms step_avg:44.90ms
+[2025-09-04 10:37:04] [Rank 0] step:6421/10000 train_time:288330ms step_avg:44.90ms
+[2025-09-04 10:37:04] [Rank 0] step:6441/10000 train_time:289093ms step_avg:44.88ms
+[2025-09-04 10:37:04] [Rank 0] step:6441/10000 train_time:289093ms step_avg:44.88ms
+[2025-09-04 10:37:05] [Rank 0] step:6461/10000 train_time:289853ms step_avg:44.86ms
+[2025-09-04 10:37:05] [Rank 0] step:6461/10000 train_time:289853ms step_avg:44.86ms
+[2025-09-04 10:37:06] [Rank 0] step:6481/10000 train_time:290898ms step_avg:44.88ms
+[2025-09-04 10:37:06] [Rank 0] step:6481/10000 train_time:290898ms step_avg:44.88ms
+[2025-09-04 10:37:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:37:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:37:07] [Rank 0] PRINT: step:6500/10000 train_loss:0.6369 val_loss:0.6237 train_time:291664ms step_avg:44.87ms
+[2025-09-04 10:37:07] [Rank 0] PRINT: step:6500/10000 train_loss:0.6369 val_loss:0.6237 train_time:291664ms step_avg:44.87ms
+[2025-09-04 10:37:07] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:37:07] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:37:07] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:37:07] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:38:46] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:38:46] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:38:46] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:38:46] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:38:46] [Rank 0]   Total Loss: 4.9646
+[2025-09-04 10:38:46] [Rank 0]   Total Loss: 4.9646
+[2025-09-04 10:38:46] [Rank 0]   Total FTA (Unweighted): 0.9587
+[2025-09-04 10:38:46] [Rank 0]   Total FTA (Unweighted): 0.9587
+[2025-09-04 10:38:46] [Rank 0]   Total FTA (Weighted):   0.9587
+[2025-09-04 10:38:46] [Rank 0]   Total FTA (Weighted):   0.9587
+[2025-09-04 10:38:46] [Rank 0]   Group 0 Loss: 5.0378
+[2025-09-04 10:38:46] [Rank 0]   Group 0 Loss: 5.0378
+[2025-09-04 10:38:46] [Rank 0]   Group 1 Loss: 4.5057
+[2025-09-04 10:38:46] [Rank 0]   Group 1 Loss: 4.5057
+[2025-09-04 10:38:46] [Rank 0]   Group 2 Loss: 4.3917
+[2025-09-04 10:38:46] [Rank 0]   Group 2 Loss: 4.3917
+[2025-09-04 10:38:46] [Rank 0]   Group 3 Loss: 4.8781
+[2025-09-04 10:38:46] [Rank 0]   Group 3 Loss: 4.8781
+[2025-09-04 10:38:46] [Rank 0]   Group 4 Loss: 4.8279
+[2025-09-04 10:38:46] [Rank 0]   Group 4 Loss: 4.8279
+[2025-09-04 10:38:46] [Rank 0]   Group 5 Loss: 4.9355
+[2025-09-04 10:38:46] [Rank 0]   Group 5 Loss: 4.9355
+[2025-09-04 10:38:46] [Rank 0]   Group 6 Loss: 4.7916
+[2025-09-04 10:38:46] [Rank 0]   Group 6 Loss: 4.7916
+[2025-09-04 10:38:46] [Rank 0]   Group 7 Loss: 4.9236
+[2025-09-04 10:38:46] [Rank 0]   Group 7 Loss: 4.9236
+[2025-09-04 10:38:46] [Rank 0]   Group 8 Loss: 5.0814
+[2025-09-04 10:38:46] [Rank 0]   Group 8 Loss: 5.0814
+[2025-09-04 10:38:46] [Rank 0]   Group 9 Loss: 5.0466
+[2025-09-04 10:38:46] [Rank 0]   Group 9 Loss: 5.0466
+[2025-09-04 10:38:46] [Rank 0]   Group 10 Loss: 5.1740
+[2025-09-04 10:38:46] [Rank 0]   Group 10 Loss: 5.1740
+[2025-09-04 10:38:46] [Rank 0]   Group 11 Loss: 5.1584
+[2025-09-04 10:38:46] [Rank 0]   Group 11 Loss: 5.1584
+[2025-09-04 10:38:46] [Rank 0]   Group 12 Loss: 5.0870
+[2025-09-04 10:38:46] [Rank 0]   Group 12 Loss: 5.0870
+[2025-09-04 10:38:46] [Rank 0]   Group 13 Loss: 5.2470
+[2025-09-04 10:38:46] [Rank 0]   Group 13 Loss: 5.2470
+[2025-09-04 10:38:46] [Rank 0]   Group 14 Loss: 5.1694
+[2025-09-04 10:38:46] [Rank 0]   Group 14 Loss: 5.1694
+[2025-09-04 10:38:46] [Rank 0]   Group 15 Loss: 5.1771
+[2025-09-04 10:38:46] [Rank 0]   Group 15 Loss: 5.1771
+[2025-09-04 10:38:46] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:38:46] [Rank 0]   Group 14 FTA: 0.8000
+[2025-09-04 10:38:46] [Rank 0]   Group 14 FTA: 0.8000
+[2025-09-04 10:38:46] [Rank 0]   Group 15 FTA: 0.5400
+[2025-09-04 10:38:46] [Rank 0]   Group 15 FTA: 0.5400
+[2025-09-04 10:38:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:38:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:38:47] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:38:47] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:38:47] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:38:47] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:38:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:38:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:38:48] [Rank 0] step:6501/10000 train_time:291681ms step_avg:44.87ms
+[2025-09-04 10:38:48] [Rank 0] step:6501/10000 train_time:291681ms step_avg:44.87ms
+[2025-09-04 10:38:48] [Rank 0] step:6521/10000 train_time:292465ms step_avg:44.85ms
+[2025-09-04 10:38:48] [Rank 0] step:6521/10000 train_time:292465ms step_avg:44.85ms
+[2025-09-04 10:38:49] [Rank 0] step:6541/10000 train_time:293228ms step_avg:44.83ms
+[2025-09-04 10:38:49] [Rank 0] step:6541/10000 train_time:293228ms step_avg:44.83ms
+[2025-09-04 10:38:50] [Rank 0] step:6561/10000 train_time:293990ms step_avg:44.81ms
+[2025-09-04 10:38:50] [Rank 0] step:6561/10000 train_time:293990ms step_avg:44.81ms
+[2025-09-04 10:38:51] [Rank 0] step:6581/10000 train_time:294752ms step_avg:44.79ms
+[2025-09-04 10:38:51] [Rank 0] step:6581/10000 train_time:294752ms step_avg:44.79ms
+[2025-09-04 10:38:51] [Rank 0] step:6601/10000 train_time:295516ms step_avg:44.77ms
+[2025-09-04 10:38:51] [Rank 0] step:6601/10000 train_time:295516ms step_avg:44.77ms
+[2025-09-04 10:38:52] [Rank 0] step:6621/10000 train_time:296281ms step_avg:44.75ms
+[2025-09-04 10:38:52] [Rank 0] step:6621/10000 train_time:296281ms step_avg:44.75ms
+[2025-09-04 10:38:53] [Rank 0] step:6641/10000 train_time:297043ms step_avg:44.73ms
+[2025-09-04 10:38:53] [Rank 0] step:6641/10000 train_time:297043ms step_avg:44.73ms
+[2025-09-04 10:38:54] [Rank 0] step:6661/10000 train_time:297809ms step_avg:44.71ms
+[2025-09-04 10:38:54] [Rank 0] step:6661/10000 train_time:297809ms step_avg:44.71ms
+[2025-09-04 10:38:55] [Rank 0] step:6681/10000 train_time:298571ms step_avg:44.69ms
+[2025-09-04 10:38:55] [Rank 0] step:6681/10000 train_time:298571ms step_avg:44.69ms
+[2025-09-04 10:38:55] [Rank 0] step:6701/10000 train_time:299333ms step_avg:44.67ms
+[2025-09-04 10:38:55] [Rank 0] step:6701/10000 train_time:299333ms step_avg:44.67ms
+[2025-09-04 10:38:56] [Rank 0] step:6721/10000 train_time:300094ms step_avg:44.65ms
+[2025-09-04 10:38:56] [Rank 0] step:6721/10000 train_time:300094ms step_avg:44.65ms
+[2025-09-04 10:38:57] [Rank 0] step:6741/10000 train_time:300857ms step_avg:44.63ms
+[2025-09-04 10:38:57] [Rank 0] step:6741/10000 train_time:300857ms step_avg:44.63ms
+[2025-09-04 10:38:58] [Rank 0] step:6761/10000 train_time:301619ms step_avg:44.61ms
+[2025-09-04 10:38:58] [Rank 0] step:6761/10000 train_time:301619ms step_avg:44.61ms
+[2025-09-04 10:38:58] [Rank 0] step:6781/10000 train_time:302382ms step_avg:44.59ms
+[2025-09-04 10:38:58] [Rank 0] step:6781/10000 train_time:302382ms step_avg:44.59ms
+[2025-09-04 10:38:59] [Rank 0] step:6801/10000 train_time:303145ms step_avg:44.57ms
+[2025-09-04 10:38:59] [Rank 0] step:6801/10000 train_time:303145ms step_avg:44.57ms
+[2025-09-04 10:39:00] [Rank 0] step:6821/10000 train_time:303909ms step_avg:44.55ms
+[2025-09-04 10:39:00] [Rank 0] step:6821/10000 train_time:303909ms step_avg:44.55ms
+[2025-09-04 10:39:01] [Rank 0] step:6841/10000 train_time:304945ms step_avg:44.58ms
+[2025-09-04 10:39:01] [Rank 0] step:6841/10000 train_time:304945ms step_avg:44.58ms
+[2025-09-04 10:39:02] [Rank 0] step:6861/10000 train_time:305709ms step_avg:44.56ms
+[2025-09-04 10:39:02] [Rank 0] step:6861/10000 train_time:305709ms step_avg:44.56ms
+[2025-09-04 10:39:02] [Rank 0] step:6881/10000 train_time:306472ms step_avg:44.54ms
+[2025-09-04 10:39:02] [Rank 0] step:6881/10000 train_time:306472ms step_avg:44.54ms
+[2025-09-04 10:39:03] [Rank 0] step:6901/10000 train_time:307235ms step_avg:44.52ms
+[2025-09-04 10:39:03] [Rank 0] step:6901/10000 train_time:307235ms step_avg:44.52ms
+[2025-09-04 10:39:04] [Rank 0] step:6921/10000 train_time:307997ms step_avg:44.50ms
+[2025-09-04 10:39:04] [Rank 0] step:6921/10000 train_time:307997ms step_avg:44.50ms
+[2025-09-04 10:39:05] [Rank 0] step:6941/10000 train_time:308759ms step_avg:44.48ms
+[2025-09-04 10:39:05] [Rank 0] step:6941/10000 train_time:308759ms step_avg:44.48ms
+[2025-09-04 10:39:05] [Rank 0] step:6961/10000 train_time:309521ms step_avg:44.47ms
+[2025-09-04 10:39:05] [Rank 0] step:6961/10000 train_time:309521ms step_avg:44.47ms
+[2025-09-04 10:39:06] [Rank 0] step:6981/10000 train_time:310283ms step_avg:44.45ms
+[2025-09-04 10:39:06] [Rank 0] step:6981/10000 train_time:310283ms step_avg:44.45ms
+[2025-09-04 10:39:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:39:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:39:07] [Rank 0] PRINT: step:7000/10000 train_loss:0.6308 val_loss:0.6194 train_time:311050ms step_avg:44.44ms
+[2025-09-04 10:39:07] [Rank 0] PRINT: step:7000/10000 train_loss:0.6308 val_loss:0.6194 train_time:311050ms step_avg:44.44ms
+[2025-09-04 10:39:07] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:39:07] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:39:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:39:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:40:46] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:40:46] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:40:46] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:40:46] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:40:46] [Rank 0]   Total Loss: 5.0290
+[2025-09-04 10:40:46] [Rank 0]   Total Loss: 5.0290
+[2025-09-04 10:40:46] [Rank 0]   Total FTA (Unweighted): 0.9587
+[2025-09-04 10:40:46] [Rank 0]   Total FTA (Unweighted): 0.9587
+[2025-09-04 10:40:46] [Rank 0]   Total FTA (Weighted):   0.9587
+[2025-09-04 10:40:46] [Rank 0]   Total FTA (Weighted):   0.9587
+[2025-09-04 10:40:46] [Rank 0]   Group 0 Loss: 5.1007
+[2025-09-04 10:40:46] [Rank 0]   Group 0 Loss: 5.1007
+[2025-09-04 10:40:46] [Rank 0]   Group 1 Loss: 4.5443
+[2025-09-04 10:40:46] [Rank 0]   Group 1 Loss: 4.5443
+[2025-09-04 10:40:46] [Rank 0]   Group 2 Loss: 4.4441
+[2025-09-04 10:40:46] [Rank 0]   Group 2 Loss: 4.4441
+[2025-09-04 10:40:46] [Rank 0]   Group 3 Loss: 4.9156
+[2025-09-04 10:40:46] [Rank 0]   Group 3 Loss: 4.9156
+[2025-09-04 10:40:46] [Rank 0]   Group 4 Loss: 4.8964
+[2025-09-04 10:40:46] [Rank 0]   Group 4 Loss: 4.8964
+[2025-09-04 10:40:46] [Rank 0]   Group 5 Loss: 4.9970
+[2025-09-04 10:40:46] [Rank 0]   Group 5 Loss: 4.9970
+[2025-09-04 10:40:46] [Rank 0]   Group 6 Loss: 4.9072
+[2025-09-04 10:40:46] [Rank 0]   Group 6 Loss: 4.9072
+[2025-09-04 10:40:46] [Rank 0]   Group 7 Loss: 4.9689
+[2025-09-04 10:40:46] [Rank 0]   Group 7 Loss: 4.9689
+[2025-09-04 10:40:46] [Rank 0]   Group 8 Loss: 5.1757
+[2025-09-04 10:40:46] [Rank 0]   Group 8 Loss: 5.1757
+[2025-09-04 10:40:46] [Rank 0]   Group 9 Loss: 5.1393
+[2025-09-04 10:40:46] [Rank 0]   Group 9 Loss: 5.1393
+[2025-09-04 10:40:46] [Rank 0]   Group 10 Loss: 5.2409
+[2025-09-04 10:40:46] [Rank 0]   Group 10 Loss: 5.2409
+[2025-09-04 10:40:46] [Rank 0]   Group 11 Loss: 5.2360
+[2025-09-04 10:40:46] [Rank 0]   Group 11 Loss: 5.2360
+[2025-09-04 10:40:46] [Rank 0]   Group 12 Loss: 5.1430
+[2025-09-04 10:40:46] [Rank 0]   Group 12 Loss: 5.1430
+[2025-09-04 10:40:46] [Rank 0]   Group 13 Loss: 5.3258
+[2025-09-04 10:40:46] [Rank 0]   Group 13 Loss: 5.3258
+[2025-09-04 10:40:46] [Rank 0]   Group 14 Loss: 5.2342
+[2025-09-04 10:40:46] [Rank 0]   Group 14 Loss: 5.2342
+[2025-09-04 10:40:47] [Rank 0]   Group 15 Loss: 5.1947
+[2025-09-04 10:40:47] [Rank 0]   Group 15 Loss: 5.1947
+[2025-09-04 10:40:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:40:47] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:40:47] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:40:47] [Rank 0]   Group 14 FTA: 0.8300
+[2025-09-04 10:40:47] [Rank 0]   Group 14 FTA: 0.8300
+[2025-09-04 10:40:47] [Rank 0]   Group 15 FTA: 0.5200
+[2025-09-04 10:40:47] [Rank 0]   Group 15 FTA: 0.5200
+[2025-09-04 10:40:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:40:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:40:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:40:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:40:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:40:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:40:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:40:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:40:48] [Rank 0] step:7001/10000 train_time:311066ms step_avg:44.43ms
+[2025-09-04 10:40:48] [Rank 0] step:7001/10000 train_time:311066ms step_avg:44.43ms
+[2025-09-04 10:40:49] [Rank 0] step:7021/10000 train_time:311835ms step_avg:44.41ms
+[2025-09-04 10:40:49] [Rank 0] step:7021/10000 train_time:311835ms step_avg:44.41ms
+[2025-09-04 10:40:50] [Rank 0] step:7041/10000 train_time:312596ms step_avg:44.40ms
+[2025-09-04 10:40:50] [Rank 0] step:7041/10000 train_time:312596ms step_avg:44.40ms
+[2025-09-04 10:40:50] [Rank 0] step:7061/10000 train_time:313357ms step_avg:44.38ms
+[2025-09-04 10:40:50] [Rank 0] step:7061/10000 train_time:313357ms step_avg:44.38ms
+[2025-09-04 10:40:51] [Rank 0] step:7081/10000 train_time:314119ms step_avg:44.36ms
+[2025-09-04 10:40:51] [Rank 0] step:7081/10000 train_time:314119ms step_avg:44.36ms
+[2025-09-04 10:40:52] [Rank 0] step:7101/10000 train_time:314880ms step_avg:44.34ms
+[2025-09-04 10:40:52] [Rank 0] step:7101/10000 train_time:314880ms step_avg:44.34ms
+[2025-09-04 10:40:53] [Rank 0] step:7121/10000 train_time:315641ms step_avg:44.33ms
+[2025-09-04 10:40:53] [Rank 0] step:7121/10000 train_time:315641ms step_avg:44.33ms
+[2025-09-04 10:40:53] [Rank 0] step:7141/10000 train_time:316403ms step_avg:44.31ms
+[2025-09-04 10:40:53] [Rank 0] step:7141/10000 train_time:316403ms step_avg:44.31ms
+[2025-09-04 10:40:54] [Rank 0] step:7161/10000 train_time:317164ms step_avg:44.29ms
+[2025-09-04 10:40:54] [Rank 0] step:7161/10000 train_time:317164ms step_avg:44.29ms
+[2025-09-04 10:40:55] [Rank 0] step:7181/10000 train_time:317926ms step_avg:44.27ms
+[2025-09-04 10:40:55] [Rank 0] step:7181/10000 train_time:317926ms step_avg:44.27ms
+[2025-09-04 10:40:56] [Rank 0] step:7201/10000 train_time:318687ms step_avg:44.26ms
+[2025-09-04 10:40:56] [Rank 0] step:7201/10000 train_time:318687ms step_avg:44.26ms
+[2025-09-04 10:40:57] [Rank 0] step:7221/10000 train_time:319449ms step_avg:44.24ms
+[2025-09-04 10:40:57] [Rank 0] step:7221/10000 train_time:319449ms step_avg:44.24ms
+[2025-09-04 10:40:57] [Rank 0] step:7241/10000 train_time:320210ms step_avg:44.22ms
+[2025-09-04 10:40:57] [Rank 0] step:7241/10000 train_time:320210ms step_avg:44.22ms
+[2025-09-04 10:40:58] [Rank 0] step:7261/10000 train_time:320972ms step_avg:44.20ms
+[2025-09-04 10:40:58] [Rank 0] step:7261/10000 train_time:320972ms step_avg:44.20ms
+[2025-09-04 10:40:59] [Rank 0] step:7281/10000 train_time:321733ms step_avg:44.19ms
+[2025-09-04 10:40:59] [Rank 0] step:7281/10000 train_time:321733ms step_avg:44.19ms
+[2025-09-04 10:41:00] [Rank 0] step:7301/10000 train_time:322495ms step_avg:44.17ms
+[2025-09-04 10:41:00] [Rank 0] step:7301/10000 train_time:322495ms step_avg:44.17ms
+[2025-09-04 10:41:00] [Rank 0] step:7321/10000 train_time:323257ms step_avg:44.15ms
+[2025-09-04 10:41:00] [Rank 0] step:7321/10000 train_time:323257ms step_avg:44.15ms
+[2025-09-04 10:41:01] [Rank 0] step:7341/10000 train_time:324019ms step_avg:44.14ms
+[2025-09-04 10:41:01] [Rank 0] step:7341/10000 train_time:324019ms step_avg:44.14ms
+[2025-09-04 10:41:02] [Rank 0] step:7361/10000 train_time:324782ms step_avg:44.12ms
+[2025-09-04 10:41:02] [Rank 0] step:7361/10000 train_time:324782ms step_avg:44.12ms
+[2025-09-04 10:41:03] [Rank 0] step:7381/10000 train_time:325545ms step_avg:44.11ms
+[2025-09-04 10:41:03] [Rank 0] step:7381/10000 train_time:325545ms step_avg:44.11ms
+[2025-09-04 10:41:03] [Rank 0] step:7401/10000 train_time:326307ms step_avg:44.09ms
+[2025-09-04 10:41:03] [Rank 0] step:7401/10000 train_time:326307ms step_avg:44.09ms
+[2025-09-04 10:41:04] [Rank 0] step:7421/10000 train_time:327070ms step_avg:44.07ms
+[2025-09-04 10:41:04] [Rank 0] step:7421/10000 train_time:327070ms step_avg:44.07ms
+[2025-09-04 10:41:05] [Rank 0] step:7441/10000 train_time:327833ms step_avg:44.06ms
+[2025-09-04 10:41:05] [Rank 0] step:7441/10000 train_time:327833ms step_avg:44.06ms
+[2025-09-04 10:41:06] [Rank 0] step:7461/10000 train_time:328595ms step_avg:44.04ms
+[2025-09-04 10:41:06] [Rank 0] step:7461/10000 train_time:328595ms step_avg:44.04ms
+[2025-09-04 10:41:06] [Rank 0] step:7481/10000 train_time:329357ms step_avg:44.03ms
+[2025-09-04 10:41:06] [Rank 0] step:7481/10000 train_time:329357ms step_avg:44.03ms
+[2025-09-04 10:41:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:41:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:41:07] [Rank 0] PRINT: step:7500/10000 train_loss:0.6256 val_loss:0.6155 train_time:330125ms step_avg:44.02ms
+[2025-09-04 10:41:07] [Rank 0] PRINT: step:7500/10000 train_loss:0.6256 val_loss:0.6155 train_time:330125ms step_avg:44.02ms
+[2025-09-04 10:41:07] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:41:07] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:41:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:41:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:42:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:42:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:42:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:42:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:42:47] [Rank 0]   Total Loss: 5.0324
+[2025-09-04 10:42:47] [Rank 0]   Total Loss: 5.0324
+[2025-09-04 10:42:47] [Rank 0]   Total FTA (Unweighted): 0.9731
+[2025-09-04 10:42:47] [Rank 0]   Total FTA (Unweighted): 0.9731
+[2025-09-04 10:42:47] [Rank 0]   Total FTA (Weighted):   0.9731
+[2025-09-04 10:42:47] [Rank 0]   Total FTA (Weighted):   0.9731
+[2025-09-04 10:42:47] [Rank 0]   Group 0 Loss: 5.1274
+[2025-09-04 10:42:47] [Rank 0]   Group 0 Loss: 5.1274
+[2025-09-04 10:42:47] [Rank 0]   Group 1 Loss: 4.5201
+[2025-09-04 10:42:47] [Rank 0]   Group 1 Loss: 4.5201
+[2025-09-04 10:42:47] [Rank 0]   Group 2 Loss: 4.4628
+[2025-09-04 10:42:47] [Rank 0]   Group 2 Loss: 4.4628
+[2025-09-04 10:42:47] [Rank 0]   Group 3 Loss: 4.9170
+[2025-09-04 10:42:47] [Rank 0]   Group 3 Loss: 4.9170
+[2025-09-04 10:42:47] [Rank 0]   Group 4 Loss: 4.8863
+[2025-09-04 10:42:47] [Rank 0]   Group 4 Loss: 4.8863
+[2025-09-04 10:42:47] [Rank 0]   Group 5 Loss: 5.0227
+[2025-09-04 10:42:47] [Rank 0]   Group 5 Loss: 5.0227
+[2025-09-04 10:42:47] [Rank 0]   Group 6 Loss: 4.8521
+[2025-09-04 10:42:47] [Rank 0]   Group 6 Loss: 4.8521
+[2025-09-04 10:42:47] [Rank 0]   Group 7 Loss: 5.0053
+[2025-09-04 10:42:47] [Rank 0]   Group 7 Loss: 5.0053
+[2025-09-04 10:42:47] [Rank 0]   Group 8 Loss: 5.1541
+[2025-09-04 10:42:47] [Rank 0]   Group 8 Loss: 5.1541
+[2025-09-04 10:42:47] [Rank 0]   Group 9 Loss: 5.1226
+[2025-09-04 10:42:47] [Rank 0]   Group 9 Loss: 5.1226
+[2025-09-04 10:42:47] [Rank 0]   Group 10 Loss: 5.2559
+[2025-09-04 10:42:47] [Rank 0]   Group 10 Loss: 5.2559
+[2025-09-04 10:42:47] [Rank 0]   Group 11 Loss: 5.2561
+[2025-09-04 10:42:47] [Rank 0]   Group 11 Loss: 5.2561
+[2025-09-04 10:42:47] [Rank 0]   Group 12 Loss: 5.1708
+[2025-09-04 10:42:47] [Rank 0]   Group 12 Loss: 5.1708
+[2025-09-04 10:42:47] [Rank 0]   Group 13 Loss: 5.3047
+[2025-09-04 10:42:47] [Rank 0]   Group 13 Loss: 5.3047
+[2025-09-04 10:42:47] [Rank 0]   Group 14 Loss: 5.2553
+[2025-09-04 10:42:47] [Rank 0]   Group 14 Loss: 5.2553
+[2025-09-04 10:42:47] [Rank 0]   Group 15 Loss: 5.2052
+[2025-09-04 10:42:47] [Rank 0]   Group 15 Loss: 5.2052
+[2025-09-04 10:42:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:42:47] [Rank 0]   Group 14 FTA: 0.9200
+[2025-09-04 10:42:47] [Rank 0]   Group 14 FTA: 0.9200
+[2025-09-04 10:42:47] [Rank 0]   Group 15 FTA: 0.6500
+[2025-09-04 10:42:47] [Rank 0]   Group 15 FTA: 0.6500
+[2025-09-04 10:42:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:42:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:42:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:42:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:42:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:42:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:42:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:42:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:42:48] [Rank 0] step:7501/10000 train_time:330140ms step_avg:44.01ms
+[2025-09-04 10:42:48] [Rank 0] step:7501/10000 train_time:330140ms step_avg:44.01ms
+[2025-09-04 10:42:49] [Rank 0] step:7521/10000 train_time:330923ms step_avg:44.00ms
+[2025-09-04 10:42:49] [Rank 0] step:7521/10000 train_time:330923ms step_avg:44.00ms
+[2025-09-04 10:42:50] [Rank 0] step:7541/10000 train_time:331684ms step_avg:43.98ms
+[2025-09-04 10:42:50] [Rank 0] step:7541/10000 train_time:331684ms step_avg:43.98ms
+[2025-09-04 10:42:51] [Rank 0] step:7561/10000 train_time:332446ms step_avg:43.97ms
+[2025-09-04 10:42:51] [Rank 0] step:7561/10000 train_time:332446ms step_avg:43.97ms
+[2025-09-04 10:42:51] [Rank 0] step:7581/10000 train_time:333207ms step_avg:43.95ms
+[2025-09-04 10:42:51] [Rank 0] step:7581/10000 train_time:333207ms step_avg:43.95ms
+[2025-09-04 10:42:52] [Rank 0] step:7601/10000 train_time:333969ms step_avg:43.94ms
+[2025-09-04 10:42:52] [Rank 0] step:7601/10000 train_time:333969ms step_avg:43.94ms
+[2025-09-04 10:42:53] [Rank 0] step:7621/10000 train_time:334730ms step_avg:43.92ms
+[2025-09-04 10:42:53] [Rank 0] step:7621/10000 train_time:334730ms step_avg:43.92ms
+[2025-09-04 10:42:54] [Rank 0] step:7641/10000 train_time:335761ms step_avg:43.94ms
+[2025-09-04 10:42:54] [Rank 0] step:7641/10000 train_time:335761ms step_avg:43.94ms
+[2025-09-04 10:42:55] [Rank 0] step:7661/10000 train_time:336524ms step_avg:43.93ms
+[2025-09-04 10:42:55] [Rank 0] step:7661/10000 train_time:336524ms step_avg:43.93ms
+[2025-09-04 10:42:55] [Rank 0] step:7681/10000 train_time:337287ms step_avg:43.91ms
+[2025-09-04 10:42:55] [Rank 0] step:7681/10000 train_time:337287ms step_avg:43.91ms
+[2025-09-04 10:42:56] [Rank 0] step:7701/10000 train_time:338049ms step_avg:43.90ms
+[2025-09-04 10:42:56] [Rank 0] step:7701/10000 train_time:338049ms step_avg:43.90ms
+[2025-09-04 10:42:57] [Rank 0] step:7721/10000 train_time:338811ms step_avg:43.88ms
+[2025-09-04 10:42:57] [Rank 0] step:7721/10000 train_time:338811ms step_avg:43.88ms
+[2025-09-04 10:42:58] [Rank 0] step:7741/10000 train_time:339573ms step_avg:43.87ms
+[2025-09-04 10:42:58] [Rank 0] step:7741/10000 train_time:339573ms step_avg:43.87ms
+[2025-09-04 10:42:58] [Rank 0] step:7761/10000 train_time:340335ms step_avg:43.85ms
+[2025-09-04 10:42:58] [Rank 0] step:7761/10000 train_time:340335ms step_avg:43.85ms
+[2025-09-04 10:42:59] [Rank 0] step:7781/10000 train_time:341096ms step_avg:43.84ms
+[2025-09-04 10:42:59] [Rank 0] step:7781/10000 train_time:341096ms step_avg:43.84ms
+[2025-09-04 10:43:00] [Rank 0] step:7801/10000 train_time:341857ms step_avg:43.82ms
+[2025-09-04 10:43:00] [Rank 0] step:7801/10000 train_time:341857ms step_avg:43.82ms
+[2025-09-04 10:43:01] [Rank 0] step:7821/10000 train_time:342618ms step_avg:43.81ms
+[2025-09-04 10:43:01] [Rank 0] step:7821/10000 train_time:342618ms step_avg:43.81ms
+[2025-09-04 10:43:01] [Rank 0] step:7841/10000 train_time:343380ms step_avg:43.79ms
+[2025-09-04 10:43:01] [Rank 0] step:7841/10000 train_time:343380ms step_avg:43.79ms
+[2025-09-04 10:43:02] [Rank 0] step:7861/10000 train_time:344142ms step_avg:43.78ms
+[2025-09-04 10:43:02] [Rank 0] step:7861/10000 train_time:344142ms step_avg:43.78ms
+[2025-09-04 10:43:03] [Rank 0] step:7881/10000 train_time:344904ms step_avg:43.76ms
+[2025-09-04 10:43:03] [Rank 0] step:7881/10000 train_time:344904ms step_avg:43.76ms
+[2025-09-04 10:43:04] [Rank 0] step:7901/10000 train_time:345666ms step_avg:43.75ms
+[2025-09-04 10:43:04] [Rank 0] step:7901/10000 train_time:345666ms step_avg:43.75ms
+[2025-09-04 10:43:05] [Rank 0] step:7921/10000 train_time:346428ms step_avg:43.74ms
+[2025-09-04 10:43:05] [Rank 0] step:7921/10000 train_time:346428ms step_avg:43.74ms
+[2025-09-04 10:43:05] [Rank 0] step:7941/10000 train_time:347190ms step_avg:43.72ms
+[2025-09-04 10:43:05] [Rank 0] step:7941/10000 train_time:347190ms step_avg:43.72ms
+[2025-09-04 10:43:06] [Rank 0] step:7961/10000 train_time:347951ms step_avg:43.71ms
+[2025-09-04 10:43:06] [Rank 0] step:7961/10000 train_time:347951ms step_avg:43.71ms
+[2025-09-04 10:43:07] [Rank 0] step:7981/10000 train_time:348713ms step_avg:43.69ms
+[2025-09-04 10:43:07] [Rank 0] step:7981/10000 train_time:348713ms step_avg:43.69ms
+[2025-09-04 10:43:08] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:43:08] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:43:08] [Rank 0] PRINT: step:8000/10000 train_loss:0.6208 val_loss:0.6122 train_time:349480ms step_avg:43.68ms
+[2025-09-04 10:43:08] [Rank 0] PRINT: step:8000/10000 train_loss:0.6208 val_loss:0.6122 train_time:349480ms step_avg:43.68ms
+[2025-09-04 10:43:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:43:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:43:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:43:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:44:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:44:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:44:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:44:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:44:47] [Rank 0]   Total Loss: 5.0036
+[2025-09-04 10:44:47] [Rank 0]   Total Loss: 5.0036
+[2025-09-04 10:44:47] [Rank 0]   Total FTA (Unweighted): 0.9856
+[2025-09-04 10:44:47] [Rank 0]   Total FTA (Unweighted): 0.9856
+[2025-09-04 10:44:47] [Rank 0]   Total FTA (Weighted):   0.9856
+[2025-09-04 10:44:47] [Rank 0]   Total FTA (Weighted):   0.9856
+[2025-09-04 10:44:47] [Rank 0]   Group 0 Loss: 5.0650
+[2025-09-04 10:44:47] [Rank 0]   Group 0 Loss: 5.0650
+[2025-09-04 10:44:47] [Rank 0]   Group 1 Loss: 4.5274
+[2025-09-04 10:44:47] [Rank 0]   Group 1 Loss: 4.5274
+[2025-09-04 10:44:47] [Rank 0]   Group 2 Loss: 4.4307
+[2025-09-04 10:44:47] [Rank 0]   Group 2 Loss: 4.4307
+[2025-09-04 10:44:47] [Rank 0]   Group 3 Loss: 4.8734
+[2025-09-04 10:44:47] [Rank 0]   Group 3 Loss: 4.8734
+[2025-09-04 10:44:47] [Rank 0]   Group 4 Loss: 4.8865
+[2025-09-04 10:44:47] [Rank 0]   Group 4 Loss: 4.8865
+[2025-09-04 10:44:47] [Rank 0]   Group 5 Loss: 4.9920
+[2025-09-04 10:44:47] [Rank 0]   Group 5 Loss: 4.9920
+[2025-09-04 10:44:47] [Rank 0]   Group 6 Loss: 4.8215
+[2025-09-04 10:44:47] [Rank 0]   Group 6 Loss: 4.8215
+[2025-09-04 10:44:47] [Rank 0]   Group 7 Loss: 4.9531
+[2025-09-04 10:44:47] [Rank 0]   Group 7 Loss: 4.9531
+[2025-09-04 10:44:47] [Rank 0]   Group 8 Loss: 5.1363
+[2025-09-04 10:44:47] [Rank 0]   Group 8 Loss: 5.1363
+[2025-09-04 10:44:47] [Rank 0]   Group 9 Loss: 5.0890
+[2025-09-04 10:44:47] [Rank 0]   Group 9 Loss: 5.0890
+[2025-09-04 10:44:47] [Rank 0]   Group 10 Loss: 5.2242
+[2025-09-04 10:44:47] [Rank 0]   Group 10 Loss: 5.2242
+[2025-09-04 10:44:47] [Rank 0]   Group 11 Loss: 5.1893
+[2025-09-04 10:44:47] [Rank 0]   Group 11 Loss: 5.1893
+[2025-09-04 10:44:47] [Rank 0]   Group 12 Loss: 5.1366
+[2025-09-04 10:44:47] [Rank 0]   Group 12 Loss: 5.1366
+[2025-09-04 10:44:47] [Rank 0]   Group 13 Loss: 5.2827
+[2025-09-04 10:44:47] [Rank 0]   Group 13 Loss: 5.2827
+[2025-09-04 10:44:47] [Rank 0]   Group 14 Loss: 5.2427
+[2025-09-04 10:44:47] [Rank 0]   Group 14 Loss: 5.2427
+[2025-09-04 10:44:47] [Rank 0]   Group 15 Loss: 5.2077
+[2025-09-04 10:44:47] [Rank 0]   Group 15 Loss: 5.2077
+[2025-09-04 10:44:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:44:47] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:44:47] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:44:47] [Rank 0]   Group 14 FTA: 0.9700
+[2025-09-04 10:44:47] [Rank 0]   Group 14 FTA: 0.9700
+[2025-09-04 10:44:47] [Rank 0]   Group 15 FTA: 0.8100
+[2025-09-04 10:44:47] [Rank 0]   Group 15 FTA: 0.8100
+[2025-09-04 10:44:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:44:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:44:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:44:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:44:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:44:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:44:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:44:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:44:49] [Rank 0] step:8001/10000 train_time:349496ms step_avg:43.68ms
+[2025-09-04 10:44:49] [Rank 0] step:8001/10000 train_time:349496ms step_avg:43.68ms
+[2025-09-04 10:44:50] [Rank 0] step:8021/10000 train_time:350539ms step_avg:43.70ms
+[2025-09-04 10:44:50] [Rank 0] step:8021/10000 train_time:350539ms step_avg:43.70ms
+[2025-09-04 10:44:51] [Rank 0] step:8041/10000 train_time:351301ms step_avg:43.69ms
+[2025-09-04 10:44:51] [Rank 0] step:8041/10000 train_time:351301ms step_avg:43.69ms
+[2025-09-04 10:44:51] [Rank 0] step:8061/10000 train_time:352063ms step_avg:43.67ms
+[2025-09-04 10:44:51] [Rank 0] step:8061/10000 train_time:352063ms step_avg:43.67ms
+[2025-09-04 10:44:52] [Rank 0] step:8081/10000 train_time:352824ms step_avg:43.66ms
+[2025-09-04 10:44:52] [Rank 0] step:8081/10000 train_time:352824ms step_avg:43.66ms
+[2025-09-04 10:44:53] [Rank 0] step:8101/10000 train_time:353586ms step_avg:43.65ms
+[2025-09-04 10:44:53] [Rank 0] step:8101/10000 train_time:353586ms step_avg:43.65ms
+[2025-09-04 10:44:54] [Rank 0] step:8121/10000 train_time:354347ms step_avg:43.63ms
+[2025-09-04 10:44:54] [Rank 0] step:8121/10000 train_time:354347ms step_avg:43.63ms
+[2025-09-04 10:44:54] [Rank 0] step:8141/10000 train_time:355109ms step_avg:43.62ms
+[2025-09-04 10:44:54] [Rank 0] step:8141/10000 train_time:355109ms step_avg:43.62ms
+[2025-09-04 10:44:55] [Rank 0] step:8161/10000 train_time:355870ms step_avg:43.61ms
+[2025-09-04 10:44:55] [Rank 0] step:8161/10000 train_time:355870ms step_avg:43.61ms
+[2025-09-04 10:44:56] [Rank 0] step:8181/10000 train_time:356631ms step_avg:43.59ms
+[2025-09-04 10:44:56] [Rank 0] step:8181/10000 train_time:356631ms step_avg:43.59ms
+[2025-09-04 10:44:57] [Rank 0] step:8201/10000 train_time:357393ms step_avg:43.58ms
+[2025-09-04 10:44:57] [Rank 0] step:8201/10000 train_time:357393ms step_avg:43.58ms
+[2025-09-04 10:44:57] [Rank 0] step:8221/10000 train_time:358154ms step_avg:43.57ms
+[2025-09-04 10:44:57] [Rank 0] step:8221/10000 train_time:358154ms step_avg:43.57ms
+[2025-09-04 10:44:58] [Rank 0] step:8241/10000 train_time:358916ms step_avg:43.55ms
+[2025-09-04 10:44:58] [Rank 0] step:8241/10000 train_time:358916ms step_avg:43.55ms
+[2025-09-04 10:44:59] [Rank 0] step:8261/10000 train_time:359677ms step_avg:43.54ms
+[2025-09-04 10:44:59] [Rank 0] step:8261/10000 train_time:359677ms step_avg:43.54ms
+[2025-09-04 10:45:00] [Rank 0] step:8281/10000 train_time:360439ms step_avg:43.53ms
+[2025-09-04 10:45:00] [Rank 0] step:8281/10000 train_time:360439ms step_avg:43.53ms
+[2025-09-04 10:45:00] [Rank 0] step:8301/10000 train_time:361201ms step_avg:43.51ms
+[2025-09-04 10:45:00] [Rank 0] step:8301/10000 train_time:361201ms step_avg:43.51ms
+[2025-09-04 10:45:01] [Rank 0] step:8321/10000 train_time:361998ms step_avg:43.50ms
+[2025-09-04 10:45:01] [Rank 0] step:8321/10000 train_time:361998ms step_avg:43.50ms
+[2025-09-04 10:45:02] [Rank 0] step:8341/10000 train_time:362802ms step_avg:43.50ms
+[2025-09-04 10:45:02] [Rank 0] step:8341/10000 train_time:362802ms step_avg:43.50ms
+[2025-09-04 10:45:03] [Rank 0] step:8361/10000 train_time:363565ms step_avg:43.48ms
+[2025-09-04 10:45:03] [Rank 0] step:8361/10000 train_time:363565ms step_avg:43.48ms
+[2025-09-04 10:45:04] [Rank 0] step:8381/10000 train_time:364327ms step_avg:43.47ms
+[2025-09-04 10:45:04] [Rank 0] step:8381/10000 train_time:364327ms step_avg:43.47ms
+[2025-09-04 10:45:04] [Rank 0] step:8401/10000 train_time:365089ms step_avg:43.46ms
+[2025-09-04 10:45:04] [Rank 0] step:8401/10000 train_time:365089ms step_avg:43.46ms
+[2025-09-04 10:45:05] [Rank 0] step:8421/10000 train_time:365851ms step_avg:43.45ms
+[2025-09-04 10:45:05] [Rank 0] step:8421/10000 train_time:365851ms step_avg:43.45ms
+[2025-09-04 10:45:06] [Rank 0] step:8441/10000 train_time:366613ms step_avg:43.43ms
+[2025-09-04 10:45:06] [Rank 0] step:8441/10000 train_time:366613ms step_avg:43.43ms
+[2025-09-04 10:45:07] [Rank 0] step:8461/10000 train_time:367375ms step_avg:43.42ms
+[2025-09-04 10:45:07] [Rank 0] step:8461/10000 train_time:367375ms step_avg:43.42ms
+[2025-09-04 10:45:07] [Rank 0] step:8481/10000 train_time:368137ms step_avg:43.41ms
+[2025-09-04 10:45:07] [Rank 0] step:8481/10000 train_time:368137ms step_avg:43.41ms
+[2025-09-04 10:45:08] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:45:08] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:45:09] [Rank 0] PRINT: step:8500/10000 train_loss:0.6168 val_loss:0.6096 train_time:368905ms step_avg:43.40ms
+[2025-09-04 10:45:09] [Rank 0] PRINT: step:8500/10000 train_loss:0.6168 val_loss:0.6096 train_time:368905ms step_avg:43.40ms
+[2025-09-04 10:45:09] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:45:09] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:45:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:45:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:46:48] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:46:48] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:46:48] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:46:48] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:46:48] [Rank 0]   Total Loss: 5.0298
+[2025-09-04 10:46:48] [Rank 0]   Total Loss: 5.0298
+[2025-09-04 10:46:48] [Rank 0]   Total FTA (Unweighted): 0.9900
+[2025-09-04 10:46:48] [Rank 0]   Total FTA (Unweighted): 0.9900
+[2025-09-04 10:46:48] [Rank 0]   Total FTA (Weighted):   0.9900
+[2025-09-04 10:46:48] [Rank 0]   Total FTA (Weighted):   0.9900
+[2025-09-04 10:46:48] [Rank 0]   Group 0 Loss: 5.0870
+[2025-09-04 10:46:48] [Rank 0]   Group 0 Loss: 5.0870
+[2025-09-04 10:46:48] [Rank 0]   Group 1 Loss: 4.5272
+[2025-09-04 10:46:48] [Rank 0]   Group 1 Loss: 4.5272
+[2025-09-04 10:46:48] [Rank 0]   Group 2 Loss: 4.4254
+[2025-09-04 10:46:48] [Rank 0]   Group 2 Loss: 4.4254
+[2025-09-04 10:46:48] [Rank 0]   Group 3 Loss: 4.9216
+[2025-09-04 10:46:48] [Rank 0]   Group 3 Loss: 4.9216
+[2025-09-04 10:46:48] [Rank 0]   Group 4 Loss: 4.8718
+[2025-09-04 10:46:48] [Rank 0]   Group 4 Loss: 4.8718
+[2025-09-04 10:46:48] [Rank 0]   Group 5 Loss: 5.0275
+[2025-09-04 10:46:48] [Rank 0]   Group 5 Loss: 5.0275
+[2025-09-04 10:46:48] [Rank 0]   Group 6 Loss: 4.8587
+[2025-09-04 10:46:48] [Rank 0]   Group 6 Loss: 4.8587
+[2025-09-04 10:46:48] [Rank 0]   Group 7 Loss: 4.9729
+[2025-09-04 10:46:48] [Rank 0]   Group 7 Loss: 4.9729
+[2025-09-04 10:46:48] [Rank 0]   Group 8 Loss: 5.1644
+[2025-09-04 10:46:48] [Rank 0]   Group 8 Loss: 5.1644
+[2025-09-04 10:46:48] [Rank 0]   Group 9 Loss: 5.1185
+[2025-09-04 10:46:48] [Rank 0]   Group 9 Loss: 5.1185
+[2025-09-04 10:46:48] [Rank 0]   Group 10 Loss: 5.2594
+[2025-09-04 10:46:48] [Rank 0]   Group 10 Loss: 5.2594
+[2025-09-04 10:46:48] [Rank 0]   Group 11 Loss: 5.2447
+[2025-09-04 10:46:48] [Rank 0]   Group 11 Loss: 5.2447
+[2025-09-04 10:46:48] [Rank 0]   Group 12 Loss: 5.1661
+[2025-09-04 10:46:48] [Rank 0]   Group 12 Loss: 5.1661
+[2025-09-04 10:46:48] [Rank 0]   Group 13 Loss: 5.3260
+[2025-09-04 10:46:48] [Rank 0]   Group 13 Loss: 5.3260
+[2025-09-04 10:46:48] [Rank 0]   Group 14 Loss: 5.2760
+[2025-09-04 10:46:48] [Rank 0]   Group 14 Loss: 5.2760
+[2025-09-04 10:46:48] [Rank 0]   Group 15 Loss: 5.2289
+[2025-09-04 10:46:48] [Rank 0]   Group 15 Loss: 5.2289
+[2025-09-04 10:46:48] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:46:48] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:46:48] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:46:48] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 10:46:48] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 10:46:48] [Rank 0]   Group 15 FTA: 0.8600
+[2025-09-04 10:46:48] [Rank 0]   Group 15 FTA: 0.8600
+[2025-09-04 10:46:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:46:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:46:49] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:46:49] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:46:49] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:46:49] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:46:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:46:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:46:49] [Rank 0] step:8501/10000 train_time:368920ms step_avg:43.40ms
+[2025-09-04 10:46:49] [Rank 0] step:8501/10000 train_time:368920ms step_avg:43.40ms
+[2025-09-04 10:46:50] [Rank 0] step:8521/10000 train_time:369685ms step_avg:43.39ms
+[2025-09-04 10:46:50] [Rank 0] step:8521/10000 train_time:369685ms step_avg:43.39ms
+[2025-09-04 10:46:51] [Rank 0] step:8541/10000 train_time:370447ms step_avg:43.37ms
+[2025-09-04 10:46:51] [Rank 0] step:8541/10000 train_time:370447ms step_avg:43.37ms
+[2025-09-04 10:46:52] [Rank 0] step:8561/10000 train_time:371208ms step_avg:43.36ms
+[2025-09-04 10:46:52] [Rank 0] step:8561/10000 train_time:371208ms step_avg:43.36ms
+[2025-09-04 10:46:52] [Rank 0] step:8581/10000 train_time:371970ms step_avg:43.35ms
+[2025-09-04 10:46:52] [Rank 0] step:8581/10000 train_time:371970ms step_avg:43.35ms
+[2025-09-04 10:46:53] [Rank 0] step:8601/10000 train_time:372730ms step_avg:43.34ms
+[2025-09-04 10:46:53] [Rank 0] step:8601/10000 train_time:372730ms step_avg:43.34ms
+[2025-09-04 10:46:54] [Rank 0] step:8621/10000 train_time:373492ms step_avg:43.32ms
+[2025-09-04 10:46:54] [Rank 0] step:8621/10000 train_time:373492ms step_avg:43.32ms
+[2025-09-04 10:46:55] [Rank 0] step:8641/10000 train_time:374253ms step_avg:43.31ms
+[2025-09-04 10:46:55] [Rank 0] step:8641/10000 train_time:374253ms step_avg:43.31ms
+[2025-09-04 10:46:55] [Rank 0] step:8661/10000 train_time:375015ms step_avg:43.30ms
+[2025-09-04 10:46:55] [Rank 0] step:8661/10000 train_time:375015ms step_avg:43.30ms
+[2025-09-04 10:46:56] [Rank 0] step:8681/10000 train_time:375775ms step_avg:43.29ms
+[2025-09-04 10:46:56] [Rank 0] step:8681/10000 train_time:375775ms step_avg:43.29ms
+[2025-09-04 10:46:57] [Rank 0] step:8701/10000 train_time:376538ms step_avg:43.28ms
+[2025-09-04 10:46:57] [Rank 0] step:8701/10000 train_time:376538ms step_avg:43.28ms
+[2025-09-04 10:46:58] [Rank 0] step:8721/10000 train_time:377298ms step_avg:43.26ms
+[2025-09-04 10:46:58] [Rank 0] step:8721/10000 train_time:377298ms step_avg:43.26ms
+[2025-09-04 10:46:58] [Rank 0] step:8741/10000 train_time:378061ms step_avg:43.25ms
+[2025-09-04 10:46:58] [Rank 0] step:8741/10000 train_time:378061ms step_avg:43.25ms
+[2025-09-04 10:46:59] [Rank 0] step:8761/10000 train_time:378821ms step_avg:43.24ms
+[2025-09-04 10:46:59] [Rank 0] step:8761/10000 train_time:378821ms step_avg:43.24ms
+[2025-09-04 10:47:00] [Rank 0] step:8781/10000 train_time:379583ms step_avg:43.23ms
+[2025-09-04 10:47:00] [Rank 0] step:8781/10000 train_time:379583ms step_avg:43.23ms
+[2025-09-04 10:47:01] [Rank 0] step:8801/10000 train_time:380343ms step_avg:43.22ms
+[2025-09-04 10:47:01] [Rank 0] step:8801/10000 train_time:380343ms step_avg:43.22ms
+[2025-09-04 10:47:01] [Rank 0] step:8821/10000 train_time:381104ms step_avg:43.20ms
+[2025-09-04 10:47:01] [Rank 0] step:8821/10000 train_time:381104ms step_avg:43.20ms
+[2025-09-04 10:47:03] [Rank 0] step:8841/10000 train_time:382139ms step_avg:43.22ms
+[2025-09-04 10:47:03] [Rank 0] step:8841/10000 train_time:382139ms step_avg:43.22ms
+[2025-09-04 10:47:03] [Rank 0] step:8861/10000 train_time:382901ms step_avg:43.21ms
+[2025-09-04 10:47:03] [Rank 0] step:8861/10000 train_time:382901ms step_avg:43.21ms
+[2025-09-04 10:47:04] [Rank 0] step:8881/10000 train_time:383662ms step_avg:43.20ms
+[2025-09-04 10:47:04] [Rank 0] step:8881/10000 train_time:383662ms step_avg:43.20ms
+[2025-09-04 10:47:05] [Rank 0] step:8901/10000 train_time:384424ms step_avg:43.19ms
+[2025-09-04 10:47:05] [Rank 0] step:8901/10000 train_time:384424ms step_avg:43.19ms
+[2025-09-04 10:47:06] [Rank 0] step:8921/10000 train_time:385185ms step_avg:43.18ms
+[2025-09-04 10:47:06] [Rank 0] step:8921/10000 train_time:385185ms step_avg:43.18ms
+[2025-09-04 10:47:06] [Rank 0] step:8941/10000 train_time:385946ms step_avg:43.17ms
+[2025-09-04 10:47:06] [Rank 0] step:8941/10000 train_time:385946ms step_avg:43.17ms
+[2025-09-04 10:47:07] [Rank 0] step:8961/10000 train_time:386708ms step_avg:43.15ms
+[2025-09-04 10:47:07] [Rank 0] step:8961/10000 train_time:386708ms step_avg:43.15ms
+[2025-09-04 10:47:08] [Rank 0] step:8981/10000 train_time:387469ms step_avg:43.14ms
+[2025-09-04 10:47:08] [Rank 0] step:8981/10000 train_time:387469ms step_avg:43.14ms
+[2025-09-04 10:47:09] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:47:09] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:47:09] [Rank 0] PRINT: step:9000/10000 train_loss:0.6133 val_loss:0.6076 train_time:388235ms step_avg:43.14ms
+[2025-09-04 10:47:09] [Rank 0] PRINT: step:9000/10000 train_loss:0.6133 val_loss:0.6076 train_time:388235ms step_avg:43.14ms
+[2025-09-04 10:47:09] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:47:09] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:47:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:47:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:48:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:48:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:48:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:48:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:48:47] [Rank 0]   Total Loss: 5.0611
+[2025-09-04 10:48:47] [Rank 0]   Total Loss: 5.0611
+[2025-09-04 10:48:47] [Rank 0]   Total FTA (Unweighted): 0.9938
+[2025-09-04 10:48:47] [Rank 0]   Total FTA (Unweighted): 0.9938
+[2025-09-04 10:48:47] [Rank 0]   Total FTA (Weighted):   0.9938
+[2025-09-04 10:48:47] [Rank 0]   Total FTA (Weighted):   0.9938
+[2025-09-04 10:48:47] [Rank 0]   Group 0 Loss: 5.0970
+[2025-09-04 10:48:47] [Rank 0]   Group 0 Loss: 5.0970
+[2025-09-04 10:48:47] [Rank 0]   Group 1 Loss: 4.5471
+[2025-09-04 10:48:47] [Rank 0]   Group 1 Loss: 4.5471
+[2025-09-04 10:48:47] [Rank 0]   Group 2 Loss: 4.4768
+[2025-09-04 10:48:47] [Rank 0]   Group 2 Loss: 4.4768
+[2025-09-04 10:48:47] [Rank 0]   Group 3 Loss: 4.9557
+[2025-09-04 10:48:47] [Rank 0]   Group 3 Loss: 4.9557
+[2025-09-04 10:48:47] [Rank 0]   Group 4 Loss: 4.8993
+[2025-09-04 10:48:47] [Rank 0]   Group 4 Loss: 4.8993
+[2025-09-04 10:48:47] [Rank 0]   Group 5 Loss: 5.0457
+[2025-09-04 10:48:47] [Rank 0]   Group 5 Loss: 5.0457
+[2025-09-04 10:48:47] [Rank 0]   Group 6 Loss: 4.8953
+[2025-09-04 10:48:47] [Rank 0]   Group 6 Loss: 4.8953
+[2025-09-04 10:48:47] [Rank 0]   Group 7 Loss: 5.0016
+[2025-09-04 10:48:47] [Rank 0]   Group 7 Loss: 5.0016
+[2025-09-04 10:48:47] [Rank 0]   Group 8 Loss: 5.2039
+[2025-09-04 10:48:47] [Rank 0]   Group 8 Loss: 5.2039
+[2025-09-04 10:48:47] [Rank 0]   Group 9 Loss: 5.1525
+[2025-09-04 10:48:47] [Rank 0]   Group 9 Loss: 5.1525
+[2025-09-04 10:48:47] [Rank 0]   Group 10 Loss: 5.2818
+[2025-09-04 10:48:47] [Rank 0]   Group 10 Loss: 5.2818
+[2025-09-04 10:48:47] [Rank 0]   Group 11 Loss: 5.2926
+[2025-09-04 10:48:47] [Rank 0]   Group 11 Loss: 5.2926
+[2025-09-04 10:48:47] [Rank 0]   Group 12 Loss: 5.1851
+[2025-09-04 10:48:47] [Rank 0]   Group 12 Loss: 5.1851
+[2025-09-04 10:48:47] [Rank 0]   Group 13 Loss: 5.3645
+[2025-09-04 10:48:47] [Rank 0]   Group 13 Loss: 5.3645
+[2025-09-04 10:48:47] [Rank 0]   Group 14 Loss: 5.3194
+[2025-09-04 10:48:47] [Rank 0]   Group 14 Loss: 5.3194
+[2025-09-04 10:48:47] [Rank 0]   Group 15 Loss: 5.2585
+[2025-09-04 10:48:47] [Rank 0]   Group 15 Loss: 5.2585
+[2025-09-04 10:48:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:48:47] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:48:47] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 10:48:47] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 10:48:47] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 10:48:47] [Rank 0]   Group 15 FTA: 0.9200
+[2025-09-04 10:48:47] [Rank 0]   Group 15 FTA: 0.9200
+[2025-09-04 10:48:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:48:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:48:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:48:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:48:49] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:48:49] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:48:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:48:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:48:49] [Rank 0] step:9001/10000 train_time:388253ms step_avg:43.13ms
+[2025-09-04 10:48:49] [Rank 0] step:9001/10000 train_time:388253ms step_avg:43.13ms
+[2025-09-04 10:48:50] [Rank 0] step:9021/10000 train_time:389023ms step_avg:43.12ms
+[2025-09-04 10:48:50] [Rank 0] step:9021/10000 train_time:389023ms step_avg:43.12ms
+[2025-09-04 10:48:50] [Rank 0] step:9041/10000 train_time:389783ms step_avg:43.11ms
+[2025-09-04 10:48:50] [Rank 0] step:9041/10000 train_time:389783ms step_avg:43.11ms
+[2025-09-04 10:48:51] [Rank 0] step:9061/10000 train_time:390544ms step_avg:43.10ms
+[2025-09-04 10:48:51] [Rank 0] step:9061/10000 train_time:390544ms step_avg:43.10ms
+[2025-09-04 10:48:52] [Rank 0] step:9081/10000 train_time:391303ms step_avg:43.09ms
+[2025-09-04 10:48:52] [Rank 0] step:9081/10000 train_time:391303ms step_avg:43.09ms
+[2025-09-04 10:48:53] [Rank 0] step:9101/10000 train_time:392074ms step_avg:43.08ms
+[2025-09-04 10:48:53] [Rank 0] step:9101/10000 train_time:392074ms step_avg:43.08ms
+[2025-09-04 10:48:53] [Rank 0] step:9121/10000 train_time:392834ms step_avg:43.07ms
+[2025-09-04 10:48:53] [Rank 0] step:9121/10000 train_time:392834ms step_avg:43.07ms
+[2025-09-04 10:48:54] [Rank 0] step:9141/10000 train_time:393594ms step_avg:43.06ms
+[2025-09-04 10:48:54] [Rank 0] step:9141/10000 train_time:393594ms step_avg:43.06ms
+[2025-09-04 10:48:55] [Rank 0] step:9161/10000 train_time:394356ms step_avg:43.05ms
+[2025-09-04 10:48:55] [Rank 0] step:9161/10000 train_time:394356ms step_avg:43.05ms
+[2025-09-04 10:48:56] [Rank 0] step:9181/10000 train_time:395117ms step_avg:43.04ms
+[2025-09-04 10:48:56] [Rank 0] step:9181/10000 train_time:395117ms step_avg:43.04ms
+[2025-09-04 10:48:57] [Rank 0] step:9201/10000 train_time:395878ms step_avg:43.03ms
+[2025-09-04 10:48:57] [Rank 0] step:9201/10000 train_time:395878ms step_avg:43.03ms
+[2025-09-04 10:48:57] [Rank 0] step:9221/10000 train_time:396639ms step_avg:43.01ms
+[2025-09-04 10:48:57] [Rank 0] step:9221/10000 train_time:396639ms step_avg:43.01ms
+[2025-09-04 10:48:58] [Rank 0] step:9241/10000 train_time:397400ms step_avg:43.00ms
+[2025-09-04 10:48:58] [Rank 0] step:9241/10000 train_time:397400ms step_avg:43.00ms
+[2025-09-04 10:48:59] [Rank 0] step:9261/10000 train_time:398160ms step_avg:42.99ms
+[2025-09-04 10:48:59] [Rank 0] step:9261/10000 train_time:398160ms step_avg:42.99ms
+[2025-09-04 10:49:00] [Rank 0] step:9281/10000 train_time:398921ms step_avg:42.98ms
+[2025-09-04 10:49:00] [Rank 0] step:9281/10000 train_time:398921ms step_avg:42.98ms
+[2025-09-04 10:49:00] [Rank 0] step:9301/10000 train_time:399682ms step_avg:42.97ms
+[2025-09-04 10:49:00] [Rank 0] step:9301/10000 train_time:399682ms step_avg:42.97ms
+[2025-09-04 10:49:01] [Rank 0] step:9321/10000 train_time:400444ms step_avg:42.96ms
+[2025-09-04 10:49:01] [Rank 0] step:9321/10000 train_time:400444ms step_avg:42.96ms
+[2025-09-04 10:49:02] [Rank 0] step:9341/10000 train_time:401206ms step_avg:42.95ms
+[2025-09-04 10:49:02] [Rank 0] step:9341/10000 train_time:401206ms step_avg:42.95ms
+[2025-09-04 10:49:03] [Rank 0] step:9361/10000 train_time:401968ms step_avg:42.94ms
+[2025-09-04 10:49:03] [Rank 0] step:9361/10000 train_time:401968ms step_avg:42.94ms
+[2025-09-04 10:49:03] [Rank 0] step:9381/10000 train_time:402729ms step_avg:42.93ms
+[2025-09-04 10:49:03] [Rank 0] step:9381/10000 train_time:402729ms step_avg:42.93ms
+[2025-09-04 10:49:04] [Rank 0] step:9401/10000 train_time:403491ms step_avg:42.92ms
+[2025-09-04 10:49:04] [Rank 0] step:9401/10000 train_time:403491ms step_avg:42.92ms
+[2025-09-04 10:49:05] [Rank 0] step:9421/10000 train_time:404252ms step_avg:42.91ms
+[2025-09-04 10:49:05] [Rank 0] step:9421/10000 train_time:404252ms step_avg:42.91ms
+[2025-09-04 10:49:06] [Rank 0] step:9441/10000 train_time:405052ms step_avg:42.90ms
+[2025-09-04 10:49:06] [Rank 0] step:9441/10000 train_time:405052ms step_avg:42.90ms
+[2025-09-04 10:49:06] [Rank 0] step:9461/10000 train_time:405812ms step_avg:42.89ms
+[2025-09-04 10:49:06] [Rank 0] step:9461/10000 train_time:405812ms step_avg:42.89ms
+[2025-09-04 10:49:07] [Rank 0] step:9481/10000 train_time:406573ms step_avg:42.88ms
+[2025-09-04 10:49:07] [Rank 0] step:9481/10000 train_time:406573ms step_avg:42.88ms
+[2025-09-04 10:49:08] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:49:08] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:49:08] [Rank 0] PRINT: step:9500/10000 train_loss:0.6107 val_loss:0.6061 train_time:407339ms step_avg:42.88ms
+[2025-09-04 10:49:08] [Rank 0] PRINT: step:9500/10000 train_loss:0.6107 val_loss:0.6061 train_time:407339ms step_avg:42.88ms
+[2025-09-04 10:49:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:49:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:49:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:49:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:50:48] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:50:48] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:50:48] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:50:48] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:50:48] [Rank 0]   Total Loss: 5.0786
+[2025-09-04 10:50:48] [Rank 0]   Total Loss: 5.0786
+[2025-09-04 10:50:48] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 10:50:48] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 10:50:48] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 10:50:48] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 10:50:48] [Rank 0]   Group 0 Loss: 5.1766
+[2025-09-04 10:50:48] [Rank 0]   Group 0 Loss: 5.1766
+[2025-09-04 10:50:48] [Rank 0]   Group 1 Loss: 4.5569
+[2025-09-04 10:50:48] [Rank 0]   Group 1 Loss: 4.5569
+[2025-09-04 10:50:48] [Rank 0]   Group 2 Loss: 4.4697
+[2025-09-04 10:50:48] [Rank 0]   Group 2 Loss: 4.4697
+[2025-09-04 10:50:48] [Rank 0]   Group 3 Loss: 4.9868
+[2025-09-04 10:50:48] [Rank 0]   Group 3 Loss: 4.9868
+[2025-09-04 10:50:48] [Rank 0]   Group 4 Loss: 4.9364
+[2025-09-04 10:50:48] [Rank 0]   Group 4 Loss: 4.9364
+[2025-09-04 10:50:48] [Rank 0]   Group 5 Loss: 5.0541
+[2025-09-04 10:50:48] [Rank 0]   Group 5 Loss: 5.0541
+[2025-09-04 10:50:48] [Rank 0]   Group 6 Loss: 4.8855
+[2025-09-04 10:50:48] [Rank 0]   Group 6 Loss: 4.8855
+[2025-09-04 10:50:48] [Rank 0]   Group 7 Loss: 5.0342
+[2025-09-04 10:50:48] [Rank 0]   Group 7 Loss: 5.0342
+[2025-09-04 10:50:48] [Rank 0]   Group 8 Loss: 5.1990
+[2025-09-04 10:50:48] [Rank 0]   Group 8 Loss: 5.1990
+[2025-09-04 10:50:48] [Rank 0]   Group 9 Loss: 5.1527
+[2025-09-04 10:50:48] [Rank 0]   Group 9 Loss: 5.1527
+[2025-09-04 10:50:48] [Rank 0]   Group 10 Loss: 5.2916
+[2025-09-04 10:50:48] [Rank 0]   Group 10 Loss: 5.2916
+[2025-09-04 10:50:48] [Rank 0]   Group 11 Loss: 5.2874
+[2025-09-04 10:50:48] [Rank 0]   Group 11 Loss: 5.2874
+[2025-09-04 10:50:48] [Rank 0]   Group 12 Loss: 5.2102
+[2025-09-04 10:50:48] [Rank 0]   Group 12 Loss: 5.2102
+[2025-09-04 10:50:48] [Rank 0]   Group 13 Loss: 5.3767
+[2025-09-04 10:50:48] [Rank 0]   Group 13 Loss: 5.3767
+[2025-09-04 10:50:48] [Rank 0]   Group 14 Loss: 5.3476
+[2025-09-04 10:50:48] [Rank 0]   Group 14 Loss: 5.3476
+[2025-09-04 10:50:48] [Rank 0]   Group 15 Loss: 5.2929
+[2025-09-04 10:50:48] [Rank 0]   Group 15 Loss: 5.2929
+[2025-09-04 10:50:48] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 10:50:48] [Rank 0]   Group 15 FTA: 0.9300
+[2025-09-04 10:50:48] [Rank 0]   Group 15 FTA: 0.9300
+[2025-09-04 10:50:49] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:50:49] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:50:49] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:50:49] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:50:49] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:50:49] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:50:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:50:49] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:50:50] [Rank 0] step:9501/10000 train_time:407357ms step_avg:42.88ms
+[2025-09-04 10:50:50] [Rank 0] step:9501/10000 train_time:407357ms step_avg:42.88ms
+[2025-09-04 10:50:50] [Rank 0] step:9521/10000 train_time:408116ms step_avg:42.86ms
+[2025-09-04 10:50:50] [Rank 0] step:9521/10000 train_time:408116ms step_avg:42.86ms
+[2025-09-04 10:50:51] [Rank 0] step:9541/10000 train_time:408876ms step_avg:42.85ms
+[2025-09-04 10:50:51] [Rank 0] step:9541/10000 train_time:408876ms step_avg:42.85ms
+[2025-09-04 10:50:52] [Rank 0] step:9561/10000 train_time:409636ms step_avg:42.84ms
+[2025-09-04 10:50:52] [Rank 0] step:9561/10000 train_time:409636ms step_avg:42.84ms
+[2025-09-04 10:50:53] [Rank 0] step:9581/10000 train_time:410396ms step_avg:42.83ms
+[2025-09-04 10:50:53] [Rank 0] step:9581/10000 train_time:410396ms step_avg:42.83ms
+[2025-09-04 10:50:53] [Rank 0] step:9601/10000 train_time:411157ms step_avg:42.82ms
+[2025-09-04 10:50:53] [Rank 0] step:9601/10000 train_time:411157ms step_avg:42.82ms
+[2025-09-04 10:50:54] [Rank 0] step:9621/10000 train_time:411917ms step_avg:42.81ms
+[2025-09-04 10:50:54] [Rank 0] step:9621/10000 train_time:411917ms step_avg:42.81ms
+[2025-09-04 10:50:55] [Rank 0] step:9641/10000 train_time:412678ms step_avg:42.80ms
+[2025-09-04 10:50:55] [Rank 0] step:9641/10000 train_time:412678ms step_avg:42.80ms
+[2025-09-04 10:50:56] [Rank 0] step:9661/10000 train_time:413719ms step_avg:42.82ms
+[2025-09-04 10:50:56] [Rank 0] step:9661/10000 train_time:413719ms step_avg:42.82ms
+[2025-09-04 10:50:57] [Rank 0] step:9681/10000 train_time:414479ms step_avg:42.81ms
+[2025-09-04 10:50:57] [Rank 0] step:9681/10000 train_time:414479ms step_avg:42.81ms
+[2025-09-04 10:50:57] [Rank 0] step:9701/10000 train_time:415239ms step_avg:42.80ms
+[2025-09-04 10:50:57] [Rank 0] step:9701/10000 train_time:415239ms step_avg:42.80ms
+[2025-09-04 10:50:58] [Rank 0] step:9721/10000 train_time:416000ms step_avg:42.79ms
+[2025-09-04 10:50:58] [Rank 0] step:9721/10000 train_time:416000ms step_avg:42.79ms
+[2025-09-04 10:50:59] [Rank 0] step:9741/10000 train_time:416760ms step_avg:42.78ms
+[2025-09-04 10:50:59] [Rank 0] step:9741/10000 train_time:416760ms step_avg:42.78ms
+[2025-09-04 10:51:00] [Rank 0] step:9761/10000 train_time:417520ms step_avg:42.77ms
+[2025-09-04 10:51:00] [Rank 0] step:9761/10000 train_time:417520ms step_avg:42.77ms
+[2025-09-04 10:51:00] [Rank 0] step:9781/10000 train_time:418280ms step_avg:42.76ms
+[2025-09-04 10:51:00] [Rank 0] step:9781/10000 train_time:418280ms step_avg:42.76ms
+[2025-09-04 10:51:01] [Rank 0] step:9801/10000 train_time:419040ms step_avg:42.75ms
+[2025-09-04 10:51:01] [Rank 0] step:9801/10000 train_time:419040ms step_avg:42.75ms
+[2025-09-04 10:51:02] [Rank 0] step:9821/10000 train_time:419801ms step_avg:42.75ms
+[2025-09-04 10:51:02] [Rank 0] step:9821/10000 train_time:419801ms step_avg:42.75ms
+[2025-09-04 10:51:03] [Rank 0] step:9841/10000 train_time:420561ms step_avg:42.74ms
+[2025-09-04 10:51:03] [Rank 0] step:9841/10000 train_time:420561ms step_avg:42.74ms
+[2025-09-04 10:51:03] [Rank 0] step:9861/10000 train_time:421322ms step_avg:42.73ms
+[2025-09-04 10:51:03] [Rank 0] step:9861/10000 train_time:421322ms step_avg:42.73ms
+[2025-09-04 10:51:04] [Rank 0] step:9881/10000 train_time:422081ms step_avg:42.72ms
+[2025-09-04 10:51:04] [Rank 0] step:9881/10000 train_time:422081ms step_avg:42.72ms
+[2025-09-04 10:51:05] [Rank 0] step:9901/10000 train_time:422841ms step_avg:42.71ms
+[2025-09-04 10:51:05] [Rank 0] step:9901/10000 train_time:422841ms step_avg:42.71ms
+[2025-09-04 10:51:06] [Rank 0] step:9921/10000 train_time:423601ms step_avg:42.70ms
+[2025-09-04 10:51:06] [Rank 0] step:9921/10000 train_time:423601ms step_avg:42.70ms
+[2025-09-04 10:51:07] [Rank 0] step:9941/10000 train_time:424362ms step_avg:42.69ms
+[2025-09-04 10:51:07] [Rank 0] step:9941/10000 train_time:424362ms step_avg:42.69ms
+[2025-09-04 10:51:07] [Rank 0] step:9961/10000 train_time:425122ms step_avg:42.68ms
+[2025-09-04 10:51:07] [Rank 0] step:9961/10000 train_time:425122ms step_avg:42.68ms
+[2025-09-04 10:51:08] [Rank 0] step:9981/10000 train_time:425882ms step_avg:42.67ms
+[2025-09-04 10:51:08] [Rank 0] step:9981/10000 train_time:425882ms step_avg:42.67ms
+[2025-09-04 10:51:09] [Rank 0] step:10000/10000 train_time:426605ms step_avg:42.66ms
+[2025-09-04 10:51:09] [Rank 0] step:10000/10000 train_time:426605ms step_avg:42.66ms
+[2025-09-04 10:51:09] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:51:09] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:51:09] [Rank 0] PRINT: step:10000/10000 train_loss:0.6087 val_loss:0.6051 train_time:426653ms step_avg:42.67ms
+[2025-09-04 10:51:09] [Rank 0] PRINT: step:10000/10000 train_loss:0.6087 val_loss:0.6051 train_time:426653ms step_avg:42.67ms
+[2025-09-04 10:51:09] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:51:09] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:51:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:51:09] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:52:49] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:52:49] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:52:49] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:52:49] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:52:49] [Rank 0]   Total Loss: 5.0852
+[2025-09-04 10:52:49] [Rank 0]   Total Loss: 5.0852
+[2025-09-04 10:52:49] [Rank 0]   Total FTA (Unweighted): 0.9969
+[2025-09-04 10:52:49] [Rank 0]   Total FTA (Unweighted): 0.9969
+[2025-09-04 10:52:49] [Rank 0]   Total FTA (Weighted):   0.9969
+[2025-09-04 10:52:49] [Rank 0]   Total FTA (Weighted):   0.9969
+[2025-09-04 10:52:49] [Rank 0]   Group 0 Loss: 5.1636
+[2025-09-04 10:52:49] [Rank 0]   Group 0 Loss: 5.1636
+[2025-09-04 10:52:49] [Rank 0]   Group 1 Loss: 4.5349
+[2025-09-04 10:52:49] [Rank 0]   Group 1 Loss: 4.5349
+[2025-09-04 10:52:49] [Rank 0]   Group 2 Loss: 4.4833
+[2025-09-04 10:52:49] [Rank 0]   Group 2 Loss: 4.4833
+[2025-09-04 10:52:49] [Rank 0]   Group 3 Loss: 4.9803
+[2025-09-04 10:52:49] [Rank 0]   Group 3 Loss: 4.9803
+[2025-09-04 10:52:49] [Rank 0]   Group 4 Loss: 4.9273
+[2025-09-04 10:52:49] [Rank 0]   Group 4 Loss: 4.9273
+[2025-09-04 10:52:49] [Rank 0]   Group 5 Loss: 5.0665
+[2025-09-04 10:52:49] [Rank 0]   Group 5 Loss: 5.0665
+[2025-09-04 10:52:49] [Rank 0]   Group 6 Loss: 4.9030
+[2025-09-04 10:52:49] [Rank 0]   Group 6 Loss: 4.9030
+[2025-09-04 10:52:49] [Rank 0]   Group 7 Loss: 5.0386
+[2025-09-04 10:52:49] [Rank 0]   Group 7 Loss: 5.0386
+[2025-09-04 10:52:49] [Rank 0]   Group 8 Loss: 5.1961
+[2025-09-04 10:52:49] [Rank 0]   Group 8 Loss: 5.1961
+[2025-09-04 10:52:49] [Rank 0]   Group 9 Loss: 5.1730
+[2025-09-04 10:52:49] [Rank 0]   Group 9 Loss: 5.1730
+[2025-09-04 10:52:49] [Rank 0]   Group 10 Loss: 5.2893
+[2025-09-04 10:52:49] [Rank 0]   Group 10 Loss: 5.2893
+[2025-09-04 10:52:49] [Rank 0]   Group 11 Loss: 5.3064
+[2025-09-04 10:52:49] [Rank 0]   Group 11 Loss: 5.3064
+[2025-09-04 10:52:49] [Rank 0]   Group 12 Loss: 5.2143
+[2025-09-04 10:52:49] [Rank 0]   Group 12 Loss: 5.2143
+[2025-09-04 10:52:49] [Rank 0]   Group 13 Loss: 5.3995
+[2025-09-04 10:52:49] [Rank 0]   Group 13 Loss: 5.3995
+[2025-09-04 10:52:49] [Rank 0]   Group 14 Loss: 5.3711
+[2025-09-04 10:52:49] [Rank 0]   Group 14 Loss: 5.3711
+[2025-09-04 10:52:49] [Rank 0]   Group 15 Loss: 5.3163
+[2025-09-04 10:52:49] [Rank 0]   Group 15 Loss: 5.3163
+[2025-09-04 10:52:49] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 10:52:49] [Rank 0]   Group 15 FTA: 0.9500
+[2025-09-04 10:52:49] [Rank 0]   Group 15 FTA: 0.9500
+[2025-09-04 10:52:49] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:52:49] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_loss_curves.png
+[2025-09-04 10:52:50] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:52:50] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/per_class_acc_curves.png
+[2025-09-04 10:52:50] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:52:50] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_loss_curve.png
+[2025-09-04 10:52:50] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:52:50] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_42/total_acc_curve.png
+[2025-09-04 10:52:50] [Rank 0] step:10001/10000 train_time:426671ms step_avg:42.66ms
+[2025-09-04 10:52:50] [Rank 0] step:10001/10000 train_time:426671ms step_avg:42.66ms
+[2025-09-04 10:52:50] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 10:52:50 2025 ---
+[2025-09-04 10:52:50] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 10:52:50 2025 ---
+[2025-09-04 10:52:50] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 6248 MiB
+[2025-09-04 10:52:50] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 6248 MiB
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/config.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d3193966b6688c62394887b67f42918a1a213ed1
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/config.json
@@ -0,0 +1,29 @@
+{
+    "cli_args": {
+        "unet": false,
+        "seed": 43,
+        "optimizer_mode": 10,
+        "model_parameterization": "qkvo",
+        "per_group_k": 100,
+        "muon_lr": 0.002,
+        "adam_lr": 0.002,
+        "base_dir": "logs_qa_muon/diff_modes",
+        "sgd_lr": 0.01,
+        "m_val": 15,
+        "qa_jsonl_path": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl"
+    },
+    "hyperparameters": {
+        "train_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin",
+        "val_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin",
+        "val_tokens": 491520,
+        "train_seq_len": 3072,
+        "val_seq_len": 16384,
+        "num_iterations": 10000,
+        "cooldown_frac": 0.8,
+        "vocab_size": 50257,
+        "val_loss_every": 500,
+        "save_checkpoint": false
+    },
+    "run_uuid_for_log": "1b12ee26-82ea-490b-a501-25a607fa9186",
+    "script_code_logged_at_start": true
+}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/fixed_eval_indices.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/fixed_eval_indices.json
new file mode 100644
index 0000000000000000000000000000000000000000..a823775225c5e592eb10700e5e0319b0491b1eb6
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/fixed_eval_indices.json
@@ -0,0 +1 @@
+{"1": [1238956, 182074, 1437575, 1061037, 383150, 1176376, 926, 823011, 832520, 1266421, 512738, 144357, 848076, 890204, 213997, 95146, 261767, 467731, 832231, 217985, 913168, 107253, 1361828, 61314, 1230420, 1133619, 146690, 429587, 419151, 58695, 1579770, 503799, 1421284, 882534, 1022637, 785343, 1154604, 67783, 1325109, 243941, 1213240, 438111, 460295, 269373, 538055, 1347006, 71775, 255496, 299906, 1227973, 815402, 190082, 1304077, 1023347, 613801, 983830, 1284420, 389321, 1625224, 717538, 1172273, 992184, 1181312, 1014039, 885952, 1538489, 158933, 1667270, 1250445, 958097, 1458224, 1306495, 62945, 733843, 1360200, 540493, 762461, 501460, 1208142, 1180559, 1333588, 690481, 355756, 618511, 733586, 650301, 799437, 165533, 1238977, 323078, 1485080, 609610, 1212241, 606952, 1253407, 1420922, 327112, 701, 777907, 1626516], "0": [1390189, 1220977, 1312259, 1201125, 1235379, 1272843, 344142, 1119560, 856330, 766456, 1145928, 550624, 684681, 1309879, 54620, 1126124, 1278077, 910610, 1532044, 73645, 460165, 462723, 457959, 476877, 320873, 1408316, 989156, 1276325, 338260, 577743, 1136972, 25115, 211215, 1296818, 968705, 436781, 318224, 385325, 1516441, 533464, 1628693, 703399, 579670, 1518978, 305154, 1625960, 1400284, 713841, 1652150, 811009, 336920, 881015, 512030, 1347531, 1005706, 412167, 173312, 1136816, 343017, 537797, 1489267, 935475, 703854, 1570584, 1504269, 1458356, 1310700, 1242505, 509976, 22096, 788120, 1616850, 1591198, 1151224, 797820, 823995, 1153364, 1060220, 727674, 580729, 148912, 334290, 686098, 1633151, 1130523, 1012440, 1374480, 953410, 1381730, 1376118, 932096, 739115, 739014, 263875, 1400308, 556177, 1341771, 348626, 575350, 254846], "10": [748592, 515936, 1645691, 329828, 1104748, 1453439, 406461, 1557507, 1146644, 1398421, 163667, 228310, 407739, 585218, 785725, 1371077, 134834, 15431, 1303784, 273099, 1530332, 1320219, 1329247, 753603, 664548, 501873, 1512467, 583734, 1355643, 600440, 428998, 1426876, 121341, 1456507, 892281, 1242339, 1304508, 968336, 1593981, 634471, 1253799, 593156, 1486078, 712425, 441285, 618217, 1024661, 1395408, 1423117, 1463918, 121752, 800727, 52402, 1185143, 1651644, 1583572, 1302889, 1040220, 1234122, 333457, 1552652, 1567291, 1405501, 222708, 1469496, 875611, 894032, 783951, 934281, 109533, 198744, 643064, 197926, 224405, 1169998, 956511, 1096712, 1213224, 828780, 204659, 564225, 257079, 754287, 1598723, 821179, 1156648, 60674, 290993, 1215838, 1556010, 1025931, 687671, 252775, 771448, 574110, 578615, 279536, 1503737, 1089787, 2338], "14": [1602125, 1388977, 709154, 666138, 1446614, 572103, 99632, 25794, 1487658, 15077, 1600205, 682155, 679850, 1462049, 970283, 81336, 378196, 1627374, 832926, 1217832, 740140, 1336941, 793290, 1441127, 297494, 303372, 339524, 1363322, 525184, 593339, 61055, 1468219, 1110310, 1061855, 1046762, 410655, 449257, 530144, 1362373, 241552, 926068, 1618007, 629406, 327589, 263223, 1525290, 621372, 606355, 534601, 126046, 934024, 1181343, 1234759, 295773, 838035, 1447427, 1395837, 1568116, 486923, 465204, 98192, 1145989, 271529, 720487, 1091877, 653313, 872526, 13890, 111971, 671086, 301528, 50698, 455373, 786098, 329199, 599178, 644737, 1404606, 33829, 279256, 161938, 309352, 1537653, 21723, 309413, 478184, 340938, 1080124, 1375458, 1301528, 331423, 212589, 1220931, 112707, 454022, 1466562, 1238127, 341561, 1497264, 247245], "3": [143966, 715765, 455823, 1414629, 814361, 278667, 700211, 1286614, 1601013, 923359, 1105960, 549680, 371512, 732263, 1236795, 432942, 1379421, 488467, 679369, 1200930, 479010, 683138, 700149, 1422714, 348019, 1263383, 1044275, 820031, 969830, 75153, 1410929, 394116, 1486865, 299414, 264214, 1416493, 428688, 107076, 753671, 674417, 1456072, 307080, 1259105, 1563344, 1198484, 456551, 1061513, 1128111, 922541, 1040277, 1360208, 1411429, 259514, 883801, 716781, 1004987, 285196, 266162, 781115, 385241, 252691, 1398963, 1440582, 1300653, 756663, 231617, 1319447, 42576, 1434337, 20658, 581284, 178420, 1626533, 694079, 1531414, 1631177, 1478512, 647215, 706568, 303323, 10540, 1010611, 1150313, 159016, 1614581, 1257722, 470491, 1333309, 94699, 756222, 1456446, 296132, 1186414, 1591257, 1155050, 789439, 1472373, 1403759, 1451276, 745193], "4": [1227524, 563788, 436964, 1380176, 1648232, 267090, 273744, 1031517, 1580586, 416846, 1309866, 114170, 1649739, 1161604, 690741, 1432068, 224570, 322835, 1559902, 479718, 709757, 410142, 1668701, 1143278, 968258, 406289, 884233, 1080971, 269480, 631196, 421463, 374502, 430292, 1309826, 1422997, 1439354, 131354, 975504, 84660, 1590390, 288064, 1419944, 695652, 320713, 260376, 222416, 937739, 1104508, 1323311, 962301, 1060820, 1117857, 280792, 692359, 656651, 358878, 828406, 598701, 127434, 1511259, 1332605, 1178325, 1281187, 203589, 792701, 684176, 543034, 980077, 982509, 640126, 1252418, 1529627, 519645, 359068, 992779, 550843, 1493488, 447632, 1512599, 361760, 626672, 89042, 604248, 574827, 87735, 1432602, 473061, 668609, 925422, 311417, 71495, 682067, 221270, 600053, 673169, 1408933, 719302, 788676, 998101, 243902], "15": [1444113, 1102795, 1149896, 1181364, 349047, 1076908, 816006, 945310, 197585, 276574, 946546, 1505399, 740223, 1154813, 1312891, 1261018, 249986, 1504592, 647427, 915615, 1245136, 1288301, 802239, 410533, 805664, 1013403, 498363, 777073, 1508176, 18111, 1147340, 880289, 215773, 924760, 1301970, 318135, 1524489, 103475, 872748, 1568512, 722849, 1364157, 838956, 1451408, 970208, 1350756, 1498244, 980461, 1098677, 1141812, 463859, 95672, 1220365, 554451, 524014, 1161048, 17478, 113548, 1027741, 1657643, 1462809, 1279104, 85588, 241936, 510707, 151817, 1191227, 574942, 735662, 1046248, 1056729, 96279, 781378, 1042711, 1602339, 1503514, 299087, 898968, 553537, 293167, 897565, 572125, 425215, 916907, 1210531, 1013275, 201532, 883447, 1641610, 1185682, 425791, 831133, 1291227, 1099064, 934983, 588125, 1284591, 1163873, 524656, 1230762], "13": [1122510, 535854, 845319, 585144, 884184, 441130, 934654, 1046629, 1512962, 1190830, 1462685, 90530, 894615, 1527783, 415323, 90852, 1556905, 1002406, 366356, 1386022, 846952, 928360, 3138, 406205, 78588, 1208032, 1210206, 257177, 137047, 711380, 693628, 105893, 392053, 1301525, 677934, 1549630, 1134520, 447564, 1194061, 672969, 740755, 1045536, 633455, 1336343, 613147, 1161931, 234095, 1246643, 643373, 834221, 1351442, 1502069, 1401555, 759927, 1320337, 666063, 291222, 956633, 400583, 963313, 683474, 534882, 1101312, 1519551, 1513194, 752124, 1028953, 1383362, 169732, 785294, 160568, 1053580, 774534, 372048, 292049, 791470, 1318537, 141966, 510573, 1440830, 930544, 1352026, 463299, 1245306, 159755, 1566413, 316835, 1135200, 266962, 112892, 253603, 1328380, 703630, 21798, 268463, 7112, 1232576, 374978, 240715, 1299198], "2": [463910, 845243, 985993, 402000, 962033, 473711, 108588, 1074932, 539045, 1162793, 1086008, 778638, 751876, 53650, 465364, 803834, 820717, 9490, 473708, 26912, 970677, 1441139, 231484, 749139, 286603, 1346834, 1641988, 699767, 816768, 1220504, 978431, 1331006, 611334, 829473, 271532, 566760, 332292, 1326050, 887731, 597530, 940973, 491159, 1150471, 122010, 225465, 382413, 1207486, 1568920, 992793, 896409, 1479386, 1539347, 759024, 293232, 1345256, 1282360, 956014, 371580, 1202087, 1638038, 892501, 38073, 1651323, 1103388, 1315239, 79005, 1045045, 1230831, 1206253, 1376976, 140037, 320204, 274836, 663525, 397448, 666367, 1501713, 570453, 836424, 694423, 157979, 279074, 1235605, 536170, 301830, 778556, 1069897, 181207, 1445408, 1082031, 969552, 1083538, 1395871, 168187, 5476, 514815, 194418, 1038889, 1288748, 1535767], "8": [258910, 518286, 80108, 811775, 937854, 1623550, 1592007, 661848, 1484674, 300, 104226, 1075215, 1190104, 8885, 1062956, 88002, 1020623, 189243, 1265153, 1472857, 301412, 21371, 441401, 1078954, 1648642, 894192, 1083738, 1182805, 1347655, 316568, 1063971, 255177, 186613, 550617, 102173, 563694, 465966, 965224, 372481, 929520, 1423441, 6575, 862309, 144894, 891435, 123628, 653301, 1378790, 1247156, 1129087, 1381991, 410876, 391895, 202136, 868149, 233421, 564464, 1174972, 1385073, 238326, 483488, 507729, 1329666, 955266, 479623, 902401, 1343890, 827390, 371073, 1480060, 1241650, 1670957, 644144, 1371709, 1324243, 670126, 23715, 1566067, 70858, 522312, 558100, 301714, 359605, 1401913, 1060057, 1027855, 987700, 720084, 324937, 562646, 390910, 1398541, 1097298, 1626769, 1461630, 1430950, 650075, 1025475, 274735, 650932], "7": [181761, 629797, 231148, 365747, 1229714, 500408, 198146, 500703, 1250728, 1578820, 753672, 84476, 1379367, 355424, 1591855, 1400949, 1194404, 912060, 884197, 320928, 1096692, 714201, 331003, 1345763, 1620998, 457231, 1542074, 1108477, 1602113, 641593, 542514, 159345, 360278, 35579, 506, 470465, 1606024, 1349783, 731258, 382125, 883800, 1652746, 382071, 99241, 20216, 1398849, 245155, 1601928, 1544592, 1632383, 1597152, 1266758, 1032571, 88287, 521379, 1423306, 579207, 1152743, 271926, 418168, 8827, 1358019, 429579, 380995, 1040420, 1067138, 283335, 1150469, 280745, 1184794, 468245, 948227, 357499, 527525, 1067827, 816862, 360563, 433490, 258517, 581805, 1089641, 1537081, 473494, 1480784, 60701, 686908, 614323, 143069, 738219, 1234467, 927103, 64088, 658420, 1353941, 1517049, 753334, 198407, 434613, 850731, 1211609], "5": [1250448, 711876, 285881, 342500, 987883, 195047, 810508, 356140, 999602, 284542, 730936, 736191, 661291, 11822, 551928, 617679, 384856, 414906, 238071, 473605, 868498, 1269087, 558407, 245083, 810169, 1456559, 1657433, 70018, 865238, 416303, 1389734, 1561961, 534163, 1098802, 1494214, 1285484, 599936, 296819, 696611, 307176, 29697, 731071, 257656, 1465441, 1668863, 166995, 1668377, 1457645, 1400335, 1525120, 33323, 437285, 459069, 620176, 231704, 846424, 1391144, 364636, 63552, 1093702, 67741, 285428, 454922, 719128, 1393391, 87651, 1418561, 1312176, 652985, 1446762, 71280, 1594253, 89848, 628629, 1117344, 882598, 1014283, 1601256, 739004, 508965, 1273914, 44245, 1267971, 1430470, 611385, 930900, 1439585, 1545682, 1300966, 1331569, 364367, 27900, 107616, 1579070, 94723, 818780, 1183546, 384575, 142627, 431847], "11": [12772, 1091320, 1136723, 560663, 715094, 1097800, 1016547, 1299126, 961155, 725788, 999768, 1277062, 650665, 1323964, 1024796, 1255124, 215111, 268904, 442949, 358645, 621958, 1666210, 1154349, 375517, 614870, 1166618, 1418680, 795635, 184985, 129300, 129366, 1343289, 1151524, 193531, 634856, 1618906, 43721, 1601844, 591772, 1322095, 1522284, 431856, 574972, 1544663, 1389257, 1113893, 1515857, 614345, 809163, 132406, 95327, 997226, 423104, 910705, 261584, 765030, 1216397, 439297, 1565563, 1595785, 1622521, 975588, 1314965, 1205327, 1501612, 1642358, 656493, 1369509, 195325, 1554913, 1661675, 790468, 511387, 1227059, 1219143, 1118001, 788451, 697651, 1575073, 1345104, 167862, 1409843, 138942, 984075, 1382364, 1299538, 214601, 1541475, 459076, 353219, 1310728, 757670, 904351, 334019, 1003259, 670104, 10386, 13556, 1097931, 1526407], "6": [295849, 644897, 1207883, 1648549, 1478750, 1486750, 1156499, 621047, 1387936, 662848, 1362507, 891859, 1188350, 1298200, 1044010, 767288, 1191230, 683760, 533545, 498960, 1421819, 366846, 54389, 1382045, 1002942, 1285004, 457209, 1074763, 823805, 1508295, 719970, 837720, 19881, 1189285, 1226048, 1078656, 1433699, 14805, 254682, 614235, 805087, 900964, 343832, 561709, 243288, 465867, 497085, 1153456, 730115, 42050, 878511, 1382508, 667820, 712856, 715963, 1499782, 1516002, 444399, 726168, 1204910, 591639, 986194, 107160, 1596752, 1239812, 591854, 536489, 1347905, 1597670, 122237, 620494, 354940, 1561646, 1352115, 74785, 163277, 590349, 1429584, 1440950, 848184, 1580723, 1631128, 979479, 758861, 84655, 1079672, 673692, 1315324, 167985, 1353485, 21815, 1175739, 1360408, 1062086, 1369999, 733863, 1371964, 61279, 1152056, 220910], "9": [780278, 1156801, 827882, 1095960, 874763, 1330006, 979412, 1154745, 1399441, 515095, 1505401, 1454104, 1563117, 1316054, 57098, 43367, 1175749, 1257252, 865038, 624410, 186727, 907758, 126168, 1353431, 167806, 1021404, 1585615, 846939, 552247, 1400471, 68525, 190495, 1556857, 181528, 47677, 136468, 189446, 269049, 262467, 503841, 1367168, 382007, 486946, 621235, 622683, 469399, 1537192, 707594, 1003091, 56021, 1354302, 799290, 1618221, 696794, 836658, 1591562, 251096, 1321886, 412122, 1352217, 946767, 1267324, 213504, 747174, 1403653, 1589967, 849743, 549269, 663800, 1315189, 451416, 1330881, 1352197, 1517840, 827169, 1228725, 1097725, 955944, 1375509, 1590213, 266039, 1558465, 1498493, 604161, 608723, 1547013, 1484549, 1530909, 1325336, 1584866, 630737, 1023780, 227549, 961233, 987591, 159267, 608424, 636687, 745397, 460687], "12": [1586709, 357745, 1274743, 202991, 1250469, 837216, 757864, 185235, 493474, 621351, 22706, 1096491, 186138, 1482718, 534525, 1157131, 1350934, 1167912, 1366603, 756644, 311945, 72417, 706444, 1464146, 1236814, 74555, 983713, 1182807, 1082263, 804850, 1028788, 1405036, 932746, 742978, 1589076, 723766, 898380, 906214, 913629, 33386, 256107, 1245317, 360574, 836359, 565572, 400905, 1269191, 1042959, 886332, 1415571, 68870, 470376, 1531, 1333957, 834102, 674685, 839119, 255119, 552181, 1239034, 490552, 468715, 963242, 1612174, 21991, 1157603, 975554, 1342832, 883818, 1475260, 1666539, 1252155, 986632, 906429, 1495805, 561464, 1219097, 567704, 1463586, 1571831, 689821, 481310, 1407520, 97442, 512704, 962995, 876670, 956570, 1292534, 914981, 1655195, 870222, 196978, 617479, 1610748, 128227, 213356, 1590232, 230140, 942835]}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..6c9cb411ab100fafc86604bf0f290e60448af4ea
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2743989c658707248cfdbd892e625b9962aafa041cf838bd2818345f439fbe92
+size 341548
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..f3ba2505d0a713a9d2d8b02e1098c189c4b30f4a
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:232d426579d8540692fda393b68f9ba4294f7e0f0c044546077d6e7e3db1f686
+size 439416
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..b7835d42ba36ed3709d10cdd7b8c41a28976ad0a
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49d7f2cb13d392d4564b6b528960964bb61d5b3ce94d85385a2a2f6f0ff49ee7
+size 96621
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..cc76157cb92ad04f1d7fe8ad19ca3420c658f913
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de98b95a9f6453c6885f0a41419a234fe1d071abdca5dd6872ea3b54bebd196d
+size 114737
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/training_log_1b12ee26-82ea-490b-a501-25a607fa9186.txt b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/training_log_1b12ee26-82ea-490b-a501-25a607fa9186.txt
new file mode 100644
index 0000000000000000000000000000000000000000..88d8cd898748284818ffba37ad8f2368e719eb89
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/training_log_1b12ee26-82ea-490b-a501-25a607fa9186.txt
@@ -0,0 +1,5236 @@
+[2025-09-04 10:53:16] [Rank 0] PRINT: --- Script Start: Thu Sep  4 10:53:16 2025 ---
+[2025-09-04 10:53:16] [Rank 0] PRINT: --- Script Start: Thu Sep  4 10:53:16 2025 ---
+[2025-09-04 10:53:16] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=43, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 10:53:16] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=43, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 10:53:16] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 10:53:16] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 10:53:16] [Rank 0] PRINT: Using fixed seed: 43
+[2025-09-04 10:53:16] [Rank 0] PRINT: Using fixed seed: 43
+[2025-09-04 10:53:16] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43
+[2025-09-04 10:53:16] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43
+[2025-09-04 10:53:16] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 10:53:16] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 10:53:16] [Rank 0] PRINT: Constructing model...
+[2025-09-04 10:53:16] [Rank 0] PRINT: Constructing model...
+[2025-09-04 10:53:17] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 10:53:17] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 10:53:17] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 10:53:17] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 10:53:17] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 10:53:17] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 10:53:22] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 10:53:22] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 10:53:22] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 10:53:22] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 10:53:22] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 10:53:22] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 10:53:22] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 10:53:22] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 10:53:22] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 10:53:22] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 10:53:22] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 10:53:22] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 10:53:22] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 10:53:22] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 10:53:22] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 10:53:22] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 10:53:22] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 10:53:22] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 10:53:22] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 10:53:22] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 10:53:22] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 10:53:22] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 10:53:26] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 10:53:26] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 10:53:26] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 10:53:26] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 10:55:20] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 10:55:20] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 10:55:20] [Rank 0] PRINT: Starting training...
+[2025-09-04 10:55:20] [Rank 0] PRINT: Starting training...
+[2025-09-04 10:55:28] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/fixed_eval_indices.json
+[2025-09-04 10:55:28] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/fixed_eval_indices.json
+[2025-09-04 10:55:28] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:55:28] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:55:32] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 10:55:32] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 10:56:22] [Rank 0] step:21/10000 train_time:50481ms step_avg:2403.87ms
+[2025-09-04 10:56:22] [Rank 0] step:21/10000 train_time:50481ms step_avg:2403.87ms
+[2025-09-04 10:56:23] [Rank 0] step:41/10000 train_time:51224ms step_avg:1249.37ms
+[2025-09-04 10:56:23] [Rank 0] step:41/10000 train_time:51224ms step_avg:1249.37ms
+[2025-09-04 10:56:24] [Rank 0] step:61/10000 train_time:51965ms step_avg:851.88ms
+[2025-09-04 10:56:24] [Rank 0] step:61/10000 train_time:51965ms step_avg:851.88ms
+[2025-09-04 10:56:25] [Rank 0] step:81/10000 train_time:52705ms step_avg:650.68ms
+[2025-09-04 10:56:25] [Rank 0] step:81/10000 train_time:52705ms step_avg:650.68ms
+[2025-09-04 10:56:25] [Rank 0] step:101/10000 train_time:53446ms step_avg:529.17ms
+[2025-09-04 10:56:25] [Rank 0] step:101/10000 train_time:53446ms step_avg:529.17ms
+[2025-09-04 10:56:26] [Rank 0] step:121/10000 train_time:54188ms step_avg:447.84ms
+[2025-09-04 10:56:26] [Rank 0] step:121/10000 train_time:54188ms step_avg:447.84ms
+[2025-09-04 10:56:27] [Rank 0] step:141/10000 train_time:54929ms step_avg:389.57ms
+[2025-09-04 10:56:27] [Rank 0] step:141/10000 train_time:54929ms step_avg:389.57ms
+[2025-09-04 10:56:28] [Rank 0] step:161/10000 train_time:55670ms step_avg:345.77ms
+[2025-09-04 10:56:28] [Rank 0] step:161/10000 train_time:55670ms step_avg:345.77ms
+[2025-09-04 10:56:28] [Rank 0] step:181/10000 train_time:56410ms step_avg:311.66ms
+[2025-09-04 10:56:28] [Rank 0] step:181/10000 train_time:56410ms step_avg:311.66ms
+[2025-09-04 10:56:29] [Rank 0] step:201/10000 train_time:57152ms step_avg:284.34ms
+[2025-09-04 10:56:29] [Rank 0] step:201/10000 train_time:57152ms step_avg:284.34ms
+[2025-09-04 10:56:30] [Rank 0] step:221/10000 train_time:57892ms step_avg:261.96ms
+[2025-09-04 10:56:30] [Rank 0] step:221/10000 train_time:57892ms step_avg:261.96ms
+[2025-09-04 10:56:31] [Rank 0] step:241/10000 train_time:58634ms step_avg:243.30ms
+[2025-09-04 10:56:31] [Rank 0] step:241/10000 train_time:58634ms step_avg:243.30ms
+[2025-09-04 10:56:31] [Rank 0] step:261/10000 train_time:59383ms step_avg:227.52ms
+[2025-09-04 10:56:31] [Rank 0] step:261/10000 train_time:59383ms step_avg:227.52ms
+[2025-09-04 10:56:32] [Rank 0] step:281/10000 train_time:60123ms step_avg:213.96ms
+[2025-09-04 10:56:32] [Rank 0] step:281/10000 train_time:60123ms step_avg:213.96ms
+[2025-09-04 10:56:33] [Rank 0] step:301/10000 train_time:60863ms step_avg:202.20ms
+[2025-09-04 10:56:33] [Rank 0] step:301/10000 train_time:60863ms step_avg:202.20ms
+[2025-09-04 10:56:34] [Rank 0] step:321/10000 train_time:61605ms step_avg:191.91ms
+[2025-09-04 10:56:34] [Rank 0] step:321/10000 train_time:61605ms step_avg:191.91ms
+[2025-09-04 10:56:34] [Rank 0] step:341/10000 train_time:62345ms step_avg:182.83ms
+[2025-09-04 10:56:34] [Rank 0] step:341/10000 train_time:62345ms step_avg:182.83ms
+[2025-09-04 10:56:35] [Rank 0] step:361/10000 train_time:63085ms step_avg:174.75ms
+[2025-09-04 10:56:35] [Rank 0] step:361/10000 train_time:63085ms step_avg:174.75ms
+[2025-09-04 10:56:36] [Rank 0] step:381/10000 train_time:63825ms step_avg:167.52ms
+[2025-09-04 10:56:36] [Rank 0] step:381/10000 train_time:63825ms step_avg:167.52ms
+[2025-09-04 10:56:37] [Rank 0] step:401/10000 train_time:64566ms step_avg:161.01ms
+[2025-09-04 10:56:37] [Rank 0] step:401/10000 train_time:64566ms step_avg:161.01ms
+[2025-09-04 10:56:37] [Rank 0] step:421/10000 train_time:65306ms step_avg:155.12ms
+[2025-09-04 10:56:37] [Rank 0] step:421/10000 train_time:65306ms step_avg:155.12ms
+[2025-09-04 10:56:38] [Rank 0] step:441/10000 train_time:66046ms step_avg:149.76ms
+[2025-09-04 10:56:38] [Rank 0] step:441/10000 train_time:66046ms step_avg:149.76ms
+[2025-09-04 10:56:39] [Rank 0] step:461/10000 train_time:66786ms step_avg:144.87ms
+[2025-09-04 10:56:39] [Rank 0] step:461/10000 train_time:66786ms step_avg:144.87ms
+[2025-09-04 10:56:39] [Rank 0] step:481/10000 train_time:67526ms step_avg:140.39ms
+[2025-09-04 10:56:39] [Rank 0] step:481/10000 train_time:67526ms step_avg:140.39ms
+[2025-09-04 10:56:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:56:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:56:41] [Rank 0] PRINT: step:500/10000 train_loss:3.1254 val_loss:1.1184 train_time:68271ms step_avg:136.54ms
+[2025-09-04 10:56:41] [Rank 0] PRINT: step:500/10000 train_loss:3.1254 val_loss:1.1184 train_time:68271ms step_avg:136.54ms
+[2025-09-04 10:56:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:56:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:56:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:56:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:58:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:58:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 10:58:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:58:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 10:58:20] [Rank 0]   Total Loss: 3.7692
+[2025-09-04 10:58:20] [Rank 0]   Total Loss: 3.7692
+[2025-09-04 10:58:20] [Rank 0]   Total FTA (Unweighted): 0.4794
+[2025-09-04 10:58:20] [Rank 0]   Total FTA (Unweighted): 0.4794
+[2025-09-04 10:58:20] [Rank 0]   Total FTA (Weighted):   0.4794
+[2025-09-04 10:58:20] [Rank 0]   Total FTA (Weighted):   0.4794
+[2025-09-04 10:58:20] [Rank 0]   Group 0 Loss: 3.2265
+[2025-09-04 10:58:20] [Rank 0]   Group 0 Loss: 3.2265
+[2025-09-04 10:58:20] [Rank 0]   Group 1 Loss: 3.0094
+[2025-09-04 10:58:20] [Rank 0]   Group 1 Loss: 3.0094
+[2025-09-04 10:58:20] [Rank 0]   Group 2 Loss: 2.9946
+[2025-09-04 10:58:20] [Rank 0]   Group 2 Loss: 2.9946
+[2025-09-04 10:58:20] [Rank 0]   Group 3 Loss: 3.3067
+[2025-09-04 10:58:20] [Rank 0]   Group 3 Loss: 3.3067
+[2025-09-04 10:58:20] [Rank 0]   Group 4 Loss: 3.3812
+[2025-09-04 10:58:20] [Rank 0]   Group 4 Loss: 3.3812
+[2025-09-04 10:58:20] [Rank 0]   Group 5 Loss: 3.4676
+[2025-09-04 10:58:20] [Rank 0]   Group 5 Loss: 3.4676
+[2025-09-04 10:58:20] [Rank 0]   Group 6 Loss: 3.4968
+[2025-09-04 10:58:20] [Rank 0]   Group 6 Loss: 3.4968
+[2025-09-04 10:58:20] [Rank 0]   Group 7 Loss: 3.6298
+[2025-09-04 10:58:20] [Rank 0]   Group 7 Loss: 3.6298
+[2025-09-04 10:58:20] [Rank 0]   Group 8 Loss: 3.8887
+[2025-09-04 10:58:20] [Rank 0]   Group 8 Loss: 3.8887
+[2025-09-04 10:58:20] [Rank 0]   Group 9 Loss: 3.9599
+[2025-09-04 10:58:20] [Rank 0]   Group 9 Loss: 3.9599
+[2025-09-04 10:58:20] [Rank 0]   Group 10 Loss: 4.1775
+[2025-09-04 10:58:20] [Rank 0]   Group 10 Loss: 4.1775
+[2025-09-04 10:58:20] [Rank 0]   Group 11 Loss: 4.2352
+[2025-09-04 10:58:20] [Rank 0]   Group 11 Loss: 4.2352
+[2025-09-04 10:58:20] [Rank 0]   Group 12 Loss: 4.3183
+[2025-09-04 10:58:20] [Rank 0]   Group 12 Loss: 4.3183
+[2025-09-04 10:58:20] [Rank 0]   Group 13 Loss: 4.4205
+[2025-09-04 10:58:20] [Rank 0]   Group 13 Loss: 4.4205
+[2025-09-04 10:58:20] [Rank 0]   Group 14 Loss: 4.3868
+[2025-09-04 10:58:20] [Rank 0]   Group 14 Loss: 4.3868
+[2025-09-04 10:58:20] [Rank 0]   Group 15 Loss: 4.4073
+[2025-09-04 10:58:20] [Rank 0]   Group 15 Loss: 4.4073
+[2025-09-04 10:58:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 10:58:20] [Rank 0]   Group 4 FTA: 0.9200
+[2025-09-04 10:58:20] [Rank 0]   Group 4 FTA: 0.9200
+[2025-09-04 10:58:20] [Rank 0]   Group 5 FTA: 0.6100
+[2025-09-04 10:58:20] [Rank 0]   Group 5 FTA: 0.6100
+[2025-09-04 10:58:20] [Rank 0]   Group 6 FTA: 0.5100
+[2025-09-04 10:58:20] [Rank 0]   Group 6 FTA: 0.5100
+[2025-09-04 10:58:20] [Rank 0]   Group 7 FTA: 0.4400
+[2025-09-04 10:58:20] [Rank 0]   Group 7 FTA: 0.4400
+[2025-09-04 10:58:20] [Rank 0]   Group 8 FTA: 0.3700
+[2025-09-04 10:58:20] [Rank 0]   Group 8 FTA: 0.3700
+[2025-09-04 10:58:20] [Rank 0]   Group 9 FTA: 0.1700
+[2025-09-04 10:58:20] [Rank 0]   Group 9 FTA: 0.1700
+[2025-09-04 10:58:20] [Rank 0]   Group 10 FTA: 0.1400
+[2025-09-04 10:58:20] [Rank 0]   Group 10 FTA: 0.1400
+[2025-09-04 10:58:20] [Rank 0]   Group 11 FTA: 0.1000
+[2025-09-04 10:58:20] [Rank 0]   Group 11 FTA: 0.1000
+[2025-09-04 10:58:20] [Rank 0]   Group 12 FTA: 0.1100
+[2025-09-04 10:58:20] [Rank 0]   Group 12 FTA: 0.1100
+[2025-09-04 10:58:20] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 10:58:20] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 10:58:20] [Rank 0]   Group 14 FTA: 0.1200
+[2025-09-04 10:58:20] [Rank 0]   Group 14 FTA: 0.1200
+[2025-09-04 10:58:20] [Rank 0]   Group 15 FTA: 0.0700
+[2025-09-04 10:58:20] [Rank 0]   Group 15 FTA: 0.0700
+[2025-09-04 10:58:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 10:58:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 10:58:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 10:58:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 10:58:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 10:58:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 10:58:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 10:58:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 10:58:21] [Rank 0] step:501/10000 train_time:68287ms step_avg:136.30ms
+[2025-09-04 10:58:21] [Rank 0] step:501/10000 train_time:68287ms step_avg:136.30ms
+[2025-09-04 10:58:22] [Rank 0] step:521/10000 train_time:69039ms step_avg:132.51ms
+[2025-09-04 10:58:22] [Rank 0] step:521/10000 train_time:69039ms step_avg:132.51ms
+[2025-09-04 10:58:23] [Rank 0] step:541/10000 train_time:69781ms step_avg:128.98ms
+[2025-09-04 10:58:23] [Rank 0] step:541/10000 train_time:69781ms step_avg:128.98ms
+[2025-09-04 10:58:24] [Rank 0] step:561/10000 train_time:70522ms step_avg:125.71ms
+[2025-09-04 10:58:24] [Rank 0] step:561/10000 train_time:70522ms step_avg:125.71ms
+[2025-09-04 10:58:24] [Rank 0] step:581/10000 train_time:71263ms step_avg:122.66ms
+[2025-09-04 10:58:24] [Rank 0] step:581/10000 train_time:71263ms step_avg:122.66ms
+[2025-09-04 10:58:25] [Rank 0] step:601/10000 train_time:72005ms step_avg:119.81ms
+[2025-09-04 10:58:25] [Rank 0] step:601/10000 train_time:72005ms step_avg:119.81ms
+[2025-09-04 10:58:26] [Rank 0] step:621/10000 train_time:72746ms step_avg:117.14ms
+[2025-09-04 10:58:26] [Rank 0] step:621/10000 train_time:72746ms step_avg:117.14ms
+[2025-09-04 10:58:27] [Rank 0] step:641/10000 train_time:73487ms step_avg:114.64ms
+[2025-09-04 10:58:27] [Rank 0] step:641/10000 train_time:73487ms step_avg:114.64ms
+[2025-09-04 10:58:27] [Rank 0] step:661/10000 train_time:74227ms step_avg:112.30ms
+[2025-09-04 10:58:27] [Rank 0] step:661/10000 train_time:74227ms step_avg:112.30ms
+[2025-09-04 10:58:28] [Rank 0] step:681/10000 train_time:74968ms step_avg:110.09ms
+[2025-09-04 10:58:28] [Rank 0] step:681/10000 train_time:74968ms step_avg:110.09ms
+[2025-09-04 10:58:29] [Rank 0] step:701/10000 train_time:75709ms step_avg:108.00ms
+[2025-09-04 10:58:29] [Rank 0] step:701/10000 train_time:75709ms step_avg:108.00ms
+[2025-09-04 10:58:29] [Rank 0] step:721/10000 train_time:76450ms step_avg:106.03ms
+[2025-09-04 10:58:29] [Rank 0] step:721/10000 train_time:76450ms step_avg:106.03ms
+[2025-09-04 10:58:30] [Rank 0] step:741/10000 train_time:77190ms step_avg:104.17ms
+[2025-09-04 10:58:30] [Rank 0] step:741/10000 train_time:77190ms step_avg:104.17ms
+[2025-09-04 10:58:31] [Rank 0] step:761/10000 train_time:77934ms step_avg:102.41ms
+[2025-09-04 10:58:31] [Rank 0] step:761/10000 train_time:77934ms step_avg:102.41ms
+[2025-09-04 10:58:32] [Rank 0] step:781/10000 train_time:78678ms step_avg:100.74ms
+[2025-09-04 10:58:32] [Rank 0] step:781/10000 train_time:78678ms step_avg:100.74ms
+[2025-09-04 10:58:32] [Rank 0] step:801/10000 train_time:79422ms step_avg:99.15ms
+[2025-09-04 10:58:32] [Rank 0] step:801/10000 train_time:79422ms step_avg:99.15ms
+[2025-09-04 10:58:33] [Rank 0] step:821/10000 train_time:80238ms step_avg:97.73ms
+[2025-09-04 10:58:33] [Rank 0] step:821/10000 train_time:80238ms step_avg:97.73ms
+[2025-09-04 10:58:34] [Rank 0] step:841/10000 train_time:80982ms step_avg:96.29ms
+[2025-09-04 10:58:34] [Rank 0] step:841/10000 train_time:80982ms step_avg:96.29ms
+[2025-09-04 10:58:35] [Rank 0] step:861/10000 train_time:81727ms step_avg:94.92ms
+[2025-09-04 10:58:35] [Rank 0] step:861/10000 train_time:81727ms step_avg:94.92ms
+[2025-09-04 10:58:35] [Rank 0] step:881/10000 train_time:82471ms step_avg:93.61ms
+[2025-09-04 10:58:35] [Rank 0] step:881/10000 train_time:82471ms step_avg:93.61ms
+[2025-09-04 10:58:36] [Rank 0] step:901/10000 train_time:83217ms step_avg:92.36ms
+[2025-09-04 10:58:36] [Rank 0] step:901/10000 train_time:83217ms step_avg:92.36ms
+[2025-09-04 10:58:37] [Rank 0] step:921/10000 train_time:83962ms step_avg:91.16ms
+[2025-09-04 10:58:37] [Rank 0] step:921/10000 train_time:83962ms step_avg:91.16ms
+[2025-09-04 10:58:38] [Rank 0] step:941/10000 train_time:84706ms step_avg:90.02ms
+[2025-09-04 10:58:38] [Rank 0] step:941/10000 train_time:84706ms step_avg:90.02ms
+[2025-09-04 10:58:38] [Rank 0] step:961/10000 train_time:85451ms step_avg:88.92ms
+[2025-09-04 10:58:38] [Rank 0] step:961/10000 train_time:85451ms step_avg:88.92ms
+[2025-09-04 10:58:39] [Rank 0] step:981/10000 train_time:86195ms step_avg:87.86ms
+[2025-09-04 10:58:39] [Rank 0] step:981/10000 train_time:86195ms step_avg:87.86ms
+[2025-09-04 10:58:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:58:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 10:58:40] [Rank 0] PRINT: step:1000/10000 train_loss:0.9711 val_loss:0.8666 train_time:86945ms step_avg:86.94ms
+[2025-09-04 10:58:40] [Rank 0] PRINT: step:1000/10000 train_loss:0.9711 val_loss:0.8666 train_time:86945ms step_avg:86.94ms
+[2025-09-04 10:58:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:58:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 10:58:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 10:58:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:00:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:00:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:00:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:00:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:00:19] [Rank 0]   Total Loss: 4.0779
+[2025-09-04 11:00:19] [Rank 0]   Total Loss: 4.0779
+[2025-09-04 11:00:19] [Rank 0]   Total FTA (Unweighted): 0.6756
+[2025-09-04 11:00:19] [Rank 0]   Total FTA (Unweighted): 0.6756
+[2025-09-04 11:00:19] [Rank 0]   Total FTA (Weighted):   0.6756
+[2025-09-04 11:00:19] [Rank 0]   Total FTA (Weighted):   0.6756
+[2025-09-04 11:00:19] [Rank 0]   Group 0 Loss: 3.8697
+[2025-09-04 11:00:19] [Rank 0]   Group 0 Loss: 3.8697
+[2025-09-04 11:00:19] [Rank 0]   Group 1 Loss: 3.7263
+[2025-09-04 11:00:19] [Rank 0]   Group 1 Loss: 3.7263
+[2025-09-04 11:00:19] [Rank 0]   Group 2 Loss: 3.5180
+[2025-09-04 11:00:19] [Rank 0]   Group 2 Loss: 3.5180
+[2025-09-04 11:00:19] [Rank 0]   Group 3 Loss: 3.8247
+[2025-09-04 11:00:19] [Rank 0]   Group 3 Loss: 3.8247
+[2025-09-04 11:00:19] [Rank 0]   Group 4 Loss: 3.8075
+[2025-09-04 11:00:19] [Rank 0]   Group 4 Loss: 3.8075
+[2025-09-04 11:00:19] [Rank 0]   Group 5 Loss: 3.8343
+[2025-09-04 11:00:19] [Rank 0]   Group 5 Loss: 3.8343
+[2025-09-04 11:00:19] [Rank 0]   Group 6 Loss: 3.8060
+[2025-09-04 11:00:19] [Rank 0]   Group 6 Loss: 3.8060
+[2025-09-04 11:00:19] [Rank 0]   Group 7 Loss: 3.8481
+[2025-09-04 11:00:19] [Rank 0]   Group 7 Loss: 3.8481
+[2025-09-04 11:00:19] [Rank 0]   Group 8 Loss: 4.0263
+[2025-09-04 11:00:19] [Rank 0]   Group 8 Loss: 4.0263
+[2025-09-04 11:00:19] [Rank 0]   Group 9 Loss: 4.0138
+[2025-09-04 11:00:19] [Rank 0]   Group 9 Loss: 4.0138
+[2025-09-04 11:00:19] [Rank 0]   Group 10 Loss: 4.2521
+[2025-09-04 11:00:19] [Rank 0]   Group 10 Loss: 4.2521
+[2025-09-04 11:00:19] [Rank 0]   Group 11 Loss: 4.3567
+[2025-09-04 11:00:19] [Rank 0]   Group 11 Loss: 4.3567
+[2025-09-04 11:00:19] [Rank 0]   Group 12 Loss: 4.4341
+[2025-09-04 11:00:19] [Rank 0]   Group 12 Loss: 4.4341
+[2025-09-04 11:00:19] [Rank 0]   Group 13 Loss: 4.5886
+[2025-09-04 11:00:19] [Rank 0]   Group 13 Loss: 4.5886
+[2025-09-04 11:00:19] [Rank 0]   Group 14 Loss: 4.6155
+[2025-09-04 11:00:19] [Rank 0]   Group 14 Loss: 4.6155
+[2025-09-04 11:00:19] [Rank 0]   Group 15 Loss: 4.7247
+[2025-09-04 11:00:19] [Rank 0]   Group 15 Loss: 4.7247
+[2025-09-04 11:00:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:00:19] [Rank 0]   Group 7 FTA: 0.8700
+[2025-09-04 11:00:19] [Rank 0]   Group 7 FTA: 0.8700
+[2025-09-04 11:00:19] [Rank 0]   Group 8 FTA: 0.7700
+[2025-09-04 11:00:19] [Rank 0]   Group 8 FTA: 0.7700
+[2025-09-04 11:00:19] [Rank 0]   Group 9 FTA: 0.6400
+[2025-09-04 11:00:19] [Rank 0]   Group 9 FTA: 0.6400
+[2025-09-04 11:00:19] [Rank 0]   Group 10 FTA: 0.6300
+[2025-09-04 11:00:19] [Rank 0]   Group 10 FTA: 0.6300
+[2025-09-04 11:00:19] [Rank 0]   Group 11 FTA: 0.3200
+[2025-09-04 11:00:19] [Rank 0]   Group 11 FTA: 0.3200
+[2025-09-04 11:00:19] [Rank 0]   Group 12 FTA: 0.2200
+[2025-09-04 11:00:19] [Rank 0]   Group 12 FTA: 0.2200
+[2025-09-04 11:00:19] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 11:00:19] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 11:00:19] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 11:00:19] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 11:00:19] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 11:00:19] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 11:00:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:00:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:00:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:00:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:00:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:00:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:00:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:00:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:00:20] [Rank 0] step:1001/10000 train_time:86961ms step_avg:86.87ms
+[2025-09-04 11:00:20] [Rank 0] step:1001/10000 train_time:86961ms step_avg:86.87ms
+[2025-09-04 11:00:21] [Rank 0] step:1021/10000 train_time:87715ms step_avg:85.91ms
+[2025-09-04 11:00:21] [Rank 0] step:1021/10000 train_time:87715ms step_avg:85.91ms
+[2025-09-04 11:00:22] [Rank 0] step:1041/10000 train_time:88459ms step_avg:84.98ms
+[2025-09-04 11:00:22] [Rank 0] step:1041/10000 train_time:88459ms step_avg:84.98ms
+[2025-09-04 11:00:23] [Rank 0] step:1061/10000 train_time:89203ms step_avg:84.07ms
+[2025-09-04 11:00:23] [Rank 0] step:1061/10000 train_time:89203ms step_avg:84.07ms
+[2025-09-04 11:00:23] [Rank 0] step:1081/10000 train_time:89955ms step_avg:83.21ms
+[2025-09-04 11:00:23] [Rank 0] step:1081/10000 train_time:89955ms step_avg:83.21ms
+[2025-09-04 11:00:24] [Rank 0] step:1101/10000 train_time:90699ms step_avg:82.38ms
+[2025-09-04 11:00:24] [Rank 0] step:1101/10000 train_time:90699ms step_avg:82.38ms
+[2025-09-04 11:00:25] [Rank 0] step:1121/10000 train_time:91444ms step_avg:81.57ms
+[2025-09-04 11:00:25] [Rank 0] step:1121/10000 train_time:91444ms step_avg:81.57ms
+[2025-09-04 11:00:26] [Rank 0] step:1141/10000 train_time:92188ms step_avg:80.80ms
+[2025-09-04 11:00:26] [Rank 0] step:1141/10000 train_time:92188ms step_avg:80.80ms
+[2025-09-04 11:00:26] [Rank 0] step:1161/10000 train_time:92932ms step_avg:80.04ms
+[2025-09-04 11:00:26] [Rank 0] step:1161/10000 train_time:92932ms step_avg:80.04ms
+[2025-09-04 11:00:27] [Rank 0] step:1181/10000 train_time:93677ms step_avg:79.32ms
+[2025-09-04 11:00:27] [Rank 0] step:1181/10000 train_time:93677ms step_avg:79.32ms
+[2025-09-04 11:00:28] [Rank 0] step:1201/10000 train_time:94421ms step_avg:78.62ms
+[2025-09-04 11:00:28] [Rank 0] step:1201/10000 train_time:94421ms step_avg:78.62ms
+[2025-09-04 11:00:29] [Rank 0] step:1221/10000 train_time:95170ms step_avg:77.94ms
+[2025-09-04 11:00:29] [Rank 0] step:1221/10000 train_time:95170ms step_avg:77.94ms
+[2025-09-04 11:00:29] [Rank 0] step:1241/10000 train_time:95915ms step_avg:77.29ms
+[2025-09-04 11:00:29] [Rank 0] step:1241/10000 train_time:95915ms step_avg:77.29ms
+[2025-09-04 11:00:30] [Rank 0] step:1261/10000 train_time:96660ms step_avg:76.65ms
+[2025-09-04 11:00:30] [Rank 0] step:1261/10000 train_time:96660ms step_avg:76.65ms
+[2025-09-04 11:00:31] [Rank 0] step:1281/10000 train_time:97404ms step_avg:76.04ms
+[2025-09-04 11:00:31] [Rank 0] step:1281/10000 train_time:97404ms step_avg:76.04ms
+[2025-09-04 11:00:32] [Rank 0] step:1301/10000 train_time:98150ms step_avg:75.44ms
+[2025-09-04 11:00:32] [Rank 0] step:1301/10000 train_time:98150ms step_avg:75.44ms
+[2025-09-04 11:00:32] [Rank 0] step:1321/10000 train_time:98892ms step_avg:74.86ms
+[2025-09-04 11:00:32] [Rank 0] step:1321/10000 train_time:98892ms step_avg:74.86ms
+[2025-09-04 11:00:33] [Rank 0] step:1341/10000 train_time:99635ms step_avg:74.30ms
+[2025-09-04 11:00:33] [Rank 0] step:1341/10000 train_time:99635ms step_avg:74.30ms
+[2025-09-04 11:00:34] [Rank 0] step:1361/10000 train_time:100379ms step_avg:73.75ms
+[2025-09-04 11:00:34] [Rank 0] step:1361/10000 train_time:100379ms step_avg:73.75ms
+[2025-09-04 11:00:35] [Rank 0] step:1381/10000 train_time:101124ms step_avg:73.23ms
+[2025-09-04 11:00:35] [Rank 0] step:1381/10000 train_time:101124ms step_avg:73.23ms
+[2025-09-04 11:00:35] [Rank 0] step:1401/10000 train_time:101868ms step_avg:72.71ms
+[2025-09-04 11:00:35] [Rank 0] step:1401/10000 train_time:101868ms step_avg:72.71ms
+[2025-09-04 11:00:36] [Rank 0] step:1421/10000 train_time:102613ms step_avg:72.21ms
+[2025-09-04 11:00:36] [Rank 0] step:1421/10000 train_time:102613ms step_avg:72.21ms
+[2025-09-04 11:00:37] [Rank 0] step:1441/10000 train_time:103357ms step_avg:71.73ms
+[2025-09-04 11:00:37] [Rank 0] step:1441/10000 train_time:103357ms step_avg:71.73ms
+[2025-09-04 11:00:38] [Rank 0] step:1461/10000 train_time:104101ms step_avg:71.25ms
+[2025-09-04 11:00:38] [Rank 0] step:1461/10000 train_time:104101ms step_avg:71.25ms
+[2025-09-04 11:00:38] [Rank 0] step:1481/10000 train_time:104845ms step_avg:70.79ms
+[2025-09-04 11:00:38] [Rank 0] step:1481/10000 train_time:104845ms step_avg:70.79ms
+[2025-09-04 11:00:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:00:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:00:39] [Rank 0] PRINT: step:1500/10000 train_loss:0.8305 val_loss:0.7841 train_time:105595ms step_avg:70.40ms
+[2025-09-04 11:00:39] [Rank 0] PRINT: step:1500/10000 train_loss:0.8305 val_loss:0.7841 train_time:105595ms step_avg:70.40ms
+[2025-09-04 11:00:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:00:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:00:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:00:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:02:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:02:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:02:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:02:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:02:18] [Rank 0]   Total Loss: 4.4455
+[2025-09-04 11:02:18] [Rank 0]   Total Loss: 4.4455
+[2025-09-04 11:02:18] [Rank 0]   Total FTA (Unweighted): 0.7425
+[2025-09-04 11:02:18] [Rank 0]   Total FTA (Unweighted): 0.7425
+[2025-09-04 11:02:18] [Rank 0]   Total FTA (Weighted):   0.7425
+[2025-09-04 11:02:18] [Rank 0]   Total FTA (Weighted):   0.7425
+[2025-09-04 11:02:18] [Rank 0]   Group 0 Loss: 4.2940
+[2025-09-04 11:02:18] [Rank 0]   Group 0 Loss: 4.2940
+[2025-09-04 11:02:18] [Rank 0]   Group 1 Loss: 3.9070
+[2025-09-04 11:02:18] [Rank 0]   Group 1 Loss: 3.9070
+[2025-09-04 11:02:18] [Rank 0]   Group 2 Loss: 3.9798
+[2025-09-04 11:02:18] [Rank 0]   Group 2 Loss: 3.9798
+[2025-09-04 11:02:18] [Rank 0]   Group 3 Loss: 4.2875
+[2025-09-04 11:02:18] [Rank 0]   Group 3 Loss: 4.2875
+[2025-09-04 11:02:18] [Rank 0]   Group 4 Loss: 4.2381
+[2025-09-04 11:02:18] [Rank 0]   Group 4 Loss: 4.2381
+[2025-09-04 11:02:18] [Rank 0]   Group 5 Loss: 4.2621
+[2025-09-04 11:02:18] [Rank 0]   Group 5 Loss: 4.2621
+[2025-09-04 11:02:18] [Rank 0]   Group 6 Loss: 4.2102
+[2025-09-04 11:02:18] [Rank 0]   Group 6 Loss: 4.2102
+[2025-09-04 11:02:18] [Rank 0]   Group 7 Loss: 4.2807
+[2025-09-04 11:02:18] [Rank 0]   Group 7 Loss: 4.2807
+[2025-09-04 11:02:18] [Rank 0]   Group 8 Loss: 4.3982
+[2025-09-04 11:02:18] [Rank 0]   Group 8 Loss: 4.3982
+[2025-09-04 11:02:18] [Rank 0]   Group 9 Loss: 4.3571
+[2025-09-04 11:02:18] [Rank 0]   Group 9 Loss: 4.3571
+[2025-09-04 11:02:18] [Rank 0]   Group 10 Loss: 4.5701
+[2025-09-04 11:02:18] [Rank 0]   Group 10 Loss: 4.5701
+[2025-09-04 11:02:18] [Rank 0]   Group 11 Loss: 4.6753
+[2025-09-04 11:02:18] [Rank 0]   Group 11 Loss: 4.6753
+[2025-09-04 11:02:18] [Rank 0]   Group 12 Loss: 4.7129
+[2025-09-04 11:02:18] [Rank 0]   Group 12 Loss: 4.7129
+[2025-09-04 11:02:18] [Rank 0]   Group 13 Loss: 4.9020
+[2025-09-04 11:02:18] [Rank 0]   Group 13 Loss: 4.9020
+[2025-09-04 11:02:18] [Rank 0]   Group 14 Loss: 4.9623
+[2025-09-04 11:02:18] [Rank 0]   Group 14 Loss: 4.9623
+[2025-09-04 11:02:18] [Rank 0]   Group 15 Loss: 5.0910
+[2025-09-04 11:02:18] [Rank 0]   Group 15 Loss: 5.0910
+[2025-09-04 11:02:18] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:02:18] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:02:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:02:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:02:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:02:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:02:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:02:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:02:19] [Rank 0]   Group 8 FTA: 0.9400
+[2025-09-04 11:02:19] [Rank 0]   Group 8 FTA: 0.9400
+[2025-09-04 11:02:19] [Rank 0]   Group 9 FTA: 0.7900
+[2025-09-04 11:02:19] [Rank 0]   Group 9 FTA: 0.7900
+[2025-09-04 11:02:19] [Rank 0]   Group 10 FTA: 0.8200
+[2025-09-04 11:02:19] [Rank 0]   Group 10 FTA: 0.8200
+[2025-09-04 11:02:19] [Rank 0]   Group 11 FTA: 0.6200
+[2025-09-04 11:02:19] [Rank 0]   Group 11 FTA: 0.6200
+[2025-09-04 11:02:19] [Rank 0]   Group 12 FTA: 0.3600
+[2025-09-04 11:02:19] [Rank 0]   Group 12 FTA: 0.3600
+[2025-09-04 11:02:19] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 11:02:19] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 11:02:19] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 11:02:19] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 11:02:19] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 11:02:19] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 11:02:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:02:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:02:19] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:02:19] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:02:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:02:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:02:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:02:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:02:20] [Rank 0] step:1501/10000 train_time:105610ms step_avg:70.36ms
+[2025-09-04 11:02:20] [Rank 0] step:1501/10000 train_time:105610ms step_avg:70.36ms
+[2025-09-04 11:02:21] [Rank 0] step:1521/10000 train_time:106527ms step_avg:70.04ms
+[2025-09-04 11:02:21] [Rank 0] step:1521/10000 train_time:106527ms step_avg:70.04ms
+[2025-09-04 11:02:22] [Rank 0] step:1541/10000 train_time:107379ms step_avg:69.68ms
+[2025-09-04 11:02:22] [Rank 0] step:1541/10000 train_time:107379ms step_avg:69.68ms
+[2025-09-04 11:02:22] [Rank 0] step:1561/10000 train_time:108123ms step_avg:69.27ms
+[2025-09-04 11:02:22] [Rank 0] step:1561/10000 train_time:108123ms step_avg:69.27ms
+[2025-09-04 11:02:23] [Rank 0] step:1581/10000 train_time:109023ms step_avg:68.96ms
+[2025-09-04 11:02:23] [Rank 0] step:1581/10000 train_time:109023ms step_avg:68.96ms
+[2025-09-04 11:02:24] [Rank 0] step:1601/10000 train_time:109883ms step_avg:68.63ms
+[2025-09-04 11:02:24] [Rank 0] step:1601/10000 train_time:109883ms step_avg:68.63ms
+[2025-09-04 11:02:25] [Rank 0] step:1621/10000 train_time:110628ms step_avg:68.25ms
+[2025-09-04 11:02:25] [Rank 0] step:1621/10000 train_time:110628ms step_avg:68.25ms
+[2025-09-04 11:02:26] [Rank 0] step:1641/10000 train_time:111647ms step_avg:68.04ms
+[2025-09-04 11:02:26] [Rank 0] step:1641/10000 train_time:111647ms step_avg:68.04ms
+[2025-09-04 11:02:27] [Rank 0] step:1661/10000 train_time:112392ms step_avg:67.67ms
+[2025-09-04 11:02:27] [Rank 0] step:1661/10000 train_time:112392ms step_avg:67.67ms
+[2025-09-04 11:02:27] [Rank 0] step:1681/10000 train_time:113137ms step_avg:67.30ms
+[2025-09-04 11:02:27] [Rank 0] step:1681/10000 train_time:113137ms step_avg:67.30ms
+[2025-09-04 11:02:28] [Rank 0] step:1701/10000 train_time:113882ms step_avg:66.95ms
+[2025-09-04 11:02:28] [Rank 0] step:1701/10000 train_time:113882ms step_avg:66.95ms
+[2025-09-04 11:02:29] [Rank 0] step:1721/10000 train_time:114628ms step_avg:66.61ms
+[2025-09-04 11:02:29] [Rank 0] step:1721/10000 train_time:114628ms step_avg:66.61ms
+[2025-09-04 11:02:30] [Rank 0] step:1741/10000 train_time:115374ms step_avg:66.27ms
+[2025-09-04 11:02:30] [Rank 0] step:1741/10000 train_time:115374ms step_avg:66.27ms
+[2025-09-04 11:02:30] [Rank 0] step:1761/10000 train_time:116119ms step_avg:65.94ms
+[2025-09-04 11:02:30] [Rank 0] step:1761/10000 train_time:116119ms step_avg:65.94ms
+[2025-09-04 11:02:31] [Rank 0] step:1781/10000 train_time:116865ms step_avg:65.62ms
+[2025-09-04 11:02:31] [Rank 0] step:1781/10000 train_time:116865ms step_avg:65.62ms
+[2025-09-04 11:02:32] [Rank 0] step:1801/10000 train_time:117610ms step_avg:65.30ms
+[2025-09-04 11:02:32] [Rank 0] step:1801/10000 train_time:117610ms step_avg:65.30ms
+[2025-09-04 11:02:33] [Rank 0] step:1821/10000 train_time:118355ms step_avg:64.99ms
+[2025-09-04 11:02:33] [Rank 0] step:1821/10000 train_time:118355ms step_avg:64.99ms
+[2025-09-04 11:02:33] [Rank 0] step:1841/10000 train_time:119099ms step_avg:64.69ms
+[2025-09-04 11:02:33] [Rank 0] step:1841/10000 train_time:119099ms step_avg:64.69ms
+[2025-09-04 11:02:34] [Rank 0] step:1861/10000 train_time:119844ms step_avg:64.40ms
+[2025-09-04 11:02:34] [Rank 0] step:1861/10000 train_time:119844ms step_avg:64.40ms
+[2025-09-04 11:02:35] [Rank 0] step:1881/10000 train_time:120589ms step_avg:64.11ms
+[2025-09-04 11:02:35] [Rank 0] step:1881/10000 train_time:120589ms step_avg:64.11ms
+[2025-09-04 11:02:36] [Rank 0] step:1901/10000 train_time:121334ms step_avg:63.83ms
+[2025-09-04 11:02:36] [Rank 0] step:1901/10000 train_time:121334ms step_avg:63.83ms
+[2025-09-04 11:02:36] [Rank 0] step:1921/10000 train_time:122079ms step_avg:63.55ms
+[2025-09-04 11:02:36] [Rank 0] step:1921/10000 train_time:122079ms step_avg:63.55ms
+[2025-09-04 11:02:37] [Rank 0] step:1941/10000 train_time:122824ms step_avg:63.28ms
+[2025-09-04 11:02:37] [Rank 0] step:1941/10000 train_time:122824ms step_avg:63.28ms
+[2025-09-04 11:02:38] [Rank 0] step:1961/10000 train_time:123572ms step_avg:63.01ms
+[2025-09-04 11:02:38] [Rank 0] step:1961/10000 train_time:123572ms step_avg:63.01ms
+[2025-09-04 11:02:39] [Rank 0] step:1981/10000 train_time:124315ms step_avg:62.75ms
+[2025-09-04 11:02:39] [Rank 0] step:1981/10000 train_time:124315ms step_avg:62.75ms
+[2025-09-04 11:02:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:02:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:02:40] [Rank 0] PRINT: step:2000/10000 train_loss:0.7720 val_loss:0.7394 train_time:125066ms step_avg:62.53ms
+[2025-09-04 11:02:40] [Rank 0] PRINT: step:2000/10000 train_loss:0.7720 val_loss:0.7394 train_time:125066ms step_avg:62.53ms
+[2025-09-04 11:02:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:02:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:02:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:02:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:04:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:04:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:04:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:04:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:04:18] [Rank 0]   Total Loss: 4.5765
+[2025-09-04 11:04:18] [Rank 0]   Total Loss: 4.5765
+[2025-09-04 11:04:18] [Rank 0]   Total FTA (Unweighted): 0.7881
+[2025-09-04 11:04:18] [Rank 0]   Total FTA (Unweighted): 0.7881
+[2025-09-04 11:04:18] [Rank 0]   Total FTA (Weighted):   0.7881
+[2025-09-04 11:04:18] [Rank 0]   Total FTA (Weighted):   0.7881
+[2025-09-04 11:04:18] [Rank 0]   Group 0 Loss: 4.4342
+[2025-09-04 11:04:18] [Rank 0]   Group 0 Loss: 4.4342
+[2025-09-04 11:04:18] [Rank 0]   Group 1 Loss: 4.2094
+[2025-09-04 11:04:18] [Rank 0]   Group 1 Loss: 4.2094
+[2025-09-04 11:04:18] [Rank 0]   Group 2 Loss: 4.0184
+[2025-09-04 11:04:18] [Rank 0]   Group 2 Loss: 4.0184
+[2025-09-04 11:04:19] [Rank 0]   Group 3 Loss: 4.4606
+[2025-09-04 11:04:19] [Rank 0]   Group 3 Loss: 4.4606
+[2025-09-04 11:04:19] [Rank 0]   Group 4 Loss: 4.3613
+[2025-09-04 11:04:19] [Rank 0]   Group 4 Loss: 4.3613
+[2025-09-04 11:04:19] [Rank 0]   Group 5 Loss: 4.4238
+[2025-09-04 11:04:19] [Rank 0]   Group 5 Loss: 4.4238
+[2025-09-04 11:04:19] [Rank 0]   Group 6 Loss: 4.4010
+[2025-09-04 11:04:19] [Rank 0]   Group 6 Loss: 4.4010
+[2025-09-04 11:04:19] [Rank 0]   Group 7 Loss: 4.4168
+[2025-09-04 11:04:19] [Rank 0]   Group 7 Loss: 4.4168
+[2025-09-04 11:04:19] [Rank 0]   Group 8 Loss: 4.5773
+[2025-09-04 11:04:19] [Rank 0]   Group 8 Loss: 4.5773
+[2025-09-04 11:04:19] [Rank 0]   Group 9 Loss: 4.5370
+[2025-09-04 11:04:19] [Rank 0]   Group 9 Loss: 4.5370
+[2025-09-04 11:04:19] [Rank 0]   Group 10 Loss: 4.7113
+[2025-09-04 11:04:19] [Rank 0]   Group 10 Loss: 4.7113
+[2025-09-04 11:04:19] [Rank 0]   Group 11 Loss: 4.8052
+[2025-09-04 11:04:19] [Rank 0]   Group 11 Loss: 4.8052
+[2025-09-04 11:04:19] [Rank 0]   Group 12 Loss: 4.7870
+[2025-09-04 11:04:19] [Rank 0]   Group 12 Loss: 4.7870
+[2025-09-04 11:04:19] [Rank 0]   Group 13 Loss: 4.9818
+[2025-09-04 11:04:19] [Rank 0]   Group 13 Loss: 4.9818
+[2025-09-04 11:04:19] [Rank 0]   Group 14 Loss: 4.9830
+[2025-09-04 11:04:19] [Rank 0]   Group 14 Loss: 4.9830
+[2025-09-04 11:04:19] [Rank 0]   Group 15 Loss: 5.1158
+[2025-09-04 11:04:19] [Rank 0]   Group 15 Loss: 5.1158
+[2025-09-04 11:04:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:04:19] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 11:04:19] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 11:04:19] [Rank 0]   Group 9 FTA: 0.8700
+[2025-09-04 11:04:19] [Rank 0]   Group 9 FTA: 0.8700
+[2025-09-04 11:04:19] [Rank 0]   Group 10 FTA: 0.9200
+[2025-09-04 11:04:19] [Rank 0]   Group 10 FTA: 0.9200
+[2025-09-04 11:04:19] [Rank 0]   Group 11 FTA: 0.8100
+[2025-09-04 11:04:19] [Rank 0]   Group 11 FTA: 0.8100
+[2025-09-04 11:04:19] [Rank 0]   Group 12 FTA: 0.5100
+[2025-09-04 11:04:19] [Rank 0]   Group 12 FTA: 0.5100
+[2025-09-04 11:04:19] [Rank 0]   Group 13 FTA: 0.2200
+[2025-09-04 11:04:19] [Rank 0]   Group 13 FTA: 0.2200
+[2025-09-04 11:04:19] [Rank 0]   Group 14 FTA: 0.1700
+[2025-09-04 11:04:19] [Rank 0]   Group 14 FTA: 0.1700
+[2025-09-04 11:04:19] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:04:19] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:04:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:04:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:04:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:04:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:04:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:04:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:04:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:04:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:04:20] [Rank 0] step:2001/10000 train_time:125080ms step_avg:62.51ms
+[2025-09-04 11:04:20] [Rank 0] step:2001/10000 train_time:125080ms step_avg:62.51ms
+[2025-09-04 11:04:21] [Rank 0] step:2021/10000 train_time:126095ms step_avg:62.39ms
+[2025-09-04 11:04:21] [Rank 0] step:2021/10000 train_time:126095ms step_avg:62.39ms
+[2025-09-04 11:04:22] [Rank 0] step:2041/10000 train_time:126840ms step_avg:62.15ms
+[2025-09-04 11:04:22] [Rank 0] step:2041/10000 train_time:126840ms step_avg:62.15ms
+[2025-09-04 11:04:23] [Rank 0] step:2061/10000 train_time:127585ms step_avg:61.90ms
+[2025-09-04 11:04:23] [Rank 0] step:2061/10000 train_time:127585ms step_avg:61.90ms
+[2025-09-04 11:04:23] [Rank 0] step:2081/10000 train_time:128329ms step_avg:61.67ms
+[2025-09-04 11:04:23] [Rank 0] step:2081/10000 train_time:128329ms step_avg:61.67ms
+[2025-09-04 11:04:24] [Rank 0] step:2101/10000 train_time:129075ms step_avg:61.43ms
+[2025-09-04 11:04:24] [Rank 0] step:2101/10000 train_time:129075ms step_avg:61.43ms
+[2025-09-04 11:04:25] [Rank 0] step:2121/10000 train_time:129820ms step_avg:61.21ms
+[2025-09-04 11:04:25] [Rank 0] step:2121/10000 train_time:129820ms step_avg:61.21ms
+[2025-09-04 11:04:26] [Rank 0] step:2141/10000 train_time:130564ms step_avg:60.98ms
+[2025-09-04 11:04:26] [Rank 0] step:2141/10000 train_time:130564ms step_avg:60.98ms
+[2025-09-04 11:04:26] [Rank 0] step:2161/10000 train_time:131309ms step_avg:60.76ms
+[2025-09-04 11:04:26] [Rank 0] step:2161/10000 train_time:131309ms step_avg:60.76ms
+[2025-09-04 11:04:27] [Rank 0] step:2181/10000 train_time:132208ms step_avg:60.62ms
+[2025-09-04 11:04:27] [Rank 0] step:2181/10000 train_time:132208ms step_avg:60.62ms
+[2025-09-04 11:04:28] [Rank 0] step:2201/10000 train_time:133022ms step_avg:60.44ms
+[2025-09-04 11:04:28] [Rank 0] step:2201/10000 train_time:133022ms step_avg:60.44ms
+[2025-09-04 11:04:29] [Rank 0] step:2221/10000 train_time:133766ms step_avg:60.23ms
+[2025-09-04 11:04:29] [Rank 0] step:2221/10000 train_time:133766ms step_avg:60.23ms
+[2025-09-04 11:04:30] [Rank 0] step:2241/10000 train_time:134696ms step_avg:60.11ms
+[2025-09-04 11:04:30] [Rank 0] step:2241/10000 train_time:134696ms step_avg:60.11ms
+[2025-09-04 11:04:31] [Rank 0] step:2261/10000 train_time:135560ms step_avg:59.96ms
+[2025-09-04 11:04:31] [Rank 0] step:2261/10000 train_time:135560ms step_avg:59.96ms
+[2025-09-04 11:04:31] [Rank 0] step:2281/10000 train_time:136315ms step_avg:59.76ms
+[2025-09-04 11:04:31] [Rank 0] step:2281/10000 train_time:136315ms step_avg:59.76ms
+[2025-09-04 11:04:32] [Rank 0] step:2301/10000 train_time:137069ms step_avg:59.57ms
+[2025-09-04 11:04:32] [Rank 0] step:2301/10000 train_time:137069ms step_avg:59.57ms
+[2025-09-04 11:04:33] [Rank 0] step:2321/10000 train_time:137824ms step_avg:59.38ms
+[2025-09-04 11:04:33] [Rank 0] step:2321/10000 train_time:137824ms step_avg:59.38ms
+[2025-09-04 11:04:34] [Rank 0] step:2341/10000 train_time:138578ms step_avg:59.20ms
+[2025-09-04 11:04:34] [Rank 0] step:2341/10000 train_time:138578ms step_avg:59.20ms
+[2025-09-04 11:04:34] [Rank 0] step:2361/10000 train_time:139332ms step_avg:59.01ms
+[2025-09-04 11:04:34] [Rank 0] step:2361/10000 train_time:139332ms step_avg:59.01ms
+[2025-09-04 11:04:35] [Rank 0] step:2381/10000 train_time:140089ms step_avg:58.84ms
+[2025-09-04 11:04:35] [Rank 0] step:2381/10000 train_time:140089ms step_avg:58.84ms
+[2025-09-04 11:04:36] [Rank 0] step:2401/10000 train_time:140843ms step_avg:58.66ms
+[2025-09-04 11:04:36] [Rank 0] step:2401/10000 train_time:140843ms step_avg:58.66ms
+[2025-09-04 11:04:37] [Rank 0] step:2421/10000 train_time:141597ms step_avg:58.49ms
+[2025-09-04 11:04:37] [Rank 0] step:2421/10000 train_time:141597ms step_avg:58.49ms
+[2025-09-04 11:04:37] [Rank 0] step:2441/10000 train_time:142352ms step_avg:58.32ms
+[2025-09-04 11:04:37] [Rank 0] step:2441/10000 train_time:142352ms step_avg:58.32ms
+[2025-09-04 11:04:38] [Rank 0] step:2461/10000 train_time:143107ms step_avg:58.15ms
+[2025-09-04 11:04:38] [Rank 0] step:2461/10000 train_time:143107ms step_avg:58.15ms
+[2025-09-04 11:04:39] [Rank 0] step:2481/10000 train_time:143861ms step_avg:57.98ms
+[2025-09-04 11:04:39] [Rank 0] step:2481/10000 train_time:143861ms step_avg:57.98ms
+[2025-09-04 11:04:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:04:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:04:40] [Rank 0] PRINT: step:2500/10000 train_loss:0.7346 val_loss:0.7067 train_time:144622ms step_avg:57.85ms
+[2025-09-04 11:04:40] [Rank 0] PRINT: step:2500/10000 train_loss:0.7346 val_loss:0.7067 train_time:144622ms step_avg:57.85ms
+[2025-09-04 11:04:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:04:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:04:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:04:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:06:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:06:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:06:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:06:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:06:19] [Rank 0]   Total Loss: 4.5916
+[2025-09-04 11:06:19] [Rank 0]   Total Loss: 4.5916
+[2025-09-04 11:06:19] [Rank 0]   Total FTA (Unweighted): 0.8269
+[2025-09-04 11:06:19] [Rank 0]   Total FTA (Unweighted): 0.8269
+[2025-09-04 11:06:19] [Rank 0]   Total FTA (Weighted):   0.8269
+[2025-09-04 11:06:19] [Rank 0]   Total FTA (Weighted):   0.8269
+[2025-09-04 11:06:19] [Rank 0]   Group 0 Loss: 4.4962
+[2025-09-04 11:06:19] [Rank 0]   Group 0 Loss: 4.4962
+[2025-09-04 11:06:19] [Rank 0]   Group 1 Loss: 4.1679
+[2025-09-04 11:06:19] [Rank 0]   Group 1 Loss: 4.1679
+[2025-09-04 11:06:19] [Rank 0]   Group 2 Loss: 4.1019
+[2025-09-04 11:06:19] [Rank 0]   Group 2 Loss: 4.1019
+[2025-09-04 11:06:19] [Rank 0]   Group 3 Loss: 4.5736
+[2025-09-04 11:06:19] [Rank 0]   Group 3 Loss: 4.5736
+[2025-09-04 11:06:19] [Rank 0]   Group 4 Loss: 4.4377
+[2025-09-04 11:06:19] [Rank 0]   Group 4 Loss: 4.4377
+[2025-09-04 11:06:19] [Rank 0]   Group 5 Loss: 4.4596
+[2025-09-04 11:06:19] [Rank 0]   Group 5 Loss: 4.4596
+[2025-09-04 11:06:19] [Rank 0]   Group 6 Loss: 4.4179
+[2025-09-04 11:06:19] [Rank 0]   Group 6 Loss: 4.4179
+[2025-09-04 11:06:19] [Rank 0]   Group 7 Loss: 4.4955
+[2025-09-04 11:06:19] [Rank 0]   Group 7 Loss: 4.4955
+[2025-09-04 11:06:19] [Rank 0]   Group 8 Loss: 4.6475
+[2025-09-04 11:06:19] [Rank 0]   Group 8 Loss: 4.6475
+[2025-09-04 11:06:19] [Rank 0]   Group 9 Loss: 4.6184
+[2025-09-04 11:06:19] [Rank 0]   Group 9 Loss: 4.6184
+[2025-09-04 11:06:19] [Rank 0]   Group 10 Loss: 4.7407
+[2025-09-04 11:06:19] [Rank 0]   Group 10 Loss: 4.7407
+[2025-09-04 11:06:19] [Rank 0]   Group 11 Loss: 4.7611
+[2025-09-04 11:06:19] [Rank 0]   Group 11 Loss: 4.7611
+[2025-09-04 11:06:19] [Rank 0]   Group 12 Loss: 4.7704
+[2025-09-04 11:06:19] [Rank 0]   Group 12 Loss: 4.7704
+[2025-09-04 11:06:19] [Rank 0]   Group 13 Loss: 4.8793
+[2025-09-04 11:06:19] [Rank 0]   Group 13 Loss: 4.8793
+[2025-09-04 11:06:19] [Rank 0]   Group 14 Loss: 4.8751
+[2025-09-04 11:06:19] [Rank 0]   Group 14 Loss: 4.8751
+[2025-09-04 11:06:19] [Rank 0]   Group 15 Loss: 5.0222
+[2025-09-04 11:06:19] [Rank 0]   Group 15 Loss: 5.0222
+[2025-09-04 11:06:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:06:19] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 11:06:19] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 11:06:19] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 11:06:19] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 11:06:19] [Rank 0]   Group 11 FTA: 0.8800
+[2025-09-04 11:06:19] [Rank 0]   Group 11 FTA: 0.8800
+[2025-09-04 11:06:19] [Rank 0]   Group 12 FTA: 0.7400
+[2025-09-04 11:06:19] [Rank 0]   Group 12 FTA: 0.7400
+[2025-09-04 11:06:19] [Rank 0]   Group 13 FTA: 0.3800
+[2025-09-04 11:06:19] [Rank 0]   Group 13 FTA: 0.3800
+[2025-09-04 11:06:20] [Rank 0]   Group 14 FTA: 0.1800
+[2025-09-04 11:06:20] [Rank 0]   Group 14 FTA: 0.1800
+[2025-09-04 11:06:20] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:06:20] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:06:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:06:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:06:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:06:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:06:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:06:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:06:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:06:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:06:21] [Rank 0] step:2501/10000 train_time:144637ms step_avg:57.83ms
+[2025-09-04 11:06:21] [Rank 0] step:2501/10000 train_time:144637ms step_avg:57.83ms
+[2025-09-04 11:06:22] [Rank 0] step:2521/10000 train_time:145405ms step_avg:57.68ms
+[2025-09-04 11:06:22] [Rank 0] step:2521/10000 train_time:145405ms step_avg:57.68ms
+[2025-09-04 11:06:22] [Rank 0] step:2541/10000 train_time:146161ms step_avg:57.52ms
+[2025-09-04 11:06:22] [Rank 0] step:2541/10000 train_time:146161ms step_avg:57.52ms
+[2025-09-04 11:06:23] [Rank 0] step:2561/10000 train_time:146917ms step_avg:57.37ms
+[2025-09-04 11:06:23] [Rank 0] step:2561/10000 train_time:146917ms step_avg:57.37ms
+[2025-09-04 11:06:24] [Rank 0] step:2581/10000 train_time:147672ms step_avg:57.22ms
+[2025-09-04 11:06:24] [Rank 0] step:2581/10000 train_time:147672ms step_avg:57.22ms
+[2025-09-04 11:06:25] [Rank 0] step:2601/10000 train_time:148428ms step_avg:57.07ms
+[2025-09-04 11:06:25] [Rank 0] step:2601/10000 train_time:148428ms step_avg:57.07ms
+[2025-09-04 11:06:26] [Rank 0] step:2621/10000 train_time:149184ms step_avg:56.92ms
+[2025-09-04 11:06:26] [Rank 0] step:2621/10000 train_time:149184ms step_avg:56.92ms
+[2025-09-04 11:06:26] [Rank 0] step:2641/10000 train_time:149940ms step_avg:56.77ms
+[2025-09-04 11:06:26] [Rank 0] step:2641/10000 train_time:149940ms step_avg:56.77ms
+[2025-09-04 11:06:27] [Rank 0] step:2661/10000 train_time:150696ms step_avg:56.63ms
+[2025-09-04 11:06:27] [Rank 0] step:2661/10000 train_time:150696ms step_avg:56.63ms
+[2025-09-04 11:06:28] [Rank 0] step:2681/10000 train_time:151454ms step_avg:56.49ms
+[2025-09-04 11:06:28] [Rank 0] step:2681/10000 train_time:151454ms step_avg:56.49ms
+[2025-09-04 11:06:29] [Rank 0] step:2701/10000 train_time:152208ms step_avg:56.35ms
+[2025-09-04 11:06:29] [Rank 0] step:2701/10000 train_time:152208ms step_avg:56.35ms
+[2025-09-04 11:06:29] [Rank 0] step:2721/10000 train_time:152964ms step_avg:56.22ms
+[2025-09-04 11:06:29] [Rank 0] step:2721/10000 train_time:152964ms step_avg:56.22ms
+[2025-09-04 11:06:30] [Rank 0] step:2741/10000 train_time:153720ms step_avg:56.08ms
+[2025-09-04 11:06:30] [Rank 0] step:2741/10000 train_time:153720ms step_avg:56.08ms
+[2025-09-04 11:06:31] [Rank 0] step:2761/10000 train_time:154475ms step_avg:55.95ms
+[2025-09-04 11:06:31] [Rank 0] step:2761/10000 train_time:154475ms step_avg:55.95ms
+[2025-09-04 11:06:32] [Rank 0] step:2781/10000 train_time:155231ms step_avg:55.82ms
+[2025-09-04 11:06:32] [Rank 0] step:2781/10000 train_time:155231ms step_avg:55.82ms
+[2025-09-04 11:06:32] [Rank 0] step:2801/10000 train_time:155986ms step_avg:55.69ms
+[2025-09-04 11:06:32] [Rank 0] step:2801/10000 train_time:155986ms step_avg:55.69ms
+[2025-09-04 11:06:33] [Rank 0] step:2821/10000 train_time:157013ms step_avg:55.66ms
+[2025-09-04 11:06:33] [Rank 0] step:2821/10000 train_time:157013ms step_avg:55.66ms
+[2025-09-04 11:06:34] [Rank 0] step:2841/10000 train_time:158032ms step_avg:55.63ms
+[2025-09-04 11:06:34] [Rank 0] step:2841/10000 train_time:158032ms step_avg:55.63ms
+[2025-09-04 11:06:35] [Rank 0] step:2861/10000 train_time:158789ms step_avg:55.50ms
+[2025-09-04 11:06:35] [Rank 0] step:2861/10000 train_time:158789ms step_avg:55.50ms
+[2025-09-04 11:06:36] [Rank 0] step:2881/10000 train_time:159545ms step_avg:55.38ms
+[2025-09-04 11:06:36] [Rank 0] step:2881/10000 train_time:159545ms step_avg:55.38ms
+[2025-09-04 11:06:37] [Rank 0] step:2901/10000 train_time:160567ms step_avg:55.35ms
+[2025-09-04 11:06:37] [Rank 0] step:2901/10000 train_time:160567ms step_avg:55.35ms
+[2025-09-04 11:06:38] [Rank 0] step:2921/10000 train_time:161321ms step_avg:55.23ms
+[2025-09-04 11:06:38] [Rank 0] step:2921/10000 train_time:161321ms step_avg:55.23ms
+[2025-09-04 11:06:38] [Rank 0] step:2941/10000 train_time:162076ms step_avg:55.11ms
+[2025-09-04 11:06:38] [Rank 0] step:2941/10000 train_time:162076ms step_avg:55.11ms
+[2025-09-04 11:06:39] [Rank 0] step:2961/10000 train_time:162832ms step_avg:54.99ms
+[2025-09-04 11:06:39] [Rank 0] step:2961/10000 train_time:162832ms step_avg:54.99ms
+[2025-09-04 11:06:40] [Rank 0] step:2981/10000 train_time:163587ms step_avg:54.88ms
+[2025-09-04 11:06:40] [Rank 0] step:2981/10000 train_time:163587ms step_avg:54.88ms
+[2025-09-04 11:06:41] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:06:41] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:06:41] [Rank 0] PRINT: step:3000/10000 train_loss:0.7080 val_loss:0.6870 train_time:164348ms step_avg:54.78ms
+[2025-09-04 11:06:41] [Rank 0] PRINT: step:3000/10000 train_loss:0.7080 val_loss:0.6870 train_time:164348ms step_avg:54.78ms
+[2025-09-04 11:06:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:06:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:06:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:06:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:08:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:08:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:08:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:08:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:08:20] [Rank 0]   Total Loss: 4.7218
+[2025-09-04 11:08:20] [Rank 0]   Total Loss: 4.7218
+[2025-09-04 11:08:20] [Rank 0]   Total FTA (Unweighted): 0.8512
+[2025-09-04 11:08:20] [Rank 0]   Total FTA (Unweighted): 0.8512
+[2025-09-04 11:08:20] [Rank 0]   Total FTA (Weighted):   0.8512
+[2025-09-04 11:08:20] [Rank 0]   Total FTA (Weighted):   0.8512
+[2025-09-04 11:08:20] [Rank 0]   Group 0 Loss: 4.6590
+[2025-09-04 11:08:20] [Rank 0]   Group 0 Loss: 4.6590
+[2025-09-04 11:08:20] [Rank 0]   Group 1 Loss: 4.2478
+[2025-09-04 11:08:20] [Rank 0]   Group 1 Loss: 4.2478
+[2025-09-04 11:08:20] [Rank 0]   Group 2 Loss: 4.2383
+[2025-09-04 11:08:20] [Rank 0]   Group 2 Loss: 4.2383
+[2025-09-04 11:08:20] [Rank 0]   Group 3 Loss: 4.7007
+[2025-09-04 11:08:20] [Rank 0]   Group 3 Loss: 4.7007
+[2025-09-04 11:08:20] [Rank 0]   Group 4 Loss: 4.5103
+[2025-09-04 11:08:20] [Rank 0]   Group 4 Loss: 4.5103
+[2025-09-04 11:08:20] [Rank 0]   Group 5 Loss: 4.6328
+[2025-09-04 11:08:20] [Rank 0]   Group 5 Loss: 4.6328
+[2025-09-04 11:08:20] [Rank 0]   Group 6 Loss: 4.5663
+[2025-09-04 11:08:20] [Rank 0]   Group 6 Loss: 4.5663
+[2025-09-04 11:08:20] [Rank 0]   Group 7 Loss: 4.6264
+[2025-09-04 11:08:20] [Rank 0]   Group 7 Loss: 4.6264
+[2025-09-04 11:08:20] [Rank 0]   Group 8 Loss: 4.7605
+[2025-09-04 11:08:20] [Rank 0]   Group 8 Loss: 4.7605
+[2025-09-04 11:08:20] [Rank 0]   Group 9 Loss: 4.7803
+[2025-09-04 11:08:20] [Rank 0]   Group 9 Loss: 4.7803
+[2025-09-04 11:08:20] [Rank 0]   Group 10 Loss: 4.9145
+[2025-09-04 11:08:20] [Rank 0]   Group 10 Loss: 4.9145
+[2025-09-04 11:08:20] [Rank 0]   Group 11 Loss: 4.9143
+[2025-09-04 11:08:20] [Rank 0]   Group 11 Loss: 4.9143
+[2025-09-04 11:08:20] [Rank 0]   Group 12 Loss: 4.8958
+[2025-09-04 11:08:20] [Rank 0]   Group 12 Loss: 4.8958
+[2025-09-04 11:08:20] [Rank 0]   Group 13 Loss: 5.0304
+[2025-09-04 11:08:20] [Rank 0]   Group 13 Loss: 5.0304
+[2025-09-04 11:08:20] [Rank 0]   Group 14 Loss: 4.9833
+[2025-09-04 11:08:20] [Rank 0]   Group 14 Loss: 4.9833
+[2025-09-04 11:08:20] [Rank 0]   Group 15 Loss: 5.0877
+[2025-09-04 11:08:20] [Rank 0]   Group 15 Loss: 5.0877
+[2025-09-04 11:08:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:08:20] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 11:08:20] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 11:08:20] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 11:08:20] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 11:08:20] [Rank 0]   Group 11 FTA: 0.9700
+[2025-09-04 11:08:20] [Rank 0]   Group 11 FTA: 0.9700
+[2025-09-04 11:08:20] [Rank 0]   Group 12 FTA: 0.8800
+[2025-09-04 11:08:20] [Rank 0]   Group 12 FTA: 0.8800
+[2025-09-04 11:08:20] [Rank 0]   Group 13 FTA: 0.4600
+[2025-09-04 11:08:20] [Rank 0]   Group 13 FTA: 0.4600
+[2025-09-04 11:08:20] [Rank 0]   Group 14 FTA: 0.2200
+[2025-09-04 11:08:20] [Rank 0]   Group 14 FTA: 0.2200
+[2025-09-04 11:08:20] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:08:20] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:08:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:08:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:08:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:08:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:08:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:08:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:08:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:08:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:08:21] [Rank 0] step:3001/10000 train_time:164363ms step_avg:54.77ms
+[2025-09-04 11:08:21] [Rank 0] step:3001/10000 train_time:164363ms step_avg:54.77ms
+[2025-09-04 11:08:22] [Rank 0] step:3021/10000 train_time:165143ms step_avg:54.66ms
+[2025-09-04 11:08:22] [Rank 0] step:3021/10000 train_time:165143ms step_avg:54.66ms
+[2025-09-04 11:08:23] [Rank 0] step:3041/10000 train_time:165903ms step_avg:54.56ms
+[2025-09-04 11:08:23] [Rank 0] step:3041/10000 train_time:165903ms step_avg:54.56ms
+[2025-09-04 11:08:24] [Rank 0] step:3061/10000 train_time:166660ms step_avg:54.45ms
+[2025-09-04 11:08:24] [Rank 0] step:3061/10000 train_time:166660ms step_avg:54.45ms
+[2025-09-04 11:08:24] [Rank 0] step:3081/10000 train_time:167416ms step_avg:54.34ms
+[2025-09-04 11:08:24] [Rank 0] step:3081/10000 train_time:167416ms step_avg:54.34ms
+[2025-09-04 11:08:25] [Rank 0] step:3101/10000 train_time:168171ms step_avg:54.23ms
+[2025-09-04 11:08:25] [Rank 0] step:3101/10000 train_time:168171ms step_avg:54.23ms
+[2025-09-04 11:08:26] [Rank 0] step:3121/10000 train_time:168927ms step_avg:54.13ms
+[2025-09-04 11:08:26] [Rank 0] step:3121/10000 train_time:168927ms step_avg:54.13ms
+[2025-09-04 11:08:27] [Rank 0] step:3141/10000 train_time:169682ms step_avg:54.02ms
+[2025-09-04 11:08:27] [Rank 0] step:3141/10000 train_time:169682ms step_avg:54.02ms
+[2025-09-04 11:08:27] [Rank 0] step:3161/10000 train_time:170438ms step_avg:53.92ms
+[2025-09-04 11:08:27] [Rank 0] step:3161/10000 train_time:170438ms step_avg:53.92ms
+[2025-09-04 11:08:28] [Rank 0] step:3181/10000 train_time:171194ms step_avg:53.82ms
+[2025-09-04 11:08:28] [Rank 0] step:3181/10000 train_time:171194ms step_avg:53.82ms
+[2025-09-04 11:08:29] [Rank 0] step:3201/10000 train_time:171949ms step_avg:53.72ms
+[2025-09-04 11:08:29] [Rank 0] step:3201/10000 train_time:171949ms step_avg:53.72ms
+[2025-09-04 11:08:30] [Rank 0] step:3221/10000 train_time:172704ms step_avg:53.62ms
+[2025-09-04 11:08:30] [Rank 0] step:3221/10000 train_time:172704ms step_avg:53.62ms
+[2025-09-04 11:08:30] [Rank 0] step:3241/10000 train_time:173460ms step_avg:53.52ms
+[2025-09-04 11:08:30] [Rank 0] step:3241/10000 train_time:173460ms step_avg:53.52ms
+[2025-09-04 11:08:31] [Rank 0] step:3261/10000 train_time:174215ms step_avg:53.42ms
+[2025-09-04 11:08:31] [Rank 0] step:3261/10000 train_time:174215ms step_avg:53.42ms
+[2025-09-04 11:08:32] [Rank 0] step:3281/10000 train_time:174970ms step_avg:53.33ms
+[2025-09-04 11:08:32] [Rank 0] step:3281/10000 train_time:174970ms step_avg:53.33ms
+[2025-09-04 11:08:33] [Rank 0] step:3301/10000 train_time:175725ms step_avg:53.23ms
+[2025-09-04 11:08:33] [Rank 0] step:3301/10000 train_time:175725ms step_avg:53.23ms
+[2025-09-04 11:08:33] [Rank 0] step:3321/10000 train_time:176480ms step_avg:53.14ms
+[2025-09-04 11:08:33] [Rank 0] step:3321/10000 train_time:176480ms step_avg:53.14ms
+[2025-09-04 11:08:34] [Rank 0] step:3341/10000 train_time:177234ms step_avg:53.05ms
+[2025-09-04 11:08:34] [Rank 0] step:3341/10000 train_time:177234ms step_avg:53.05ms
+[2025-09-04 11:08:35] [Rank 0] step:3361/10000 train_time:177990ms step_avg:52.96ms
+[2025-09-04 11:08:35] [Rank 0] step:3361/10000 train_time:177990ms step_avg:52.96ms
+[2025-09-04 11:08:36] [Rank 0] step:3381/10000 train_time:178744ms step_avg:52.87ms
+[2025-09-04 11:08:36] [Rank 0] step:3381/10000 train_time:178744ms step_avg:52.87ms
+[2025-09-04 11:08:36] [Rank 0] step:3401/10000 train_time:179499ms step_avg:52.78ms
+[2025-09-04 11:08:36] [Rank 0] step:3401/10000 train_time:179499ms step_avg:52.78ms
+[2025-09-04 11:08:37] [Rank 0] step:3421/10000 train_time:180254ms step_avg:52.69ms
+[2025-09-04 11:08:37] [Rank 0] step:3421/10000 train_time:180254ms step_avg:52.69ms
+[2025-09-04 11:08:38] [Rank 0] step:3441/10000 train_time:181010ms step_avg:52.60ms
+[2025-09-04 11:08:38] [Rank 0] step:3441/10000 train_time:181010ms step_avg:52.60ms
+[2025-09-04 11:08:39] [Rank 0] step:3461/10000 train_time:181765ms step_avg:52.52ms
+[2025-09-04 11:08:39] [Rank 0] step:3461/10000 train_time:181765ms step_avg:52.52ms
+[2025-09-04 11:08:39] [Rank 0] step:3481/10000 train_time:182520ms step_avg:52.43ms
+[2025-09-04 11:08:39] [Rank 0] step:3481/10000 train_time:182520ms step_avg:52.43ms
+[2025-09-04 11:08:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:08:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:08:41] [Rank 0] PRINT: step:3500/10000 train_loss:0.6912 val_loss:0.6732 train_time:183532ms step_avg:52.44ms
+[2025-09-04 11:08:41] [Rank 0] PRINT: step:3500/10000 train_loss:0.6912 val_loss:0.6732 train_time:183532ms step_avg:52.44ms
+[2025-09-04 11:08:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:08:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:08:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:08:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:10:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:10:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:10:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:10:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:10:19] [Rank 0]   Total Loss: 4.7185
+[2025-09-04 11:10:19] [Rank 0]   Total Loss: 4.7185
+[2025-09-04 11:10:19] [Rank 0]   Total FTA (Unweighted): 0.8694
+[2025-09-04 11:10:19] [Rank 0]   Total FTA (Unweighted): 0.8694
+[2025-09-04 11:10:19] [Rank 0]   Total FTA (Weighted):   0.8694
+[2025-09-04 11:10:19] [Rank 0]   Total FTA (Weighted):   0.8694
+[2025-09-04 11:10:19] [Rank 0]   Group 0 Loss: 4.6092
+[2025-09-04 11:10:19] [Rank 0]   Group 0 Loss: 4.6092
+[2025-09-04 11:10:19] [Rank 0]   Group 1 Loss: 4.3303
+[2025-09-04 11:10:19] [Rank 0]   Group 1 Loss: 4.3303
+[2025-09-04 11:10:19] [Rank 0]   Group 2 Loss: 4.2820
+[2025-09-04 11:10:19] [Rank 0]   Group 2 Loss: 4.2820
+[2025-09-04 11:10:19] [Rank 0]   Group 3 Loss: 4.6637
+[2025-09-04 11:10:19] [Rank 0]   Group 3 Loss: 4.6637
+[2025-09-04 11:10:20] [Rank 0]   Group 4 Loss: 4.5205
+[2025-09-04 11:10:20] [Rank 0]   Group 4 Loss: 4.5205
+[2025-09-04 11:10:20] [Rank 0]   Group 5 Loss: 4.6204
+[2025-09-04 11:10:20] [Rank 0]   Group 5 Loss: 4.6204
+[2025-09-04 11:10:20] [Rank 0]   Group 6 Loss: 4.5511
+[2025-09-04 11:10:20] [Rank 0]   Group 6 Loss: 4.5511
+[2025-09-04 11:10:20] [Rank 0]   Group 7 Loss: 4.6462
+[2025-09-04 11:10:20] [Rank 0]   Group 7 Loss: 4.6462
+[2025-09-04 11:10:20] [Rank 0]   Group 8 Loss: 4.7864
+[2025-09-04 11:10:20] [Rank 0]   Group 8 Loss: 4.7864
+[2025-09-04 11:10:20] [Rank 0]   Group 9 Loss: 4.7670
+[2025-09-04 11:10:20] [Rank 0]   Group 9 Loss: 4.7670
+[2025-09-04 11:10:20] [Rank 0]   Group 10 Loss: 4.9265
+[2025-09-04 11:10:20] [Rank 0]   Group 10 Loss: 4.9265
+[2025-09-04 11:10:20] [Rank 0]   Group 11 Loss: 4.9388
+[2025-09-04 11:10:20] [Rank 0]   Group 11 Loss: 4.9388
+[2025-09-04 11:10:20] [Rank 0]   Group 12 Loss: 4.8787
+[2025-09-04 11:10:20] [Rank 0]   Group 12 Loss: 4.8787
+[2025-09-04 11:10:20] [Rank 0]   Group 13 Loss: 4.9379
+[2025-09-04 11:10:20] [Rank 0]   Group 13 Loss: 4.9379
+[2025-09-04 11:10:20] [Rank 0]   Group 14 Loss: 4.9777
+[2025-09-04 11:10:20] [Rank 0]   Group 14 Loss: 4.9777
+[2025-09-04 11:10:20] [Rank 0]   Group 15 Loss: 5.0595
+[2025-09-04 11:10:20] [Rank 0]   Group 15 Loss: 5.0595
+[2025-09-04 11:10:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:10:20] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 11:10:20] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 11:10:20] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 11:10:20] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 11:10:20] [Rank 0]   Group 12 FTA: 0.9200
+[2025-09-04 11:10:20] [Rank 0]   Group 12 FTA: 0.9200
+[2025-09-04 11:10:20] [Rank 0]   Group 13 FTA: 0.6600
+[2025-09-04 11:10:20] [Rank 0]   Group 13 FTA: 0.6600
+[2025-09-04 11:10:20] [Rank 0]   Group 14 FTA: 0.2500
+[2025-09-04 11:10:20] [Rank 0]   Group 14 FTA: 0.2500
+[2025-09-04 11:10:20] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 11:10:20] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 11:10:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:10:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:10:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:10:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:10:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:10:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:10:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:10:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:10:21] [Rank 0] step:3501/10000 train_time:183547ms step_avg:52.43ms
+[2025-09-04 11:10:21] [Rank 0] step:3501/10000 train_time:183547ms step_avg:52.43ms
+[2025-09-04 11:10:22] [Rank 0] step:3521/10000 train_time:184314ms step_avg:52.35ms
+[2025-09-04 11:10:22] [Rank 0] step:3521/10000 train_time:184314ms step_avg:52.35ms
+[2025-09-04 11:10:23] [Rank 0] step:3541/10000 train_time:185069ms step_avg:52.26ms
+[2025-09-04 11:10:23] [Rank 0] step:3541/10000 train_time:185069ms step_avg:52.26ms
+[2025-09-04 11:10:23] [Rank 0] step:3561/10000 train_time:185823ms step_avg:52.18ms
+[2025-09-04 11:10:23] [Rank 0] step:3561/10000 train_time:185823ms step_avg:52.18ms
+[2025-09-04 11:10:24] [Rank 0] step:3581/10000 train_time:186578ms step_avg:52.10ms
+[2025-09-04 11:10:24] [Rank 0] step:3581/10000 train_time:186578ms step_avg:52.10ms
+[2025-09-04 11:10:25] [Rank 0] step:3601/10000 train_time:187332ms step_avg:52.02ms
+[2025-09-04 11:10:25] [Rank 0] step:3601/10000 train_time:187332ms step_avg:52.02ms
+[2025-09-04 11:10:26] [Rank 0] step:3621/10000 train_time:188088ms step_avg:51.94ms
+[2025-09-04 11:10:26] [Rank 0] step:3621/10000 train_time:188088ms step_avg:51.94ms
+[2025-09-04 11:10:27] [Rank 0] step:3641/10000 train_time:189122ms step_avg:51.94ms
+[2025-09-04 11:10:27] [Rank 0] step:3641/10000 train_time:189122ms step_avg:51.94ms
+[2025-09-04 11:10:27] [Rank 0] step:3661/10000 train_time:189876ms step_avg:51.86ms
+[2025-09-04 11:10:27] [Rank 0] step:3661/10000 train_time:189876ms step_avg:51.86ms
+[2025-09-04 11:10:28] [Rank 0] step:3681/10000 train_time:190631ms step_avg:51.79ms
+[2025-09-04 11:10:28] [Rank 0] step:3681/10000 train_time:190631ms step_avg:51.79ms
+[2025-09-04 11:10:29] [Rank 0] step:3701/10000 train_time:191387ms step_avg:51.71ms
+[2025-09-04 11:10:29] [Rank 0] step:3701/10000 train_time:191387ms step_avg:51.71ms
+[2025-09-04 11:10:30] [Rank 0] step:3721/10000 train_time:192143ms step_avg:51.64ms
+[2025-09-04 11:10:30] [Rank 0] step:3721/10000 train_time:192143ms step_avg:51.64ms
+[2025-09-04 11:10:31] [Rank 0] step:3741/10000 train_time:192898ms step_avg:51.56ms
+[2025-09-04 11:10:31] [Rank 0] step:3741/10000 train_time:192898ms step_avg:51.56ms
+[2025-09-04 11:10:31] [Rank 0] step:3761/10000 train_time:193652ms step_avg:51.49ms
+[2025-09-04 11:10:31] [Rank 0] step:3761/10000 train_time:193652ms step_avg:51.49ms
+[2025-09-04 11:10:32] [Rank 0] step:3781/10000 train_time:194407ms step_avg:51.42ms
+[2025-09-04 11:10:32] [Rank 0] step:3781/10000 train_time:194407ms step_avg:51.42ms
+[2025-09-04 11:10:33] [Rank 0] step:3801/10000 train_time:195162ms step_avg:51.34ms
+[2025-09-04 11:10:33] [Rank 0] step:3801/10000 train_time:195162ms step_avg:51.34ms
+[2025-09-04 11:10:34] [Rank 0] step:3821/10000 train_time:195917ms step_avg:51.27ms
+[2025-09-04 11:10:34] [Rank 0] step:3821/10000 train_time:195917ms step_avg:51.27ms
+[2025-09-04 11:10:34] [Rank 0] step:3841/10000 train_time:196671ms step_avg:51.20ms
+[2025-09-04 11:10:34] [Rank 0] step:3841/10000 train_time:196671ms step_avg:51.20ms
+[2025-09-04 11:10:35] [Rank 0] step:3861/10000 train_time:197427ms step_avg:51.13ms
+[2025-09-04 11:10:35] [Rank 0] step:3861/10000 train_time:197427ms step_avg:51.13ms
+[2025-09-04 11:10:36] [Rank 0] step:3881/10000 train_time:198181ms step_avg:51.06ms
+[2025-09-04 11:10:36] [Rank 0] step:3881/10000 train_time:198181ms step_avg:51.06ms
+[2025-09-04 11:10:37] [Rank 0] step:3901/10000 train_time:198940ms step_avg:51.00ms
+[2025-09-04 11:10:37] [Rank 0] step:3901/10000 train_time:198940ms step_avg:51.00ms
+[2025-09-04 11:10:37] [Rank 0] step:3921/10000 train_time:199695ms step_avg:50.93ms
+[2025-09-04 11:10:37] [Rank 0] step:3921/10000 train_time:199695ms step_avg:50.93ms
+[2025-09-04 11:10:38] [Rank 0] step:3941/10000 train_time:200454ms step_avg:50.86ms
+[2025-09-04 11:10:38] [Rank 0] step:3941/10000 train_time:200454ms step_avg:50.86ms
+[2025-09-04 11:10:39] [Rank 0] step:3961/10000 train_time:201208ms step_avg:50.80ms
+[2025-09-04 11:10:39] [Rank 0] step:3961/10000 train_time:201208ms step_avg:50.80ms
+[2025-09-04 11:10:40] [Rank 0] step:3981/10000 train_time:201963ms step_avg:50.73ms
+[2025-09-04 11:10:40] [Rank 0] step:3981/10000 train_time:201963ms step_avg:50.73ms
+[2025-09-04 11:10:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:10:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:10:41] [Rank 0] PRINT: step:4000/10000 train_loss:0.6789 val_loss:0.6618 train_time:202723ms step_avg:50.68ms
+[2025-09-04 11:10:41] [Rank 0] PRINT: step:4000/10000 train_loss:0.6789 val_loss:0.6618 train_time:202723ms step_avg:50.68ms
+[2025-09-04 11:10:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:10:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:10:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:10:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:12:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:12:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:12:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:12:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:12:19] [Rank 0]   Total Loss: 4.7978
+[2025-09-04 11:12:19] [Rank 0]   Total Loss: 4.7978
+[2025-09-04 11:12:19] [Rank 0]   Total FTA (Unweighted): 0.8937
+[2025-09-04 11:12:19] [Rank 0]   Total FTA (Unweighted): 0.8937
+[2025-09-04 11:12:19] [Rank 0]   Total FTA (Weighted):   0.8938
+[2025-09-04 11:12:19] [Rank 0]   Total FTA (Weighted):   0.8938
+[2025-09-04 11:12:19] [Rank 0]   Group 0 Loss: 4.7143
+[2025-09-04 11:12:19] [Rank 0]   Group 0 Loss: 4.7143
+[2025-09-04 11:12:19] [Rank 0]   Group 1 Loss: 4.3112
+[2025-09-04 11:12:19] [Rank 0]   Group 1 Loss: 4.3112
+[2025-09-04 11:12:19] [Rank 0]   Group 2 Loss: 4.2675
+[2025-09-04 11:12:19] [Rank 0]   Group 2 Loss: 4.2675
+[2025-09-04 11:12:19] [Rank 0]   Group 3 Loss: 4.7611
+[2025-09-04 11:12:19] [Rank 0]   Group 3 Loss: 4.7611
+[2025-09-04 11:12:19] [Rank 0]   Group 4 Loss: 4.6681
+[2025-09-04 11:12:19] [Rank 0]   Group 4 Loss: 4.6681
+[2025-09-04 11:12:19] [Rank 0]   Group 5 Loss: 4.6856
+[2025-09-04 11:12:19] [Rank 0]   Group 5 Loss: 4.6856
+[2025-09-04 11:12:19] [Rank 0]   Group 6 Loss: 4.6603
+[2025-09-04 11:12:19] [Rank 0]   Group 6 Loss: 4.6603
+[2025-09-04 11:12:19] [Rank 0]   Group 7 Loss: 4.7200
+[2025-09-04 11:12:19] [Rank 0]   Group 7 Loss: 4.7200
+[2025-09-04 11:12:19] [Rank 0]   Group 8 Loss: 4.8879
+[2025-09-04 11:12:19] [Rank 0]   Group 8 Loss: 4.8879
+[2025-09-04 11:12:19] [Rank 0]   Group 9 Loss: 4.8554
+[2025-09-04 11:12:19] [Rank 0]   Group 9 Loss: 4.8554
+[2025-09-04 11:12:19] [Rank 0]   Group 10 Loss: 5.0055
+[2025-09-04 11:12:19] [Rank 0]   Group 10 Loss: 5.0055
+[2025-09-04 11:12:19] [Rank 0]   Group 11 Loss: 5.0251
+[2025-09-04 11:12:19] [Rank 0]   Group 11 Loss: 5.0251
+[2025-09-04 11:12:19] [Rank 0]   Group 12 Loss: 4.9761
+[2025-09-04 11:12:19] [Rank 0]   Group 12 Loss: 4.9761
+[2025-09-04 11:12:19] [Rank 0]   Group 13 Loss: 5.0530
+[2025-09-04 11:12:19] [Rank 0]   Group 13 Loss: 5.0530
+[2025-09-04 11:12:19] [Rank 0]   Group 14 Loss: 5.0425
+[2025-09-04 11:12:19] [Rank 0]   Group 14 Loss: 5.0425
+[2025-09-04 11:12:19] [Rank 0]   Group 15 Loss: 5.1313
+[2025-09-04 11:12:19] [Rank 0]   Group 15 Loss: 5.1313
+[2025-09-04 11:12:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:12:19] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 11:12:19] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 11:12:19] [Rank 0]   Group 12 FTA: 0.9700
+[2025-09-04 11:12:19] [Rank 0]   Group 12 FTA: 0.9700
+[2025-09-04 11:12:19] [Rank 0]   Group 13 FTA: 0.7600
+[2025-09-04 11:12:19] [Rank 0]   Group 13 FTA: 0.7600
+[2025-09-04 11:12:19] [Rank 0]   Group 14 FTA: 0.4000
+[2025-09-04 11:12:19] [Rank 0]   Group 14 FTA: 0.4000
+[2025-09-04 11:12:19] [Rank 0]   Group 15 FTA: 0.1800
+[2025-09-04 11:12:19] [Rank 0]   Group 15 FTA: 0.1800
+[2025-09-04 11:12:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:12:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:12:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:12:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:12:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:12:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:12:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:12:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:12:21] [Rank 0] step:4001/10000 train_time:202738ms step_avg:50.67ms
+[2025-09-04 11:12:21] [Rank 0] step:4001/10000 train_time:202738ms step_avg:50.67ms
+[2025-09-04 11:12:22] [Rank 0] step:4021/10000 train_time:203762ms step_avg:50.67ms
+[2025-09-04 11:12:22] [Rank 0] step:4021/10000 train_time:203762ms step_avg:50.67ms
+[2025-09-04 11:12:23] [Rank 0] step:4041/10000 train_time:204516ms step_avg:50.61ms
+[2025-09-04 11:12:23] [Rank 0] step:4041/10000 train_time:204516ms step_avg:50.61ms
+[2025-09-04 11:12:24] [Rank 0] step:4061/10000 train_time:205271ms step_avg:50.55ms
+[2025-09-04 11:12:24] [Rank 0] step:4061/10000 train_time:205271ms step_avg:50.55ms
+[2025-09-04 11:12:25] [Rank 0] step:4081/10000 train_time:206026ms step_avg:50.48ms
+[2025-09-04 11:12:25] [Rank 0] step:4081/10000 train_time:206026ms step_avg:50.48ms
+[2025-09-04 11:12:25] [Rank 0] step:4101/10000 train_time:206781ms step_avg:50.42ms
+[2025-09-04 11:12:25] [Rank 0] step:4101/10000 train_time:206781ms step_avg:50.42ms
+[2025-09-04 11:12:26] [Rank 0] step:4121/10000 train_time:207537ms step_avg:50.36ms
+[2025-09-04 11:12:26] [Rank 0] step:4121/10000 train_time:207537ms step_avg:50.36ms
+[2025-09-04 11:12:27] [Rank 0] step:4141/10000 train_time:208290ms step_avg:50.30ms
+[2025-09-04 11:12:27] [Rank 0] step:4141/10000 train_time:208290ms step_avg:50.30ms
+[2025-09-04 11:12:28] [Rank 0] step:4161/10000 train_time:209045ms step_avg:50.24ms
+[2025-09-04 11:12:28] [Rank 0] step:4161/10000 train_time:209045ms step_avg:50.24ms
+[2025-09-04 11:12:28] [Rank 0] step:4181/10000 train_time:209800ms step_avg:50.18ms
+[2025-09-04 11:12:28] [Rank 0] step:4181/10000 train_time:209800ms step_avg:50.18ms
+[2025-09-04 11:12:29] [Rank 0] step:4201/10000 train_time:210554ms step_avg:50.12ms
+[2025-09-04 11:12:29] [Rank 0] step:4201/10000 train_time:210554ms step_avg:50.12ms
+[2025-09-04 11:12:30] [Rank 0] step:4221/10000 train_time:211310ms step_avg:50.06ms
+[2025-09-04 11:12:30] [Rank 0] step:4221/10000 train_time:211310ms step_avg:50.06ms
+[2025-09-04 11:12:31] [Rank 0] step:4241/10000 train_time:212066ms step_avg:50.00ms
+[2025-09-04 11:12:31] [Rank 0] step:4241/10000 train_time:212066ms step_avg:50.00ms
+[2025-09-04 11:12:31] [Rank 0] step:4261/10000 train_time:212820ms step_avg:49.95ms
+[2025-09-04 11:12:31] [Rank 0] step:4261/10000 train_time:212820ms step_avg:49.95ms
+[2025-09-04 11:12:32] [Rank 0] step:4281/10000 train_time:213574ms step_avg:49.89ms
+[2025-09-04 11:12:32] [Rank 0] step:4281/10000 train_time:213574ms step_avg:49.89ms
+[2025-09-04 11:12:33] [Rank 0] step:4301/10000 train_time:214328ms step_avg:49.83ms
+[2025-09-04 11:12:33] [Rank 0] step:4301/10000 train_time:214328ms step_avg:49.83ms
+[2025-09-04 11:12:34] [Rank 0] step:4321/10000 train_time:215083ms step_avg:49.78ms
+[2025-09-04 11:12:34] [Rank 0] step:4321/10000 train_time:215083ms step_avg:49.78ms
+[2025-09-04 11:12:34] [Rank 0] step:4341/10000 train_time:215837ms step_avg:49.72ms
+[2025-09-04 11:12:34] [Rank 0] step:4341/10000 train_time:215837ms step_avg:49.72ms
+[2025-09-04 11:12:35] [Rank 0] step:4361/10000 train_time:216592ms step_avg:49.67ms
+[2025-09-04 11:12:35] [Rank 0] step:4361/10000 train_time:216592ms step_avg:49.67ms
+[2025-09-04 11:12:36] [Rank 0] step:4381/10000 train_time:217347ms step_avg:49.61ms
+[2025-09-04 11:12:36] [Rank 0] step:4381/10000 train_time:217347ms step_avg:49.61ms
+[2025-09-04 11:12:37] [Rank 0] step:4401/10000 train_time:218101ms step_avg:49.56ms
+[2025-09-04 11:12:37] [Rank 0] step:4401/10000 train_time:218101ms step_avg:49.56ms
+[2025-09-04 11:12:37] [Rank 0] step:4421/10000 train_time:218856ms step_avg:49.50ms
+[2025-09-04 11:12:37] [Rank 0] step:4421/10000 train_time:218856ms step_avg:49.50ms
+[2025-09-04 11:12:38] [Rank 0] step:4441/10000 train_time:219611ms step_avg:49.45ms
+[2025-09-04 11:12:38] [Rank 0] step:4441/10000 train_time:219611ms step_avg:49.45ms
+[2025-09-04 11:12:39] [Rank 0] step:4461/10000 train_time:220366ms step_avg:49.40ms
+[2025-09-04 11:12:39] [Rank 0] step:4461/10000 train_time:220366ms step_avg:49.40ms
+[2025-09-04 11:12:40] [Rank 0] step:4481/10000 train_time:221121ms step_avg:49.35ms
+[2025-09-04 11:12:40] [Rank 0] step:4481/10000 train_time:221121ms step_avg:49.35ms
+[2025-09-04 11:12:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:12:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:12:41] [Rank 0] PRINT: step:4500/10000 train_loss:0.6686 val_loss:0.6521 train_time:221881ms step_avg:49.31ms
+[2025-09-04 11:12:41] [Rank 0] PRINT: step:4500/10000 train_loss:0.6686 val_loss:0.6521 train_time:221881ms step_avg:49.31ms
+[2025-09-04 11:12:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:12:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:12:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:12:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:14:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:14:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:14:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:14:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:14:19] [Rank 0]   Total Loss: 4.8643
+[2025-09-04 11:14:19] [Rank 0]   Total Loss: 4.8643
+[2025-09-04 11:14:19] [Rank 0]   Total FTA (Unweighted): 0.9106
+[2025-09-04 11:14:19] [Rank 0]   Total FTA (Unweighted): 0.9106
+[2025-09-04 11:14:19] [Rank 0]   Total FTA (Weighted):   0.9106
+[2025-09-04 11:14:19] [Rank 0]   Total FTA (Weighted):   0.9106
+[2025-09-04 11:14:19] [Rank 0]   Group 0 Loss: 4.7578
+[2025-09-04 11:14:19] [Rank 0]   Group 0 Loss: 4.7578
+[2025-09-04 11:14:19] [Rank 0]   Group 1 Loss: 4.3758
+[2025-09-04 11:14:19] [Rank 0]   Group 1 Loss: 4.3758
+[2025-09-04 11:14:19] [Rank 0]   Group 2 Loss: 4.3686
+[2025-09-04 11:14:19] [Rank 0]   Group 2 Loss: 4.3686
+[2025-09-04 11:14:19] [Rank 0]   Group 3 Loss: 4.7740
+[2025-09-04 11:14:19] [Rank 0]   Group 3 Loss: 4.7740
+[2025-09-04 11:14:19] [Rank 0]   Group 4 Loss: 4.6920
+[2025-09-04 11:14:19] [Rank 0]   Group 4 Loss: 4.6920
+[2025-09-04 11:14:19] [Rank 0]   Group 5 Loss: 4.7919
+[2025-09-04 11:14:19] [Rank 0]   Group 5 Loss: 4.7919
+[2025-09-04 11:14:19] [Rank 0]   Group 6 Loss: 4.7080
+[2025-09-04 11:14:19] [Rank 0]   Group 6 Loss: 4.7080
+[2025-09-04 11:14:19] [Rank 0]   Group 7 Loss: 4.8236
+[2025-09-04 11:14:19] [Rank 0]   Group 7 Loss: 4.8236
+[2025-09-04 11:14:19] [Rank 0]   Group 8 Loss: 4.9562
+[2025-09-04 11:14:19] [Rank 0]   Group 8 Loss: 4.9562
+[2025-09-04 11:14:19] [Rank 0]   Group 9 Loss: 4.9233
+[2025-09-04 11:14:19] [Rank 0]   Group 9 Loss: 4.9233
+[2025-09-04 11:14:19] [Rank 0]   Group 10 Loss: 5.0917
+[2025-09-04 11:14:19] [Rank 0]   Group 10 Loss: 5.0917
+[2025-09-04 11:14:19] [Rank 0]   Group 11 Loss: 5.0794
+[2025-09-04 11:14:19] [Rank 0]   Group 11 Loss: 5.0794
+[2025-09-04 11:14:19] [Rank 0]   Group 12 Loss: 5.0762
+[2025-09-04 11:14:19] [Rank 0]   Group 12 Loss: 5.0762
+[2025-09-04 11:14:19] [Rank 0]   Group 13 Loss: 5.0978
+[2025-09-04 11:14:19] [Rank 0]   Group 13 Loss: 5.0978
+[2025-09-04 11:14:19] [Rank 0]   Group 14 Loss: 5.1189
+[2025-09-04 11:14:19] [Rank 0]   Group 14 Loss: 5.1189
+[2025-09-04 11:14:19] [Rank 0]   Group 15 Loss: 5.1930
+[2025-09-04 11:14:19] [Rank 0]   Group 15 Loss: 5.1930
+[2025-09-04 11:14:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:14:19] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 11:14:19] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 11:14:19] [Rank 0]   Group 13 FTA: 0.8700
+[2025-09-04 11:14:19] [Rank 0]   Group 13 FTA: 0.8700
+[2025-09-04 11:14:19] [Rank 0]   Group 14 FTA: 0.4800
+[2025-09-04 11:14:19] [Rank 0]   Group 14 FTA: 0.4800
+[2025-09-04 11:14:19] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 11:14:19] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 11:14:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:14:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:14:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:14:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:14:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:14:20] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:14:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:14:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:14:21] [Rank 0] step:4501/10000 train_time:221897ms step_avg:49.30ms
+[2025-09-04 11:14:21] [Rank 0] step:4501/10000 train_time:221897ms step_avg:49.30ms
+[2025-09-04 11:14:22] [Rank 0] step:4521/10000 train_time:222654ms step_avg:49.25ms
+[2025-09-04 11:14:22] [Rank 0] step:4521/10000 train_time:222654ms step_avg:49.25ms
+[2025-09-04 11:14:22] [Rank 0] step:4541/10000 train_time:223409ms step_avg:49.20ms
+[2025-09-04 11:14:22] [Rank 0] step:4541/10000 train_time:223409ms step_avg:49.20ms
+[2025-09-04 11:14:23] [Rank 0] step:4561/10000 train_time:224165ms step_avg:49.15ms
+[2025-09-04 11:14:23] [Rank 0] step:4561/10000 train_time:224165ms step_avg:49.15ms
+[2025-09-04 11:14:24] [Rank 0] step:4581/10000 train_time:224919ms step_avg:49.10ms
+[2025-09-04 11:14:24] [Rank 0] step:4581/10000 train_time:224919ms step_avg:49.10ms
+[2025-09-04 11:14:25] [Rank 0] step:4601/10000 train_time:225674ms step_avg:49.05ms
+[2025-09-04 11:14:25] [Rank 0] step:4601/10000 train_time:225674ms step_avg:49.05ms
+[2025-09-04 11:14:25] [Rank 0] step:4621/10000 train_time:226429ms step_avg:49.00ms
+[2025-09-04 11:14:25] [Rank 0] step:4621/10000 train_time:226429ms step_avg:49.00ms
+[2025-09-04 11:14:26] [Rank 0] step:4641/10000 train_time:227184ms step_avg:48.95ms
+[2025-09-04 11:14:26] [Rank 0] step:4641/10000 train_time:227184ms step_avg:48.95ms
+[2025-09-04 11:14:27] [Rank 0] step:4661/10000 train_time:227938ms step_avg:48.90ms
+[2025-09-04 11:14:27] [Rank 0] step:4661/10000 train_time:227938ms step_avg:48.90ms
+[2025-09-04 11:14:28] [Rank 0] step:4681/10000 train_time:228694ms step_avg:48.86ms
+[2025-09-04 11:14:28] [Rank 0] step:4681/10000 train_time:228694ms step_avg:48.86ms
+[2025-09-04 11:14:28] [Rank 0] step:4701/10000 train_time:229449ms step_avg:48.81ms
+[2025-09-04 11:14:28] [Rank 0] step:4701/10000 train_time:229449ms step_avg:48.81ms
+[2025-09-04 11:14:29] [Rank 0] step:4721/10000 train_time:230203ms step_avg:48.76ms
+[2025-09-04 11:14:29] [Rank 0] step:4721/10000 train_time:230203ms step_avg:48.76ms
+[2025-09-04 11:14:30] [Rank 0] step:4741/10000 train_time:230958ms step_avg:48.72ms
+[2025-09-04 11:14:30] [Rank 0] step:4741/10000 train_time:230958ms step_avg:48.72ms
+[2025-09-04 11:14:31] [Rank 0] step:4761/10000 train_time:231713ms step_avg:48.67ms
+[2025-09-04 11:14:31] [Rank 0] step:4761/10000 train_time:231713ms step_avg:48.67ms
+[2025-09-04 11:14:31] [Rank 0] step:4781/10000 train_time:232469ms step_avg:48.62ms
+[2025-09-04 11:14:31] [Rank 0] step:4781/10000 train_time:232469ms step_avg:48.62ms
+[2025-09-04 11:14:32] [Rank 0] step:4801/10000 train_time:233224ms step_avg:48.58ms
+[2025-09-04 11:14:32] [Rank 0] step:4801/10000 train_time:233224ms step_avg:48.58ms
+[2025-09-04 11:14:33] [Rank 0] step:4821/10000 train_time:233979ms step_avg:48.53ms
+[2025-09-04 11:14:33] [Rank 0] step:4821/10000 train_time:233979ms step_avg:48.53ms
+[2025-09-04 11:14:34] [Rank 0] step:4841/10000 train_time:235046ms step_avg:48.55ms
+[2025-09-04 11:14:34] [Rank 0] step:4841/10000 train_time:235046ms step_avg:48.55ms
+[2025-09-04 11:14:35] [Rank 0] step:4861/10000 train_time:235801ms step_avg:48.51ms
+[2025-09-04 11:14:35] [Rank 0] step:4861/10000 train_time:235801ms step_avg:48.51ms
+[2025-09-04 11:14:35] [Rank 0] step:4881/10000 train_time:236556ms step_avg:48.46ms
+[2025-09-04 11:14:35] [Rank 0] step:4881/10000 train_time:236556ms step_avg:48.46ms
+[2025-09-04 11:14:36] [Rank 0] step:4901/10000 train_time:237311ms step_avg:48.42ms
+[2025-09-04 11:14:36] [Rank 0] step:4901/10000 train_time:237311ms step_avg:48.42ms
+[2025-09-04 11:14:37] [Rank 0] step:4921/10000 train_time:238066ms step_avg:48.38ms
+[2025-09-04 11:14:37] [Rank 0] step:4921/10000 train_time:238066ms step_avg:48.38ms
+[2025-09-04 11:14:38] [Rank 0] step:4941/10000 train_time:238822ms step_avg:48.33ms
+[2025-09-04 11:14:38] [Rank 0] step:4941/10000 train_time:238822ms step_avg:48.33ms
+[2025-09-04 11:14:39] [Rank 0] step:4961/10000 train_time:239578ms step_avg:48.29ms
+[2025-09-04 11:14:39] [Rank 0] step:4961/10000 train_time:239578ms step_avg:48.29ms
+[2025-09-04 11:14:39] [Rank 0] step:4981/10000 train_time:240334ms step_avg:48.25ms
+[2025-09-04 11:14:39] [Rank 0] step:4981/10000 train_time:240334ms step_avg:48.25ms
+[2025-09-04 11:14:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:14:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:14:40] [Rank 0] PRINT: step:5000/10000 train_loss:0.6598 val_loss:0.6437 train_time:241094ms step_avg:48.22ms
+[2025-09-04 11:14:40] [Rank 0] PRINT: step:5000/10000 train_loss:0.6598 val_loss:0.6437 train_time:241094ms step_avg:48.22ms
+[2025-09-04 11:14:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:14:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:14:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:14:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:16:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:16:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:16:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:16:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:16:19] [Rank 0]   Total Loss: 4.9122
+[2025-09-04 11:16:19] [Rank 0]   Total Loss: 4.9122
+[2025-09-04 11:16:19] [Rank 0]   Total FTA (Unweighted): 0.9206
+[2025-09-04 11:16:19] [Rank 0]   Total FTA (Unweighted): 0.9206
+[2025-09-04 11:16:19] [Rank 0]   Total FTA (Weighted):   0.9206
+[2025-09-04 11:16:19] [Rank 0]   Total FTA (Weighted):   0.9206
+[2025-09-04 11:16:19] [Rank 0]   Group 0 Loss: 4.7374
+[2025-09-04 11:16:19] [Rank 0]   Group 0 Loss: 4.7374
+[2025-09-04 11:16:19] [Rank 0]   Group 1 Loss: 4.5137
+[2025-09-04 11:16:19] [Rank 0]   Group 1 Loss: 4.5137
+[2025-09-04 11:16:19] [Rank 0]   Group 2 Loss: 4.3642
+[2025-09-04 11:16:19] [Rank 0]   Group 2 Loss: 4.3642
+[2025-09-04 11:16:19] [Rank 0]   Group 3 Loss: 4.8439
+[2025-09-04 11:16:19] [Rank 0]   Group 3 Loss: 4.8439
+[2025-09-04 11:16:19] [Rank 0]   Group 4 Loss: 4.7355
+[2025-09-04 11:16:19] [Rank 0]   Group 4 Loss: 4.7355
+[2025-09-04 11:16:19] [Rank 0]   Group 5 Loss: 4.8157
+[2025-09-04 11:16:19] [Rank 0]   Group 5 Loss: 4.8157
+[2025-09-04 11:16:20] [Rank 0]   Group 6 Loss: 4.7626
+[2025-09-04 11:16:20] [Rank 0]   Group 6 Loss: 4.7626
+[2025-09-04 11:16:20] [Rank 0]   Group 7 Loss: 4.8653
+[2025-09-04 11:16:20] [Rank 0]   Group 7 Loss: 4.8653
+[2025-09-04 11:16:20] [Rank 0]   Group 8 Loss: 5.0044
+[2025-09-04 11:16:20] [Rank 0]   Group 8 Loss: 5.0044
+[2025-09-04 11:16:20] [Rank 0]   Group 9 Loss: 4.9769
+[2025-09-04 11:16:20] [Rank 0]   Group 9 Loss: 4.9769
+[2025-09-04 11:16:20] [Rank 0]   Group 10 Loss: 5.1766
+[2025-09-04 11:16:20] [Rank 0]   Group 10 Loss: 5.1766
+[2025-09-04 11:16:20] [Rank 0]   Group 11 Loss: 5.1457
+[2025-09-04 11:16:20] [Rank 0]   Group 11 Loss: 5.1457
+[2025-09-04 11:16:20] [Rank 0]   Group 12 Loss: 5.1081
+[2025-09-04 11:16:20] [Rank 0]   Group 12 Loss: 5.1081
+[2025-09-04 11:16:20] [Rank 0]   Group 13 Loss: 5.1745
+[2025-09-04 11:16:20] [Rank 0]   Group 13 Loss: 5.1745
+[2025-09-04 11:16:20] [Rank 0]   Group 14 Loss: 5.1639
+[2025-09-04 11:16:20] [Rank 0]   Group 14 Loss: 5.1639
+[2025-09-04 11:16:20] [Rank 0]   Group 15 Loss: 5.2067
+[2025-09-04 11:16:20] [Rank 0]   Group 15 Loss: 5.2067
+[2025-09-04 11:16:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:16:20] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 11:16:20] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 11:16:20] [Rank 0]   Group 13 FTA: 0.9400
+[2025-09-04 11:16:20] [Rank 0]   Group 13 FTA: 0.9400
+[2025-09-04 11:16:20] [Rank 0]   Group 14 FTA: 0.5600
+[2025-09-04 11:16:20] [Rank 0]   Group 14 FTA: 0.5600
+[2025-09-04 11:16:20] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 11:16:20] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 11:16:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:16:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:16:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:16:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:16:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:16:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:16:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:16:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:16:21] [Rank 0] step:5001/10000 train_time:241109ms step_avg:48.21ms
+[2025-09-04 11:16:21] [Rank 0] step:5001/10000 train_time:241109ms step_avg:48.21ms
+[2025-09-04 11:16:22] [Rank 0] step:5021/10000 train_time:241878ms step_avg:48.17ms
+[2025-09-04 11:16:22] [Rank 0] step:5021/10000 train_time:241878ms step_avg:48.17ms
+[2025-09-04 11:16:23] [Rank 0] step:5041/10000 train_time:242633ms step_avg:48.13ms
+[2025-09-04 11:16:23] [Rank 0] step:5041/10000 train_time:242633ms step_avg:48.13ms
+[2025-09-04 11:16:23] [Rank 0] step:5061/10000 train_time:243388ms step_avg:48.09ms
+[2025-09-04 11:16:23] [Rank 0] step:5061/10000 train_time:243388ms step_avg:48.09ms
+[2025-09-04 11:16:24] [Rank 0] step:5081/10000 train_time:244142ms step_avg:48.05ms
+[2025-09-04 11:16:24] [Rank 0] step:5081/10000 train_time:244142ms step_avg:48.05ms
+[2025-09-04 11:16:25] [Rank 0] step:5101/10000 train_time:244897ms step_avg:48.01ms
+[2025-09-04 11:16:25] [Rank 0] step:5101/10000 train_time:244897ms step_avg:48.01ms
+[2025-09-04 11:16:26] [Rank 0] step:5121/10000 train_time:245653ms step_avg:47.97ms
+[2025-09-04 11:16:26] [Rank 0] step:5121/10000 train_time:245653ms step_avg:47.97ms
+[2025-09-04 11:16:26] [Rank 0] step:5141/10000 train_time:246413ms step_avg:47.93ms
+[2025-09-04 11:16:26] [Rank 0] step:5141/10000 train_time:246413ms step_avg:47.93ms
+[2025-09-04 11:16:27] [Rank 0] step:5161/10000 train_time:247168ms step_avg:47.89ms
+[2025-09-04 11:16:27] [Rank 0] step:5161/10000 train_time:247168ms step_avg:47.89ms
+[2025-09-04 11:16:28] [Rank 0] step:5181/10000 train_time:247922ms step_avg:47.85ms
+[2025-09-04 11:16:28] [Rank 0] step:5181/10000 train_time:247922ms step_avg:47.85ms
+[2025-09-04 11:16:29] [Rank 0] step:5201/10000 train_time:248677ms step_avg:47.81ms
+[2025-09-04 11:16:29] [Rank 0] step:5201/10000 train_time:248677ms step_avg:47.81ms
+[2025-09-04 11:16:29] [Rank 0] step:5221/10000 train_time:249432ms step_avg:47.77ms
+[2025-09-04 11:16:29] [Rank 0] step:5221/10000 train_time:249432ms step_avg:47.77ms
+[2025-09-04 11:16:30] [Rank 0] step:5241/10000 train_time:250187ms step_avg:47.74ms
+[2025-09-04 11:16:30] [Rank 0] step:5241/10000 train_time:250187ms step_avg:47.74ms
+[2025-09-04 11:16:31] [Rank 0] step:5261/10000 train_time:250942ms step_avg:47.70ms
+[2025-09-04 11:16:31] [Rank 0] step:5261/10000 train_time:250942ms step_avg:47.70ms
+[2025-09-04 11:16:32] [Rank 0] step:5281/10000 train_time:251696ms step_avg:47.66ms
+[2025-09-04 11:16:32] [Rank 0] step:5281/10000 train_time:251696ms step_avg:47.66ms
+[2025-09-04 11:16:32] [Rank 0] step:5301/10000 train_time:252451ms step_avg:47.62ms
+[2025-09-04 11:16:32] [Rank 0] step:5301/10000 train_time:252451ms step_avg:47.62ms
+[2025-09-04 11:16:33] [Rank 0] step:5321/10000 train_time:253206ms step_avg:47.59ms
+[2025-09-04 11:16:33] [Rank 0] step:5321/10000 train_time:253206ms step_avg:47.59ms
+[2025-09-04 11:16:34] [Rank 0] step:5341/10000 train_time:253961ms step_avg:47.55ms
+[2025-09-04 11:16:34] [Rank 0] step:5341/10000 train_time:253961ms step_avg:47.55ms
+[2025-09-04 11:16:35] [Rank 0] step:5361/10000 train_time:254716ms step_avg:47.51ms
+[2025-09-04 11:16:35] [Rank 0] step:5361/10000 train_time:254716ms step_avg:47.51ms
+[2025-09-04 11:16:35] [Rank 0] step:5381/10000 train_time:255472ms step_avg:47.48ms
+[2025-09-04 11:16:35] [Rank 0] step:5381/10000 train_time:255472ms step_avg:47.48ms
+[2025-09-04 11:16:36] [Rank 0] step:5401/10000 train_time:256227ms step_avg:47.44ms
+[2025-09-04 11:16:36] [Rank 0] step:5401/10000 train_time:256227ms step_avg:47.44ms
+[2025-09-04 11:16:37] [Rank 0] step:5421/10000 train_time:256982ms step_avg:47.40ms
+[2025-09-04 11:16:37] [Rank 0] step:5421/10000 train_time:256982ms step_avg:47.40ms
+[2025-09-04 11:16:38] [Rank 0] step:5441/10000 train_time:257736ms step_avg:47.37ms
+[2025-09-04 11:16:38] [Rank 0] step:5441/10000 train_time:257736ms step_avg:47.37ms
+[2025-09-04 11:16:39] [Rank 0] step:5461/10000 train_time:258492ms step_avg:47.33ms
+[2025-09-04 11:16:39] [Rank 0] step:5461/10000 train_time:258492ms step_avg:47.33ms
+[2025-09-04 11:16:39] [Rank 0] step:5481/10000 train_time:259246ms step_avg:47.30ms
+[2025-09-04 11:16:39] [Rank 0] step:5481/10000 train_time:259246ms step_avg:47.30ms
+[2025-09-04 11:16:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:16:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:16:40] [Rank 0] PRINT: step:5500/10000 train_loss:0.6519 val_loss:0.6371 train_time:260005ms step_avg:47.27ms
+[2025-09-04 11:16:40] [Rank 0] PRINT: step:5500/10000 train_loss:0.6519 val_loss:0.6371 train_time:260005ms step_avg:47.27ms
+[2025-09-04 11:16:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:16:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:16:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:16:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:18:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:18:19] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:18:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:18:19] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:18:19] [Rank 0]   Total Loss: 4.9807
+[2025-09-04 11:18:19] [Rank 0]   Total Loss: 4.9807
+[2025-09-04 11:18:19] [Rank 0]   Total FTA (Unweighted): 0.9281
+[2025-09-04 11:18:19] [Rank 0]   Total FTA (Unweighted): 0.9281
+[2025-09-04 11:18:19] [Rank 0]   Total FTA (Weighted):   0.9281
+[2025-09-04 11:18:19] [Rank 0]   Total FTA (Weighted):   0.9281
+[2025-09-04 11:18:19] [Rank 0]   Group 0 Loss: 4.8258
+[2025-09-04 11:18:19] [Rank 0]   Group 0 Loss: 4.8258
+[2025-09-04 11:18:19] [Rank 0]   Group 1 Loss: 4.4762
+[2025-09-04 11:18:19] [Rank 0]   Group 1 Loss: 4.4762
+[2025-09-04 11:18:19] [Rank 0]   Group 2 Loss: 4.5017
+[2025-09-04 11:18:19] [Rank 0]   Group 2 Loss: 4.5017
+[2025-09-04 11:18:19] [Rank 0]   Group 3 Loss: 4.9108
+[2025-09-04 11:18:19] [Rank 0]   Group 3 Loss: 4.9108
+[2025-09-04 11:18:19] [Rank 0]   Group 4 Loss: 4.8099
+[2025-09-04 11:18:19] [Rank 0]   Group 4 Loss: 4.8099
+[2025-09-04 11:18:19] [Rank 0]   Group 5 Loss: 4.9133
+[2025-09-04 11:18:19] [Rank 0]   Group 5 Loss: 4.9133
+[2025-09-04 11:18:19] [Rank 0]   Group 6 Loss: 4.8589
+[2025-09-04 11:18:19] [Rank 0]   Group 6 Loss: 4.8589
+[2025-09-04 11:18:19] [Rank 0]   Group 7 Loss: 4.9393
+[2025-09-04 11:18:19] [Rank 0]   Group 7 Loss: 4.9393
+[2025-09-04 11:18:19] [Rank 0]   Group 8 Loss: 5.0694
+[2025-09-04 11:18:19] [Rank 0]   Group 8 Loss: 5.0694
+[2025-09-04 11:18:19] [Rank 0]   Group 9 Loss: 5.0515
+[2025-09-04 11:18:19] [Rank 0]   Group 9 Loss: 5.0515
+[2025-09-04 11:18:19] [Rank 0]   Group 10 Loss: 5.2800
+[2025-09-04 11:18:19] [Rank 0]   Group 10 Loss: 5.2800
+[2025-09-04 11:18:19] [Rank 0]   Group 11 Loss: 5.2332
+[2025-09-04 11:18:19] [Rank 0]   Group 11 Loss: 5.2332
+[2025-09-04 11:18:19] [Rank 0]   Group 12 Loss: 5.1451
+[2025-09-04 11:18:19] [Rank 0]   Group 12 Loss: 5.1451
+[2025-09-04 11:18:19] [Rank 0]   Group 13 Loss: 5.2205
+[2025-09-04 11:18:19] [Rank 0]   Group 13 Loss: 5.2205
+[2025-09-04 11:18:19] [Rank 0]   Group 14 Loss: 5.2135
+[2025-09-04 11:18:19] [Rank 0]   Group 14 Loss: 5.2135
+[2025-09-04 11:18:19] [Rank 0]   Group 15 Loss: 5.2429
+[2025-09-04 11:18:19] [Rank 0]   Group 15 Loss: 5.2429
+[2025-09-04 11:18:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:18:19] [Rank 0]   Group 13 FTA: 0.9400
+[2025-09-04 11:18:19] [Rank 0]   Group 13 FTA: 0.9400
+[2025-09-04 11:18:19] [Rank 0]   Group 14 FTA: 0.6400
+[2025-09-04 11:18:19] [Rank 0]   Group 14 FTA: 0.6400
+[2025-09-04 11:18:19] [Rank 0]   Group 15 FTA: 0.2700
+[2025-09-04 11:18:19] [Rank 0]   Group 15 FTA: 0.2700
+[2025-09-04 11:18:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:18:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:18:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:18:20] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:18:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:18:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:18:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:18:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:18:21] [Rank 0] step:5501/10000 train_time:260021ms step_avg:47.27ms
+[2025-09-04 11:18:21] [Rank 0] step:5501/10000 train_time:260021ms step_avg:47.27ms
+[2025-09-04 11:18:22] [Rank 0] step:5521/10000 train_time:260793ms step_avg:47.24ms
+[2025-09-04 11:18:22] [Rank 0] step:5521/10000 train_time:260793ms step_avg:47.24ms
+[2025-09-04 11:18:22] [Rank 0] step:5541/10000 train_time:261548ms step_avg:47.20ms
+[2025-09-04 11:18:22] [Rank 0] step:5541/10000 train_time:261548ms step_avg:47.20ms
+[2025-09-04 11:18:23] [Rank 0] step:5561/10000 train_time:262304ms step_avg:47.17ms
+[2025-09-04 11:18:23] [Rank 0] step:5561/10000 train_time:262304ms step_avg:47.17ms
+[2025-09-04 11:18:24] [Rank 0] step:5581/10000 train_time:263061ms step_avg:47.14ms
+[2025-09-04 11:18:24] [Rank 0] step:5581/10000 train_time:263061ms step_avg:47.14ms
+[2025-09-04 11:18:25] [Rank 0] step:5601/10000 train_time:263817ms step_avg:47.10ms
+[2025-09-04 11:18:25] [Rank 0] step:5601/10000 train_time:263817ms step_avg:47.10ms
+[2025-09-04 11:18:25] [Rank 0] step:5621/10000 train_time:264573ms step_avg:47.07ms
+[2025-09-04 11:18:25] [Rank 0] step:5621/10000 train_time:264573ms step_avg:47.07ms
+[2025-09-04 11:18:26] [Rank 0] step:5641/10000 train_time:265600ms step_avg:47.08ms
+[2025-09-04 11:18:26] [Rank 0] step:5641/10000 train_time:265600ms step_avg:47.08ms
+[2025-09-04 11:18:27] [Rank 0] step:5661/10000 train_time:266356ms step_avg:47.05ms
+[2025-09-04 11:18:27] [Rank 0] step:5661/10000 train_time:266356ms step_avg:47.05ms
+[2025-09-04 11:18:28] [Rank 0] step:5681/10000 train_time:267111ms step_avg:47.02ms
+[2025-09-04 11:18:28] [Rank 0] step:5681/10000 train_time:267111ms step_avg:47.02ms
+[2025-09-04 11:18:29] [Rank 0] step:5701/10000 train_time:267866ms step_avg:46.99ms
+[2025-09-04 11:18:29] [Rank 0] step:5701/10000 train_time:267866ms step_avg:46.99ms
+[2025-09-04 11:18:29] [Rank 0] step:5721/10000 train_time:268621ms step_avg:46.95ms
+[2025-09-04 11:18:29] [Rank 0] step:5721/10000 train_time:268621ms step_avg:46.95ms
+[2025-09-04 11:18:30] [Rank 0] step:5741/10000 train_time:269377ms step_avg:46.92ms
+[2025-09-04 11:18:30] [Rank 0] step:5741/10000 train_time:269377ms step_avg:46.92ms
+[2025-09-04 11:18:31] [Rank 0] step:5761/10000 train_time:270134ms step_avg:46.89ms
+[2025-09-04 11:18:31] [Rank 0] step:5761/10000 train_time:270134ms step_avg:46.89ms
+[2025-09-04 11:18:32] [Rank 0] step:5781/10000 train_time:270890ms step_avg:46.86ms
+[2025-09-04 11:18:32] [Rank 0] step:5781/10000 train_time:270890ms step_avg:46.86ms
+[2025-09-04 11:18:32] [Rank 0] step:5801/10000 train_time:271646ms step_avg:46.83ms
+[2025-09-04 11:18:32] [Rank 0] step:5801/10000 train_time:271646ms step_avg:46.83ms
+[2025-09-04 11:18:33] [Rank 0] step:5821/10000 train_time:272402ms step_avg:46.80ms
+[2025-09-04 11:18:33] [Rank 0] step:5821/10000 train_time:272402ms step_avg:46.80ms
+[2025-09-04 11:18:34] [Rank 0] step:5841/10000 train_time:273159ms step_avg:46.77ms
+[2025-09-04 11:18:34] [Rank 0] step:5841/10000 train_time:273159ms step_avg:46.77ms
+[2025-09-04 11:18:35] [Rank 0] step:5861/10000 train_time:273914ms step_avg:46.74ms
+[2025-09-04 11:18:35] [Rank 0] step:5861/10000 train_time:273914ms step_avg:46.74ms
+[2025-09-04 11:18:35] [Rank 0] step:5881/10000 train_time:274670ms step_avg:46.70ms
+[2025-09-04 11:18:35] [Rank 0] step:5881/10000 train_time:274670ms step_avg:46.70ms
+[2025-09-04 11:18:36] [Rank 0] step:5901/10000 train_time:275425ms step_avg:46.67ms
+[2025-09-04 11:18:36] [Rank 0] step:5901/10000 train_time:275425ms step_avg:46.67ms
+[2025-09-04 11:18:37] [Rank 0] step:5921/10000 train_time:276181ms step_avg:46.64ms
+[2025-09-04 11:18:37] [Rank 0] step:5921/10000 train_time:276181ms step_avg:46.64ms
+[2025-09-04 11:18:38] [Rank 0] step:5941/10000 train_time:276936ms step_avg:46.61ms
+[2025-09-04 11:18:38] [Rank 0] step:5941/10000 train_time:276936ms step_avg:46.61ms
+[2025-09-04 11:18:39] [Rank 0] step:5961/10000 train_time:277691ms step_avg:46.58ms
+[2025-09-04 11:18:39] [Rank 0] step:5961/10000 train_time:277691ms step_avg:46.58ms
+[2025-09-04 11:18:39] [Rank 0] step:5981/10000 train_time:278446ms step_avg:46.56ms
+[2025-09-04 11:18:39] [Rank 0] step:5981/10000 train_time:278446ms step_avg:46.56ms
+[2025-09-04 11:18:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:18:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:18:40] [Rank 0] PRINT: step:6000/10000 train_loss:0.6450 val_loss:0.6307 train_time:279207ms step_avg:46.53ms
+[2025-09-04 11:18:40] [Rank 0] PRINT: step:6000/10000 train_loss:0.6450 val_loss:0.6307 train_time:279207ms step_avg:46.53ms
+[2025-09-04 11:18:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:18:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:18:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:18:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:20:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:20:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:20:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:20:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:20:20] [Rank 0]   Total Loss: 4.8748
+[2025-09-04 11:20:20] [Rank 0]   Total Loss: 4.8748
+[2025-09-04 11:20:20] [Rank 0]   Total FTA (Unweighted): 0.9394
+[2025-09-04 11:20:20] [Rank 0]   Total FTA (Unweighted): 0.9394
+[2025-09-04 11:20:20] [Rank 0]   Total FTA (Weighted):   0.9394
+[2025-09-04 11:20:20] [Rank 0]   Total FTA (Weighted):   0.9394
+[2025-09-04 11:20:20] [Rank 0]   Group 0 Loss: 4.7330
+[2025-09-04 11:20:20] [Rank 0]   Group 0 Loss: 4.7330
+[2025-09-04 11:20:20] [Rank 0]   Group 1 Loss: 4.4384
+[2025-09-04 11:20:20] [Rank 0]   Group 1 Loss: 4.4384
+[2025-09-04 11:20:20] [Rank 0]   Group 2 Loss: 4.4270
+[2025-09-04 11:20:20] [Rank 0]   Group 2 Loss: 4.4270
+[2025-09-04 11:20:20] [Rank 0]   Group 3 Loss: 4.7792
+[2025-09-04 11:20:20] [Rank 0]   Group 3 Loss: 4.7792
+[2025-09-04 11:20:20] [Rank 0]   Group 4 Loss: 4.7132
+[2025-09-04 11:20:20] [Rank 0]   Group 4 Loss: 4.7132
+[2025-09-04 11:20:20] [Rank 0]   Group 5 Loss: 4.7844
+[2025-09-04 11:20:20] [Rank 0]   Group 5 Loss: 4.7844
+[2025-09-04 11:20:20] [Rank 0]   Group 6 Loss: 4.7043
+[2025-09-04 11:20:20] [Rank 0]   Group 6 Loss: 4.7043
+[2025-09-04 11:20:20] [Rank 0]   Group 7 Loss: 4.8281
+[2025-09-04 11:20:20] [Rank 0]   Group 7 Loss: 4.8281
+[2025-09-04 11:20:20] [Rank 0]   Group 8 Loss: 4.9786
+[2025-09-04 11:20:20] [Rank 0]   Group 8 Loss: 4.9786
+[2025-09-04 11:20:20] [Rank 0]   Group 9 Loss: 4.9387
+[2025-09-04 11:20:20] [Rank 0]   Group 9 Loss: 4.9387
+[2025-09-04 11:20:20] [Rank 0]   Group 10 Loss: 5.1402
+[2025-09-04 11:20:20] [Rank 0]   Group 10 Loss: 5.1402
+[2025-09-04 11:20:20] [Rank 0]   Group 11 Loss: 5.0970
+[2025-09-04 11:20:20] [Rank 0]   Group 11 Loss: 5.0970
+[2025-09-04 11:20:20] [Rank 0]   Group 12 Loss: 5.0447
+[2025-09-04 11:20:20] [Rank 0]   Group 12 Loss: 5.0447
+[2025-09-04 11:20:20] [Rank 0]   Group 13 Loss: 5.1101
+[2025-09-04 11:20:20] [Rank 0]   Group 13 Loss: 5.1101
+[2025-09-04 11:20:20] [Rank 0]   Group 14 Loss: 5.1396
+[2025-09-04 11:20:20] [Rank 0]   Group 14 Loss: 5.1396
+[2025-09-04 11:20:20] [Rank 0]   Group 15 Loss: 5.1401
+[2025-09-04 11:20:20] [Rank 0]   Group 15 Loss: 5.1401
+[2025-09-04 11:20:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:20:20] [Rank 0]   Group 13 FTA: 0.9500
+[2025-09-04 11:20:20] [Rank 0]   Group 13 FTA: 0.9500
+[2025-09-04 11:20:20] [Rank 0]   Group 14 FTA: 0.7300
+[2025-09-04 11:20:20] [Rank 0]   Group 14 FTA: 0.7300
+[2025-09-04 11:20:20] [Rank 0]   Group 15 FTA: 0.3500
+[2025-09-04 11:20:20] [Rank 0]   Group 15 FTA: 0.3500
+[2025-09-04 11:20:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:20:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:20:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:20:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:20:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:20:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:20:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:20:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:20:21] [Rank 0] step:6001/10000 train_time:279223ms step_avg:46.53ms
+[2025-09-04 11:20:21] [Rank 0] step:6001/10000 train_time:279223ms step_avg:46.53ms
+[2025-09-04 11:20:22] [Rank 0] step:6021/10000 train_time:280266ms step_avg:46.55ms
+[2025-09-04 11:20:22] [Rank 0] step:6021/10000 train_time:280266ms step_avg:46.55ms
+[2025-09-04 11:20:23] [Rank 0] step:6041/10000 train_time:281021ms step_avg:46.52ms
+[2025-09-04 11:20:23] [Rank 0] step:6041/10000 train_time:281021ms step_avg:46.52ms
+[2025-09-04 11:20:24] [Rank 0] step:6061/10000 train_time:281776ms step_avg:46.49ms
+[2025-09-04 11:20:24] [Rank 0] step:6061/10000 train_time:281776ms step_avg:46.49ms
+[2025-09-04 11:20:25] [Rank 0] step:6081/10000 train_time:282531ms step_avg:46.46ms
+[2025-09-04 11:20:25] [Rank 0] step:6081/10000 train_time:282531ms step_avg:46.46ms
+[2025-09-04 11:20:25] [Rank 0] step:6101/10000 train_time:283285ms step_avg:46.43ms
+[2025-09-04 11:20:25] [Rank 0] step:6101/10000 train_time:283285ms step_avg:46.43ms
+[2025-09-04 11:20:26] [Rank 0] step:6121/10000 train_time:284040ms step_avg:46.40ms
+[2025-09-04 11:20:26] [Rank 0] step:6121/10000 train_time:284040ms step_avg:46.40ms
+[2025-09-04 11:20:27] [Rank 0] step:6141/10000 train_time:284795ms step_avg:46.38ms
+[2025-09-04 11:20:27] [Rank 0] step:6141/10000 train_time:284795ms step_avg:46.38ms
+[2025-09-04 11:20:28] [Rank 0] step:6161/10000 train_time:285550ms step_avg:46.35ms
+[2025-09-04 11:20:28] [Rank 0] step:6161/10000 train_time:285550ms step_avg:46.35ms
+[2025-09-04 11:20:28] [Rank 0] step:6181/10000 train_time:286305ms step_avg:46.32ms
+[2025-09-04 11:20:28] [Rank 0] step:6181/10000 train_time:286305ms step_avg:46.32ms
+[2025-09-04 11:20:29] [Rank 0] step:6201/10000 train_time:287060ms step_avg:46.29ms
+[2025-09-04 11:20:29] [Rank 0] step:6201/10000 train_time:287060ms step_avg:46.29ms
+[2025-09-04 11:20:30] [Rank 0] step:6221/10000 train_time:287815ms step_avg:46.27ms
+[2025-09-04 11:20:30] [Rank 0] step:6221/10000 train_time:287815ms step_avg:46.27ms
+[2025-09-04 11:20:31] [Rank 0] step:6241/10000 train_time:288569ms step_avg:46.24ms
+[2025-09-04 11:20:31] [Rank 0] step:6241/10000 train_time:288569ms step_avg:46.24ms
+[2025-09-04 11:20:32] [Rank 0] step:6261/10000 train_time:289324ms step_avg:46.21ms
+[2025-09-04 11:20:32] [Rank 0] step:6261/10000 train_time:289324ms step_avg:46.21ms
+[2025-09-04 11:20:32] [Rank 0] step:6281/10000 train_time:290078ms step_avg:46.18ms
+[2025-09-04 11:20:32] [Rank 0] step:6281/10000 train_time:290078ms step_avg:46.18ms
+[2025-09-04 11:20:33] [Rank 0] step:6301/10000 train_time:290835ms step_avg:46.16ms
+[2025-09-04 11:20:33] [Rank 0] step:6301/10000 train_time:290835ms step_avg:46.16ms
+[2025-09-04 11:20:34] [Rank 0] step:6321/10000 train_time:291589ms step_avg:46.13ms
+[2025-09-04 11:20:34] [Rank 0] step:6321/10000 train_time:291589ms step_avg:46.13ms
+[2025-09-04 11:20:35] [Rank 0] step:6341/10000 train_time:292343ms step_avg:46.10ms
+[2025-09-04 11:20:35] [Rank 0] step:6341/10000 train_time:292343ms step_avg:46.10ms
+[2025-09-04 11:20:35] [Rank 0] step:6361/10000 train_time:293098ms step_avg:46.08ms
+[2025-09-04 11:20:35] [Rank 0] step:6361/10000 train_time:293098ms step_avg:46.08ms
+[2025-09-04 11:20:36] [Rank 0] step:6381/10000 train_time:293852ms step_avg:46.05ms
+[2025-09-04 11:20:36] [Rank 0] step:6381/10000 train_time:293852ms step_avg:46.05ms
+[2025-09-04 11:20:37] [Rank 0] step:6401/10000 train_time:294607ms step_avg:46.03ms
+[2025-09-04 11:20:37] [Rank 0] step:6401/10000 train_time:294607ms step_avg:46.03ms
+[2025-09-04 11:20:38] [Rank 0] step:6421/10000 train_time:295361ms step_avg:46.00ms
+[2025-09-04 11:20:38] [Rank 0] step:6421/10000 train_time:295361ms step_avg:46.00ms
+[2025-09-04 11:20:38] [Rank 0] step:6441/10000 train_time:296116ms step_avg:45.97ms
+[2025-09-04 11:20:38] [Rank 0] step:6441/10000 train_time:296116ms step_avg:45.97ms
+[2025-09-04 11:20:39] [Rank 0] step:6461/10000 train_time:296871ms step_avg:45.95ms
+[2025-09-04 11:20:39] [Rank 0] step:6461/10000 train_time:296871ms step_avg:45.95ms
+[2025-09-04 11:20:40] [Rank 0] step:6481/10000 train_time:297626ms step_avg:45.92ms
+[2025-09-04 11:20:40] [Rank 0] step:6481/10000 train_time:297626ms step_avg:45.92ms
+[2025-09-04 11:20:41] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:20:41] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:20:41] [Rank 0] PRINT: step:6500/10000 train_loss:0.6386 val_loss:0.6252 train_time:298386ms step_avg:45.91ms
+[2025-09-04 11:20:41] [Rank 0] PRINT: step:6500/10000 train_loss:0.6386 val_loss:0.6252 train_time:298386ms step_avg:45.91ms
+[2025-09-04 11:20:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:20:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:20:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:20:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:22:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:22:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:22:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:22:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:22:21] [Rank 0]   Total Loss: 4.8703
+[2025-09-04 11:22:21] [Rank 0]   Total Loss: 4.8703
+[2025-09-04 11:22:21] [Rank 0]   Total FTA (Unweighted): 0.9494
+[2025-09-04 11:22:21] [Rank 0]   Total FTA (Unweighted): 0.9494
+[2025-09-04 11:22:21] [Rank 0]   Total FTA (Weighted):   0.9494
+[2025-09-04 11:22:21] [Rank 0]   Total FTA (Weighted):   0.9494
+[2025-09-04 11:22:21] [Rank 0]   Group 0 Loss: 4.7526
+[2025-09-04 11:22:21] [Rank 0]   Group 0 Loss: 4.7526
+[2025-09-04 11:22:21] [Rank 0]   Group 1 Loss: 4.4257
+[2025-09-04 11:22:21] [Rank 0]   Group 1 Loss: 4.4257
+[2025-09-04 11:22:21] [Rank 0]   Group 2 Loss: 4.4780
+[2025-09-04 11:22:21] [Rank 0]   Group 2 Loss: 4.4780
+[2025-09-04 11:22:21] [Rank 0]   Group 3 Loss: 4.7851
+[2025-09-04 11:22:21] [Rank 0]   Group 3 Loss: 4.7851
+[2025-09-04 11:22:21] [Rank 0]   Group 4 Loss: 4.7234
+[2025-09-04 11:22:21] [Rank 0]   Group 4 Loss: 4.7234
+[2025-09-04 11:22:21] [Rank 0]   Group 5 Loss: 4.7747
+[2025-09-04 11:22:21] [Rank 0]   Group 5 Loss: 4.7747
+[2025-09-04 11:22:21] [Rank 0]   Group 6 Loss: 4.7161
+[2025-09-04 11:22:21] [Rank 0]   Group 6 Loss: 4.7161
+[2025-09-04 11:22:21] [Rank 0]   Group 7 Loss: 4.8011
+[2025-09-04 11:22:21] [Rank 0]   Group 7 Loss: 4.8011
+[2025-09-04 11:22:21] [Rank 0]   Group 8 Loss: 4.9841
+[2025-09-04 11:22:21] [Rank 0]   Group 8 Loss: 4.9841
+[2025-09-04 11:22:21] [Rank 0]   Group 9 Loss: 4.9334
+[2025-09-04 11:22:21] [Rank 0]   Group 9 Loss: 4.9334
+[2025-09-04 11:22:21] [Rank 0]   Group 10 Loss: 5.1290
+[2025-09-04 11:22:21] [Rank 0]   Group 10 Loss: 5.1290
+[2025-09-04 11:22:21] [Rank 0]   Group 11 Loss: 5.0865
+[2025-09-04 11:22:21] [Rank 0]   Group 11 Loss: 5.0865
+[2025-09-04 11:22:21] [Rank 0]   Group 12 Loss: 5.0302
+[2025-09-04 11:22:21] [Rank 0]   Group 12 Loss: 5.0302
+[2025-09-04 11:22:21] [Rank 0]   Group 13 Loss: 5.0871
+[2025-09-04 11:22:21] [Rank 0]   Group 13 Loss: 5.0871
+[2025-09-04 11:22:21] [Rank 0]   Group 14 Loss: 5.1017
+[2025-09-04 11:22:21] [Rank 0]   Group 14 Loss: 5.1017
+[2025-09-04 11:22:21] [Rank 0]   Group 15 Loss: 5.1160
+[2025-09-04 11:22:21] [Rank 0]   Group 15 Loss: 5.1160
+[2025-09-04 11:22:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:22:21] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 11:22:21] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 11:22:21] [Rank 0]   Group 14 FTA: 0.7300
+[2025-09-04 11:22:21] [Rank 0]   Group 14 FTA: 0.7300
+[2025-09-04 11:22:21] [Rank 0]   Group 15 FTA: 0.4800
+[2025-09-04 11:22:21] [Rank 0]   Group 15 FTA: 0.4800
+[2025-09-04 11:22:21] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:22:21] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:22:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:22:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:22:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:22:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:22:22] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:22:22] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:22:22] [Rank 0] step:6501/10000 train_time:298401ms step_avg:45.90ms
+[2025-09-04 11:22:22] [Rank 0] step:6501/10000 train_time:298401ms step_avg:45.90ms
+[2025-09-04 11:22:23] [Rank 0] step:6521/10000 train_time:299178ms step_avg:45.88ms
+[2025-09-04 11:22:23] [Rank 0] step:6521/10000 train_time:299178ms step_avg:45.88ms
+[2025-09-04 11:22:24] [Rank 0] step:6541/10000 train_time:299934ms step_avg:45.85ms
+[2025-09-04 11:22:24] [Rank 0] step:6541/10000 train_time:299934ms step_avg:45.85ms
+[2025-09-04 11:22:25] [Rank 0] step:6561/10000 train_time:300868ms step_avg:45.86ms
+[2025-09-04 11:22:25] [Rank 0] step:6561/10000 train_time:300868ms step_avg:45.86ms
+[2025-09-04 11:22:26] [Rank 0] step:6581/10000 train_time:301676ms step_avg:45.84ms
+[2025-09-04 11:22:26] [Rank 0] step:6581/10000 train_time:301676ms step_avg:45.84ms
+[2025-09-04 11:22:26] [Rank 0] step:6601/10000 train_time:302431ms step_avg:45.82ms
+[2025-09-04 11:22:26] [Rank 0] step:6601/10000 train_time:302431ms step_avg:45.82ms
+[2025-09-04 11:22:27] [Rank 0] step:6621/10000 train_time:303185ms step_avg:45.79ms
+[2025-09-04 11:22:27] [Rank 0] step:6621/10000 train_time:303185ms step_avg:45.79ms
+[2025-09-04 11:22:28] [Rank 0] step:6641/10000 train_time:303940ms step_avg:45.77ms
+[2025-09-04 11:22:28] [Rank 0] step:6641/10000 train_time:303940ms step_avg:45.77ms
+[2025-09-04 11:22:29] [Rank 0] step:6661/10000 train_time:304695ms step_avg:45.74ms
+[2025-09-04 11:22:29] [Rank 0] step:6661/10000 train_time:304695ms step_avg:45.74ms
+[2025-09-04 11:22:29] [Rank 0] step:6681/10000 train_time:305449ms step_avg:45.72ms
+[2025-09-04 11:22:29] [Rank 0] step:6681/10000 train_time:305449ms step_avg:45.72ms
+[2025-09-04 11:22:30] [Rank 0] step:6701/10000 train_time:306205ms step_avg:45.70ms
+[2025-09-04 11:22:30] [Rank 0] step:6701/10000 train_time:306205ms step_avg:45.70ms
+[2025-09-04 11:22:31] [Rank 0] step:6721/10000 train_time:306961ms step_avg:45.67ms
+[2025-09-04 11:22:31] [Rank 0] step:6721/10000 train_time:306961ms step_avg:45.67ms
+[2025-09-04 11:22:32] [Rank 0] step:6741/10000 train_time:307716ms step_avg:45.65ms
+[2025-09-04 11:22:32] [Rank 0] step:6741/10000 train_time:307716ms step_avg:45.65ms
+[2025-09-04 11:22:33] [Rank 0] step:6761/10000 train_time:308471ms step_avg:45.63ms
+[2025-09-04 11:22:33] [Rank 0] step:6761/10000 train_time:308471ms step_avg:45.63ms
+[2025-09-04 11:22:33] [Rank 0] step:6781/10000 train_time:309226ms step_avg:45.60ms
+[2025-09-04 11:22:33] [Rank 0] step:6781/10000 train_time:309226ms step_avg:45.60ms
+[2025-09-04 11:22:34] [Rank 0] step:6801/10000 train_time:309982ms step_avg:45.58ms
+[2025-09-04 11:22:34] [Rank 0] step:6801/10000 train_time:309982ms step_avg:45.58ms
+[2025-09-04 11:22:35] [Rank 0] step:6821/10000 train_time:310737ms step_avg:45.56ms
+[2025-09-04 11:22:35] [Rank 0] step:6821/10000 train_time:310737ms step_avg:45.56ms
+[2025-09-04 11:22:36] [Rank 0] step:6841/10000 train_time:311763ms step_avg:45.57ms
+[2025-09-04 11:22:36] [Rank 0] step:6841/10000 train_time:311763ms step_avg:45.57ms
+[2025-09-04 11:22:37] [Rank 0] step:6861/10000 train_time:312518ms step_avg:45.55ms
+[2025-09-04 11:22:37] [Rank 0] step:6861/10000 train_time:312518ms step_avg:45.55ms
+[2025-09-04 11:22:37] [Rank 0] step:6881/10000 train_time:313273ms step_avg:45.53ms
+[2025-09-04 11:22:37] [Rank 0] step:6881/10000 train_time:313273ms step_avg:45.53ms
+[2025-09-04 11:22:38] [Rank 0] step:6901/10000 train_time:314028ms step_avg:45.50ms
+[2025-09-04 11:22:38] [Rank 0] step:6901/10000 train_time:314028ms step_avg:45.50ms
+[2025-09-04 11:22:39] [Rank 0] step:6921/10000 train_time:314783ms step_avg:45.48ms
+[2025-09-04 11:22:39] [Rank 0] step:6921/10000 train_time:314783ms step_avg:45.48ms
+[2025-09-04 11:22:40] [Rank 0] step:6941/10000 train_time:315538ms step_avg:45.46ms
+[2025-09-04 11:22:40] [Rank 0] step:6941/10000 train_time:315538ms step_avg:45.46ms
+[2025-09-04 11:22:40] [Rank 0] step:6961/10000 train_time:316293ms step_avg:45.44ms
+[2025-09-04 11:22:40] [Rank 0] step:6961/10000 train_time:316293ms step_avg:45.44ms
+[2025-09-04 11:22:41] [Rank 0] step:6981/10000 train_time:317048ms step_avg:45.42ms
+[2025-09-04 11:22:41] [Rank 0] step:6981/10000 train_time:317048ms step_avg:45.42ms
+[2025-09-04 11:22:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:22:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:22:42] [Rank 0] PRINT: step:7000/10000 train_loss:0.6323 val_loss:0.6201 train_time:317808ms step_avg:45.40ms
+[2025-09-04 11:22:42] [Rank 0] PRINT: step:7000/10000 train_loss:0.6323 val_loss:0.6201 train_time:317808ms step_avg:45.40ms
+[2025-09-04 11:22:42] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:22:42] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:22:42] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:22:42] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:24:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:24:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:24:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:24:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:24:21] [Rank 0]   Total Loss: 4.9150
+[2025-09-04 11:24:21] [Rank 0]   Total Loss: 4.9150
+[2025-09-04 11:24:21] [Rank 0]   Total FTA (Unweighted): 0.9631
+[2025-09-04 11:24:21] [Rank 0]   Total FTA (Unweighted): 0.9631
+[2025-09-04 11:24:21] [Rank 0]   Total FTA (Weighted):   0.9631
+[2025-09-04 11:24:21] [Rank 0]   Total FTA (Weighted):   0.9631
+[2025-09-04 11:24:21] [Rank 0]   Group 0 Loss: 4.8125
+[2025-09-04 11:24:21] [Rank 0]   Group 0 Loss: 4.8125
+[2025-09-04 11:24:21] [Rank 0]   Group 1 Loss: 4.5152
+[2025-09-04 11:24:21] [Rank 0]   Group 1 Loss: 4.5152
+[2025-09-04 11:24:21] [Rank 0]   Group 2 Loss: 4.4567
+[2025-09-04 11:24:21] [Rank 0]   Group 2 Loss: 4.4567
+[2025-09-04 11:24:21] [Rank 0]   Group 3 Loss: 4.8299
+[2025-09-04 11:24:21] [Rank 0]   Group 3 Loss: 4.8299
+[2025-09-04 11:24:21] [Rank 0]   Group 4 Loss: 4.7526
+[2025-09-04 11:24:21] [Rank 0]   Group 4 Loss: 4.7526
+[2025-09-04 11:24:21] [Rank 0]   Group 5 Loss: 4.8148
+[2025-09-04 11:24:21] [Rank 0]   Group 5 Loss: 4.8148
+[2025-09-04 11:24:21] [Rank 0]   Group 6 Loss: 4.7629
+[2025-09-04 11:24:21] [Rank 0]   Group 6 Loss: 4.7629
+[2025-09-04 11:24:21] [Rank 0]   Group 7 Loss: 4.8339
+[2025-09-04 11:24:21] [Rank 0]   Group 7 Loss: 4.8339
+[2025-09-04 11:24:21] [Rank 0]   Group 8 Loss: 5.0340
+[2025-09-04 11:24:21] [Rank 0]   Group 8 Loss: 5.0340
+[2025-09-04 11:24:21] [Rank 0]   Group 9 Loss: 4.9829
+[2025-09-04 11:24:21] [Rank 0]   Group 9 Loss: 4.9829
+[2025-09-04 11:24:21] [Rank 0]   Group 10 Loss: 5.1517
+[2025-09-04 11:24:21] [Rank 0]   Group 10 Loss: 5.1517
+[2025-09-04 11:24:21] [Rank 0]   Group 11 Loss: 5.1524
+[2025-09-04 11:24:21] [Rank 0]   Group 11 Loss: 5.1524
+[2025-09-04 11:24:21] [Rank 0]   Group 12 Loss: 5.0846
+[2025-09-04 11:24:21] [Rank 0]   Group 12 Loss: 5.0846
+[2025-09-04 11:24:21] [Rank 0]   Group 13 Loss: 5.1514
+[2025-09-04 11:24:21] [Rank 0]   Group 13 Loss: 5.1514
+[2025-09-04 11:24:21] [Rank 0]   Group 14 Loss: 5.1475
+[2025-09-04 11:24:21] [Rank 0]   Group 14 Loss: 5.1475
+[2025-09-04 11:24:21] [Rank 0]   Group 15 Loss: 5.1574
+[2025-09-04 11:24:21] [Rank 0]   Group 15 Loss: 5.1574
+[2025-09-04 11:24:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:24:21] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 11:24:21] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 11:24:21] [Rank 0]   Group 14 FTA: 0.8700
+[2025-09-04 11:24:21] [Rank 0]   Group 14 FTA: 0.8700
+[2025-09-04 11:24:21] [Rank 0]   Group 15 FTA: 0.5500
+[2025-09-04 11:24:21] [Rank 0]   Group 15 FTA: 0.5500
+[2025-09-04 11:24:21] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:24:21] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:24:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:24:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:24:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:24:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:24:22] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:24:22] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:24:22] [Rank 0] step:7001/10000 train_time:317824ms step_avg:45.40ms
+[2025-09-04 11:24:22] [Rank 0] step:7001/10000 train_time:317824ms step_avg:45.40ms
+[2025-09-04 11:24:23] [Rank 0] step:7021/10000 train_time:318583ms step_avg:45.38ms
+[2025-09-04 11:24:23] [Rank 0] step:7021/10000 train_time:318583ms step_avg:45.38ms
+[2025-09-04 11:24:24] [Rank 0] step:7041/10000 train_time:319340ms step_avg:45.35ms
+[2025-09-04 11:24:24] [Rank 0] step:7041/10000 train_time:319340ms step_avg:45.35ms
+[2025-09-04 11:24:25] [Rank 0] step:7061/10000 train_time:320094ms step_avg:45.33ms
+[2025-09-04 11:24:25] [Rank 0] step:7061/10000 train_time:320094ms step_avg:45.33ms
+[2025-09-04 11:24:25] [Rank 0] step:7081/10000 train_time:320849ms step_avg:45.31ms
+[2025-09-04 11:24:25] [Rank 0] step:7081/10000 train_time:320849ms step_avg:45.31ms
+[2025-09-04 11:24:26] [Rank 0] step:7101/10000 train_time:321603ms step_avg:45.29ms
+[2025-09-04 11:24:26] [Rank 0] step:7101/10000 train_time:321603ms step_avg:45.29ms
+[2025-09-04 11:24:27] [Rank 0] step:7121/10000 train_time:322357ms step_avg:45.27ms
+[2025-09-04 11:24:27] [Rank 0] step:7121/10000 train_time:322357ms step_avg:45.27ms
+[2025-09-04 11:24:28] [Rank 0] step:7141/10000 train_time:323112ms step_avg:45.25ms
+[2025-09-04 11:24:28] [Rank 0] step:7141/10000 train_time:323112ms step_avg:45.25ms
+[2025-09-04 11:24:28] [Rank 0] step:7161/10000 train_time:323867ms step_avg:45.23ms
+[2025-09-04 11:24:28] [Rank 0] step:7161/10000 train_time:323867ms step_avg:45.23ms
+[2025-09-04 11:24:29] [Rank 0] step:7181/10000 train_time:324913ms step_avg:45.25ms
+[2025-09-04 11:24:29] [Rank 0] step:7181/10000 train_time:324913ms step_avg:45.25ms
+[2025-09-04 11:24:30] [Rank 0] step:7201/10000 train_time:325669ms step_avg:45.23ms
+[2025-09-04 11:24:30] [Rank 0] step:7201/10000 train_time:325669ms step_avg:45.23ms
+[2025-09-04 11:24:31] [Rank 0] step:7221/10000 train_time:326424ms step_avg:45.20ms
+[2025-09-04 11:24:31] [Rank 0] step:7221/10000 train_time:326424ms step_avg:45.20ms
+[2025-09-04 11:24:32] [Rank 0] step:7241/10000 train_time:327447ms step_avg:45.22ms
+[2025-09-04 11:24:32] [Rank 0] step:7241/10000 train_time:327447ms step_avg:45.22ms
+[2025-09-04 11:24:33] [Rank 0] step:7261/10000 train_time:328202ms step_avg:45.20ms
+[2025-09-04 11:24:33] [Rank 0] step:7261/10000 train_time:328202ms step_avg:45.20ms
+[2025-09-04 11:24:34] [Rank 0] step:7281/10000 train_time:328957ms step_avg:45.18ms
+[2025-09-04 11:24:34] [Rank 0] step:7281/10000 train_time:328957ms step_avg:45.18ms
+[2025-09-04 11:24:34] [Rank 0] step:7301/10000 train_time:329713ms step_avg:45.16ms
+[2025-09-04 11:24:34] [Rank 0] step:7301/10000 train_time:329713ms step_avg:45.16ms
+[2025-09-04 11:24:35] [Rank 0] step:7321/10000 train_time:330468ms step_avg:45.14ms
+[2025-09-04 11:24:35] [Rank 0] step:7321/10000 train_time:330468ms step_avg:45.14ms
+[2025-09-04 11:24:36] [Rank 0] step:7341/10000 train_time:331224ms step_avg:45.12ms
+[2025-09-04 11:24:36] [Rank 0] step:7341/10000 train_time:331224ms step_avg:45.12ms
+[2025-09-04 11:24:37] [Rank 0] step:7361/10000 train_time:331980ms step_avg:45.10ms
+[2025-09-04 11:24:37] [Rank 0] step:7361/10000 train_time:331980ms step_avg:45.10ms
+[2025-09-04 11:24:37] [Rank 0] step:7381/10000 train_time:332736ms step_avg:45.08ms
+[2025-09-04 11:24:37] [Rank 0] step:7381/10000 train_time:332736ms step_avg:45.08ms
+[2025-09-04 11:24:38] [Rank 0] step:7401/10000 train_time:333492ms step_avg:45.06ms
+[2025-09-04 11:24:38] [Rank 0] step:7401/10000 train_time:333492ms step_avg:45.06ms
+[2025-09-04 11:24:39] [Rank 0] step:7421/10000 train_time:334246ms step_avg:45.04ms
+[2025-09-04 11:24:39] [Rank 0] step:7421/10000 train_time:334246ms step_avg:45.04ms
+[2025-09-04 11:24:40] [Rank 0] step:7441/10000 train_time:335004ms step_avg:45.02ms
+[2025-09-04 11:24:40] [Rank 0] step:7441/10000 train_time:335004ms step_avg:45.02ms
+[2025-09-04 11:24:40] [Rank 0] step:7461/10000 train_time:335760ms step_avg:45.00ms
+[2025-09-04 11:24:40] [Rank 0] step:7461/10000 train_time:335760ms step_avg:45.00ms
+[2025-09-04 11:24:41] [Rank 0] step:7481/10000 train_time:336515ms step_avg:44.98ms
+[2025-09-04 11:24:41] [Rank 0] step:7481/10000 train_time:336515ms step_avg:44.98ms
+[2025-09-04 11:24:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:24:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:24:42] [Rank 0] PRINT: step:7500/10000 train_loss:0.6269 val_loss:0.6164 train_time:337277ms step_avg:44.97ms
+[2025-09-04 11:24:42] [Rank 0] PRINT: step:7500/10000 train_loss:0.6269 val_loss:0.6164 train_time:337277ms step_avg:44.97ms
+[2025-09-04 11:24:42] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:24:42] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:24:42] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:24:42] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:26:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:26:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:26:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:26:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:26:21] [Rank 0]   Total Loss: 4.8919
+[2025-09-04 11:26:21] [Rank 0]   Total Loss: 4.8919
+[2025-09-04 11:26:21] [Rank 0]   Total FTA (Unweighted): 0.9700
+[2025-09-04 11:26:21] [Rank 0]   Total FTA (Unweighted): 0.9700
+[2025-09-04 11:26:21] [Rank 0]   Total FTA (Weighted):   0.9700
+[2025-09-04 11:26:21] [Rank 0]   Total FTA (Weighted):   0.9700
+[2025-09-04 11:26:21] [Rank 0]   Group 0 Loss: 4.8333
+[2025-09-04 11:26:21] [Rank 0]   Group 0 Loss: 4.8333
+[2025-09-04 11:26:21] [Rank 0]   Group 1 Loss: 4.4891
+[2025-09-04 11:26:21] [Rank 0]   Group 1 Loss: 4.4891
+[2025-09-04 11:26:21] [Rank 0]   Group 2 Loss: 4.3710
+[2025-09-04 11:26:21] [Rank 0]   Group 2 Loss: 4.3710
+[2025-09-04 11:26:21] [Rank 0]   Group 3 Loss: 4.8071
+[2025-09-04 11:26:21] [Rank 0]   Group 3 Loss: 4.8071
+[2025-09-04 11:26:21] [Rank 0]   Group 4 Loss: 4.7413
+[2025-09-04 11:26:21] [Rank 0]   Group 4 Loss: 4.7413
+[2025-09-04 11:26:21] [Rank 0]   Group 5 Loss: 4.8089
+[2025-09-04 11:26:21] [Rank 0]   Group 5 Loss: 4.8089
+[2025-09-04 11:26:21] [Rank 0]   Group 6 Loss: 4.7236
+[2025-09-04 11:26:21] [Rank 0]   Group 6 Loss: 4.7236
+[2025-09-04 11:26:21] [Rank 0]   Group 7 Loss: 4.8227
+[2025-09-04 11:26:21] [Rank 0]   Group 7 Loss: 4.8227
+[2025-09-04 11:26:21] [Rank 0]   Group 8 Loss: 5.0139
+[2025-09-04 11:26:21] [Rank 0]   Group 8 Loss: 5.0139
+[2025-09-04 11:26:21] [Rank 0]   Group 9 Loss: 4.9476
+[2025-09-04 11:26:21] [Rank 0]   Group 9 Loss: 4.9476
+[2025-09-04 11:26:21] [Rank 0]   Group 10 Loss: 5.1504
+[2025-09-04 11:26:21] [Rank 0]   Group 10 Loss: 5.1504
+[2025-09-04 11:26:21] [Rank 0]   Group 11 Loss: 5.1092
+[2025-09-04 11:26:21] [Rank 0]   Group 11 Loss: 5.1092
+[2025-09-04 11:26:21] [Rank 0]   Group 12 Loss: 5.0663
+[2025-09-04 11:26:21] [Rank 0]   Group 12 Loss: 5.0663
+[2025-09-04 11:26:21] [Rank 0]   Group 13 Loss: 5.1205
+[2025-09-04 11:26:21] [Rank 0]   Group 13 Loss: 5.1205
+[2025-09-04 11:26:21] [Rank 0]   Group 14 Loss: 5.1168
+[2025-09-04 11:26:21] [Rank 0]   Group 14 Loss: 5.1168
+[2025-09-04 11:26:21] [Rank 0]   Group 15 Loss: 5.1479
+[2025-09-04 11:26:21] [Rank 0]   Group 15 Loss: 5.1479
+[2025-09-04 11:26:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:26:21] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 11:26:21] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 11:26:21] [Rank 0]   Group 14 FTA: 0.9300
+[2025-09-04 11:26:21] [Rank 0]   Group 14 FTA: 0.9300
+[2025-09-04 11:26:21] [Rank 0]   Group 15 FTA: 0.6000
+[2025-09-04 11:26:21] [Rank 0]   Group 15 FTA: 0.6000
+[2025-09-04 11:26:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:26:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:26:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:26:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:26:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:26:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:26:23] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:26:23] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:26:23] [Rank 0] step:7501/10000 train_time:337291ms step_avg:44.97ms
+[2025-09-04 11:26:23] [Rank 0] step:7501/10000 train_time:337291ms step_avg:44.97ms
+[2025-09-04 11:26:24] [Rank 0] step:7521/10000 train_time:338043ms step_avg:44.95ms
+[2025-09-04 11:26:24] [Rank 0] step:7521/10000 train_time:338043ms step_avg:44.95ms
+[2025-09-04 11:26:24] [Rank 0] step:7541/10000 train_time:338797ms step_avg:44.93ms
+[2025-09-04 11:26:24] [Rank 0] step:7541/10000 train_time:338797ms step_avg:44.93ms
+[2025-09-04 11:26:25] [Rank 0] step:7561/10000 train_time:339552ms step_avg:44.91ms
+[2025-09-04 11:26:25] [Rank 0] step:7561/10000 train_time:339552ms step_avg:44.91ms
+[2025-09-04 11:26:26] [Rank 0] step:7581/10000 train_time:340307ms step_avg:44.89ms
+[2025-09-04 11:26:26] [Rank 0] step:7581/10000 train_time:340307ms step_avg:44.89ms
+[2025-09-04 11:26:27] [Rank 0] step:7601/10000 train_time:341063ms step_avg:44.87ms
+[2025-09-04 11:26:27] [Rank 0] step:7601/10000 train_time:341063ms step_avg:44.87ms
+[2025-09-04 11:26:27] [Rank 0] step:7621/10000 train_time:341817ms step_avg:44.85ms
+[2025-09-04 11:26:27] [Rank 0] step:7621/10000 train_time:341817ms step_avg:44.85ms
+[2025-09-04 11:26:28] [Rank 0] step:7641/10000 train_time:342841ms step_avg:44.87ms
+[2025-09-04 11:26:28] [Rank 0] step:7641/10000 train_time:342841ms step_avg:44.87ms
+[2025-09-04 11:26:29] [Rank 0] step:7661/10000 train_time:343597ms step_avg:44.85ms
+[2025-09-04 11:26:29] [Rank 0] step:7661/10000 train_time:343597ms step_avg:44.85ms
+[2025-09-04 11:26:30] [Rank 0] step:7681/10000 train_time:344353ms step_avg:44.83ms
+[2025-09-04 11:26:30] [Rank 0] step:7681/10000 train_time:344353ms step_avg:44.83ms
+[2025-09-04 11:26:31] [Rank 0] step:7701/10000 train_time:345107ms step_avg:44.81ms
+[2025-09-04 11:26:31] [Rank 0] step:7701/10000 train_time:345107ms step_avg:44.81ms
+[2025-09-04 11:26:31] [Rank 0] step:7721/10000 train_time:345862ms step_avg:44.80ms
+[2025-09-04 11:26:31] [Rank 0] step:7721/10000 train_time:345862ms step_avg:44.80ms
+[2025-09-04 11:26:32] [Rank 0] step:7741/10000 train_time:346617ms step_avg:44.78ms
+[2025-09-04 11:26:32] [Rank 0] step:7741/10000 train_time:346617ms step_avg:44.78ms
+[2025-09-04 11:26:33] [Rank 0] step:7761/10000 train_time:347374ms step_avg:44.76ms
+[2025-09-04 11:26:33] [Rank 0] step:7761/10000 train_time:347374ms step_avg:44.76ms
+[2025-09-04 11:26:34] [Rank 0] step:7781/10000 train_time:348129ms step_avg:44.74ms
+[2025-09-04 11:26:34] [Rank 0] step:7781/10000 train_time:348129ms step_avg:44.74ms
+[2025-09-04 11:26:34] [Rank 0] step:7801/10000 train_time:348884ms step_avg:44.72ms
+[2025-09-04 11:26:34] [Rank 0] step:7801/10000 train_time:348884ms step_avg:44.72ms
+[2025-09-04 11:26:35] [Rank 0] step:7821/10000 train_time:349902ms step_avg:44.74ms
+[2025-09-04 11:26:35] [Rank 0] step:7821/10000 train_time:349902ms step_avg:44.74ms
+[2025-09-04 11:26:36] [Rank 0] step:7841/10000 train_time:350657ms step_avg:44.72ms
+[2025-09-04 11:26:36] [Rank 0] step:7841/10000 train_time:350657ms step_avg:44.72ms
+[2025-09-04 11:26:37] [Rank 0] step:7861/10000 train_time:351412ms step_avg:44.70ms
+[2025-09-04 11:26:37] [Rank 0] step:7861/10000 train_time:351412ms step_avg:44.70ms
+[2025-09-04 11:26:38] [Rank 0] step:7881/10000 train_time:352432ms step_avg:44.72ms
+[2025-09-04 11:26:38] [Rank 0] step:7881/10000 train_time:352432ms step_avg:44.72ms
+[2025-09-04 11:26:39] [Rank 0] step:7901/10000 train_time:353187ms step_avg:44.70ms
+[2025-09-04 11:26:39] [Rank 0] step:7901/10000 train_time:353187ms step_avg:44.70ms
+[2025-09-04 11:26:39] [Rank 0] step:7921/10000 train_time:353941ms step_avg:44.68ms
+[2025-09-04 11:26:39] [Rank 0] step:7921/10000 train_time:353941ms step_avg:44.68ms
+[2025-09-04 11:26:40] [Rank 0] step:7941/10000 train_time:354695ms step_avg:44.67ms
+[2025-09-04 11:26:40] [Rank 0] step:7941/10000 train_time:354695ms step_avg:44.67ms
+[2025-09-04 11:26:41] [Rank 0] step:7961/10000 train_time:355450ms step_avg:44.65ms
+[2025-09-04 11:26:41] [Rank 0] step:7961/10000 train_time:355450ms step_avg:44.65ms
+[2025-09-04 11:26:42] [Rank 0] step:7981/10000 train_time:356205ms step_avg:44.63ms
+[2025-09-04 11:26:42] [Rank 0] step:7981/10000 train_time:356205ms step_avg:44.63ms
+[2025-09-04 11:26:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:26:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:26:43] [Rank 0] PRINT: step:8000/10000 train_loss:0.6222 val_loss:0.6130 train_time:356965ms step_avg:44.62ms
+[2025-09-04 11:26:43] [Rank 0] PRINT: step:8000/10000 train_loss:0.6222 val_loss:0.6130 train_time:356965ms step_avg:44.62ms
+[2025-09-04 11:26:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:26:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:26:43] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:26:43] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:28:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:28:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:28:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:28:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:28:21] [Rank 0]   Total Loss: 4.9385
+[2025-09-04 11:28:21] [Rank 0]   Total Loss: 4.9385
+[2025-09-04 11:28:21] [Rank 0]   Total FTA (Unweighted): 0.9731
+[2025-09-04 11:28:21] [Rank 0]   Total FTA (Unweighted): 0.9731
+[2025-09-04 11:28:21] [Rank 0]   Total FTA (Weighted):   0.9731
+[2025-09-04 11:28:21] [Rank 0]   Total FTA (Weighted):   0.9731
+[2025-09-04 11:28:21] [Rank 0]   Group 0 Loss: 4.8406
+[2025-09-04 11:28:21] [Rank 0]   Group 0 Loss: 4.8406
+[2025-09-04 11:28:21] [Rank 0]   Group 1 Loss: 4.5139
+[2025-09-04 11:28:21] [Rank 0]   Group 1 Loss: 4.5139
+[2025-09-04 11:28:21] [Rank 0]   Group 2 Loss: 4.4373
+[2025-09-04 11:28:21] [Rank 0]   Group 2 Loss: 4.4373
+[2025-09-04 11:28:21] [Rank 0]   Group 3 Loss: 4.8378
+[2025-09-04 11:28:21] [Rank 0]   Group 3 Loss: 4.8378
+[2025-09-04 11:28:21] [Rank 0]   Group 4 Loss: 4.7774
+[2025-09-04 11:28:21] [Rank 0]   Group 4 Loss: 4.7774
+[2025-09-04 11:28:22] [Rank 0]   Group 5 Loss: 4.8518
+[2025-09-04 11:28:22] [Rank 0]   Group 5 Loss: 4.8518
+[2025-09-04 11:28:22] [Rank 0]   Group 6 Loss: 4.7711
+[2025-09-04 11:28:22] [Rank 0]   Group 6 Loss: 4.7711
+[2025-09-04 11:28:22] [Rank 0]   Group 7 Loss: 4.8857
+[2025-09-04 11:28:22] [Rank 0]   Group 7 Loss: 4.8857
+[2025-09-04 11:28:22] [Rank 0]   Group 8 Loss: 5.0655
+[2025-09-04 11:28:22] [Rank 0]   Group 8 Loss: 5.0655
+[2025-09-04 11:28:22] [Rank 0]   Group 9 Loss: 5.0074
+[2025-09-04 11:28:22] [Rank 0]   Group 9 Loss: 5.0074
+[2025-09-04 11:28:22] [Rank 0]   Group 10 Loss: 5.1994
+[2025-09-04 11:28:22] [Rank 0]   Group 10 Loss: 5.1994
+[2025-09-04 11:28:22] [Rank 0]   Group 11 Loss: 5.1499
+[2025-09-04 11:28:22] [Rank 0]   Group 11 Loss: 5.1499
+[2025-09-04 11:28:22] [Rank 0]   Group 12 Loss: 5.1120
+[2025-09-04 11:28:22] [Rank 0]   Group 12 Loss: 5.1120
+[2025-09-04 11:28:22] [Rank 0]   Group 13 Loss: 5.1886
+[2025-09-04 11:28:22] [Rank 0]   Group 13 Loss: 5.1886
+[2025-09-04 11:28:22] [Rank 0]   Group 14 Loss: 5.1915
+[2025-09-04 11:28:22] [Rank 0]   Group 14 Loss: 5.1915
+[2025-09-04 11:28:22] [Rank 0]   Group 15 Loss: 5.1856
+[2025-09-04 11:28:22] [Rank 0]   Group 15 Loss: 5.1856
+[2025-09-04 11:28:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:28:22] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 11:28:22] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 11:28:22] [Rank 0]   Group 14 FTA: 0.9100
+[2025-09-04 11:28:22] [Rank 0]   Group 14 FTA: 0.9100
+[2025-09-04 11:28:22] [Rank 0]   Group 15 FTA: 0.6700
+[2025-09-04 11:28:22] [Rank 0]   Group 15 FTA: 0.6700
+[2025-09-04 11:28:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:28:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:28:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:28:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:28:23] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:28:23] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:28:23] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:28:23] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:28:23] [Rank 0] step:8001/10000 train_time:356981ms step_avg:44.62ms
+[2025-09-04 11:28:23] [Rank 0] step:8001/10000 train_time:356981ms step_avg:44.62ms
+[2025-09-04 11:28:24] [Rank 0] step:8021/10000 train_time:358023ms step_avg:44.64ms
+[2025-09-04 11:28:24] [Rank 0] step:8021/10000 train_time:358023ms step_avg:44.64ms
+[2025-09-04 11:28:25] [Rank 0] step:8041/10000 train_time:358779ms step_avg:44.62ms
+[2025-09-04 11:28:25] [Rank 0] step:8041/10000 train_time:358779ms step_avg:44.62ms
+[2025-09-04 11:28:26] [Rank 0] step:8061/10000 train_time:359532ms step_avg:44.60ms
+[2025-09-04 11:28:26] [Rank 0] step:8061/10000 train_time:359532ms step_avg:44.60ms
+[2025-09-04 11:28:26] [Rank 0] step:8081/10000 train_time:360288ms step_avg:44.58ms
+[2025-09-04 11:28:26] [Rank 0] step:8081/10000 train_time:360288ms step_avg:44.58ms
+[2025-09-04 11:28:27] [Rank 0] step:8101/10000 train_time:361042ms step_avg:44.57ms
+[2025-09-04 11:28:27] [Rank 0] step:8101/10000 train_time:361042ms step_avg:44.57ms
+[2025-09-04 11:28:28] [Rank 0] step:8121/10000 train_time:361798ms step_avg:44.55ms
+[2025-09-04 11:28:28] [Rank 0] step:8121/10000 train_time:361798ms step_avg:44.55ms
+[2025-09-04 11:28:29] [Rank 0] step:8141/10000 train_time:362552ms step_avg:44.53ms
+[2025-09-04 11:28:29] [Rank 0] step:8141/10000 train_time:362552ms step_avg:44.53ms
+[2025-09-04 11:28:29] [Rank 0] step:8161/10000 train_time:363308ms step_avg:44.52ms
+[2025-09-04 11:28:29] [Rank 0] step:8161/10000 train_time:363308ms step_avg:44.52ms
+[2025-09-04 11:28:30] [Rank 0] step:8181/10000 train_time:364063ms step_avg:44.50ms
+[2025-09-04 11:28:30] [Rank 0] step:8181/10000 train_time:364063ms step_avg:44.50ms
+[2025-09-04 11:28:31] [Rank 0] step:8201/10000 train_time:364817ms step_avg:44.48ms
+[2025-09-04 11:28:31] [Rank 0] step:8201/10000 train_time:364817ms step_avg:44.48ms
+[2025-09-04 11:28:32] [Rank 0] step:8221/10000 train_time:365572ms step_avg:44.47ms
+[2025-09-04 11:28:32] [Rank 0] step:8221/10000 train_time:365572ms step_avg:44.47ms
+[2025-09-04 11:28:32] [Rank 0] step:8241/10000 train_time:366327ms step_avg:44.45ms
+[2025-09-04 11:28:32] [Rank 0] step:8241/10000 train_time:366327ms step_avg:44.45ms
+[2025-09-04 11:28:33] [Rank 0] step:8261/10000 train_time:367082ms step_avg:44.44ms
+[2025-09-04 11:28:33] [Rank 0] step:8261/10000 train_time:367082ms step_avg:44.44ms
+[2025-09-04 11:28:34] [Rank 0] step:8281/10000 train_time:367837ms step_avg:44.42ms
+[2025-09-04 11:28:34] [Rank 0] step:8281/10000 train_time:367837ms step_avg:44.42ms
+[2025-09-04 11:28:35] [Rank 0] step:8301/10000 train_time:368592ms step_avg:44.40ms
+[2025-09-04 11:28:35] [Rank 0] step:8301/10000 train_time:368592ms step_avg:44.40ms
+[2025-09-04 11:28:35] [Rank 0] step:8321/10000 train_time:369347ms step_avg:44.39ms
+[2025-09-04 11:28:35] [Rank 0] step:8321/10000 train_time:369347ms step_avg:44.39ms
+[2025-09-04 11:28:36] [Rank 0] step:8341/10000 train_time:370102ms step_avg:44.37ms
+[2025-09-04 11:28:36] [Rank 0] step:8341/10000 train_time:370102ms step_avg:44.37ms
+[2025-09-04 11:28:37] [Rank 0] step:8361/10000 train_time:370858ms step_avg:44.36ms
+[2025-09-04 11:28:37] [Rank 0] step:8361/10000 train_time:370858ms step_avg:44.36ms
+[2025-09-04 11:28:38] [Rank 0] step:8381/10000 train_time:371613ms step_avg:44.34ms
+[2025-09-04 11:28:38] [Rank 0] step:8381/10000 train_time:371613ms step_avg:44.34ms
+[2025-09-04 11:28:38] [Rank 0] step:8401/10000 train_time:372368ms step_avg:44.32ms
+[2025-09-04 11:28:38] [Rank 0] step:8401/10000 train_time:372368ms step_avg:44.32ms
+[2025-09-04 11:28:39] [Rank 0] step:8421/10000 train_time:373123ms step_avg:44.31ms
+[2025-09-04 11:28:39] [Rank 0] step:8421/10000 train_time:373123ms step_avg:44.31ms
+[2025-09-04 11:28:40] [Rank 0] step:8441/10000 train_time:373878ms step_avg:44.29ms
+[2025-09-04 11:28:40] [Rank 0] step:8441/10000 train_time:373878ms step_avg:44.29ms
+[2025-09-04 11:28:41] [Rank 0] step:8461/10000 train_time:374633ms step_avg:44.28ms
+[2025-09-04 11:28:41] [Rank 0] step:8461/10000 train_time:374633ms step_avg:44.28ms
+[2025-09-04 11:28:42] [Rank 0] step:8481/10000 train_time:375562ms step_avg:44.28ms
+[2025-09-04 11:28:42] [Rank 0] step:8481/10000 train_time:375562ms step_avg:44.28ms
+[2025-09-04 11:28:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:28:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:28:43] [Rank 0] PRINT: step:8500/10000 train_loss:0.6180 val_loss:0.6101 train_time:376409ms step_avg:44.28ms
+[2025-09-04 11:28:43] [Rank 0] PRINT: step:8500/10000 train_loss:0.6180 val_loss:0.6101 train_time:376409ms step_avg:44.28ms
+[2025-09-04 11:28:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:28:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:28:43] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:28:43] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:30:22] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:30:22] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:30:22] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:30:22] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:30:22] [Rank 0]   Total Loss: 4.9111
+[2025-09-04 11:30:22] [Rank 0]   Total Loss: 4.9111
+[2025-09-04 11:30:22] [Rank 0]   Total FTA (Unweighted): 0.9813
+[2025-09-04 11:30:22] [Rank 0]   Total FTA (Unweighted): 0.9813
+[2025-09-04 11:30:22] [Rank 0]   Total FTA (Weighted):   0.9812
+[2025-09-04 11:30:22] [Rank 0]   Total FTA (Weighted):   0.9812
+[2025-09-04 11:30:22] [Rank 0]   Group 0 Loss: 4.8728
+[2025-09-04 11:30:22] [Rank 0]   Group 0 Loss: 4.8728
+[2025-09-04 11:30:22] [Rank 0]   Group 1 Loss: 4.4751
+[2025-09-04 11:30:22] [Rank 0]   Group 1 Loss: 4.4751
+[2025-09-04 11:30:22] [Rank 0]   Group 2 Loss: 4.4406
+[2025-09-04 11:30:22] [Rank 0]   Group 2 Loss: 4.4406
+[2025-09-04 11:30:22] [Rank 0]   Group 3 Loss: 4.8072
+[2025-09-04 11:30:22] [Rank 0]   Group 3 Loss: 4.8072
+[2025-09-04 11:30:22] [Rank 0]   Group 4 Loss: 4.7578
+[2025-09-04 11:30:22] [Rank 0]   Group 4 Loss: 4.7578
+[2025-09-04 11:30:22] [Rank 0]   Group 5 Loss: 4.8266
+[2025-09-04 11:30:22] [Rank 0]   Group 5 Loss: 4.8266
+[2025-09-04 11:30:22] [Rank 0]   Group 6 Loss: 4.7392
+[2025-09-04 11:30:22] [Rank 0]   Group 6 Loss: 4.7392
+[2025-09-04 11:30:22] [Rank 0]   Group 7 Loss: 4.8380
+[2025-09-04 11:30:22] [Rank 0]   Group 7 Loss: 4.8380
+[2025-09-04 11:30:22] [Rank 0]   Group 8 Loss: 5.0425
+[2025-09-04 11:30:22] [Rank 0]   Group 8 Loss: 5.0425
+[2025-09-04 11:30:22] [Rank 0]   Group 9 Loss: 4.9669
+[2025-09-04 11:30:22] [Rank 0]   Group 9 Loss: 4.9669
+[2025-09-04 11:30:22] [Rank 0]   Group 10 Loss: 5.1444
+[2025-09-04 11:30:22] [Rank 0]   Group 10 Loss: 5.1444
+[2025-09-04 11:30:22] [Rank 0]   Group 11 Loss: 5.1254
+[2025-09-04 11:30:22] [Rank 0]   Group 11 Loss: 5.1254
+[2025-09-04 11:30:22] [Rank 0]   Group 12 Loss: 5.0878
+[2025-09-04 11:30:22] [Rank 0]   Group 12 Loss: 5.0878
+[2025-09-04 11:30:22] [Rank 0]   Group 13 Loss: 5.1508
+[2025-09-04 11:30:22] [Rank 0]   Group 13 Loss: 5.1508
+[2025-09-04 11:30:22] [Rank 0]   Group 14 Loss: 5.1337
+[2025-09-04 11:30:22] [Rank 0]   Group 14 Loss: 5.1337
+[2025-09-04 11:30:22] [Rank 0]   Group 15 Loss: 5.1691
+[2025-09-04 11:30:22] [Rank 0]   Group 15 Loss: 5.1691
+[2025-09-04 11:30:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:30:22] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:30:23] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:30:23] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:30:23] [Rank 0]   Group 14 FTA: 0.9600
+[2025-09-04 11:30:23] [Rank 0]   Group 14 FTA: 0.9600
+[2025-09-04 11:30:23] [Rank 0]   Group 15 FTA: 0.7400
+[2025-09-04 11:30:23] [Rank 0]   Group 15 FTA: 0.7400
+[2025-09-04 11:30:23] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:30:23] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:30:23] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:30:23] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:30:24] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:30:24] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:30:24] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:30:24] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:30:24] [Rank 0] step:8501/10000 train_time:376426ms step_avg:44.28ms
+[2025-09-04 11:30:24] [Rank 0] step:8501/10000 train_time:376426ms step_avg:44.28ms
+[2025-09-04 11:30:25] [Rank 0] step:8521/10000 train_time:377185ms step_avg:44.27ms
+[2025-09-04 11:30:25] [Rank 0] step:8521/10000 train_time:377185ms step_avg:44.27ms
+[2025-09-04 11:30:25] [Rank 0] step:8541/10000 train_time:377939ms step_avg:44.25ms
+[2025-09-04 11:30:25] [Rank 0] step:8541/10000 train_time:377939ms step_avg:44.25ms
+[2025-09-04 11:30:26] [Rank 0] step:8561/10000 train_time:378694ms step_avg:44.23ms
+[2025-09-04 11:30:26] [Rank 0] step:8561/10000 train_time:378694ms step_avg:44.23ms
+[2025-09-04 11:30:27] [Rank 0] step:8581/10000 train_time:379449ms step_avg:44.22ms
+[2025-09-04 11:30:27] [Rank 0] step:8581/10000 train_time:379449ms step_avg:44.22ms
+[2025-09-04 11:30:28] [Rank 0] step:8601/10000 train_time:380203ms step_avg:44.20ms
+[2025-09-04 11:30:28] [Rank 0] step:8601/10000 train_time:380203ms step_avg:44.20ms
+[2025-09-04 11:30:29] [Rank 0] step:8621/10000 train_time:380958ms step_avg:44.19ms
+[2025-09-04 11:30:29] [Rank 0] step:8621/10000 train_time:380958ms step_avg:44.19ms
+[2025-09-04 11:30:29] [Rank 0] step:8641/10000 train_time:381712ms step_avg:44.17ms
+[2025-09-04 11:30:29] [Rank 0] step:8641/10000 train_time:381712ms step_avg:44.17ms
+[2025-09-04 11:30:30] [Rank 0] step:8661/10000 train_time:382467ms step_avg:44.16ms
+[2025-09-04 11:30:30] [Rank 0] step:8661/10000 train_time:382467ms step_avg:44.16ms
+[2025-09-04 11:30:31] [Rank 0] step:8681/10000 train_time:383222ms step_avg:44.14ms
+[2025-09-04 11:30:31] [Rank 0] step:8681/10000 train_time:383222ms step_avg:44.14ms
+[2025-09-04 11:30:32] [Rank 0] step:8701/10000 train_time:383978ms step_avg:44.13ms
+[2025-09-04 11:30:32] [Rank 0] step:8701/10000 train_time:383978ms step_avg:44.13ms
+[2025-09-04 11:30:32] [Rank 0] step:8721/10000 train_time:384733ms step_avg:44.12ms
+[2025-09-04 11:30:32] [Rank 0] step:8721/10000 train_time:384733ms step_avg:44.12ms
+[2025-09-04 11:30:33] [Rank 0] step:8741/10000 train_time:385488ms step_avg:44.10ms
+[2025-09-04 11:30:33] [Rank 0] step:8741/10000 train_time:385488ms step_avg:44.10ms
+[2025-09-04 11:30:34] [Rank 0] step:8761/10000 train_time:386243ms step_avg:44.09ms
+[2025-09-04 11:30:34] [Rank 0] step:8761/10000 train_time:386243ms step_avg:44.09ms
+[2025-09-04 11:30:35] [Rank 0] step:8781/10000 train_time:386998ms step_avg:44.07ms
+[2025-09-04 11:30:35] [Rank 0] step:8781/10000 train_time:386998ms step_avg:44.07ms
+[2025-09-04 11:30:35] [Rank 0] step:8801/10000 train_time:387753ms step_avg:44.06ms
+[2025-09-04 11:30:35] [Rank 0] step:8801/10000 train_time:387753ms step_avg:44.06ms
+[2025-09-04 11:30:36] [Rank 0] step:8821/10000 train_time:388508ms step_avg:44.04ms
+[2025-09-04 11:30:36] [Rank 0] step:8821/10000 train_time:388508ms step_avg:44.04ms
+[2025-09-04 11:30:37] [Rank 0] step:8841/10000 train_time:389534ms step_avg:44.06ms
+[2025-09-04 11:30:37] [Rank 0] step:8841/10000 train_time:389534ms step_avg:44.06ms
+[2025-09-04 11:30:38] [Rank 0] step:8861/10000 train_time:390289ms step_avg:44.05ms
+[2025-09-04 11:30:38] [Rank 0] step:8861/10000 train_time:390289ms step_avg:44.05ms
+[2025-09-04 11:30:39] [Rank 0] step:8881/10000 train_time:391044ms step_avg:44.03ms
+[2025-09-04 11:30:39] [Rank 0] step:8881/10000 train_time:391044ms step_avg:44.03ms
+[2025-09-04 11:30:39] [Rank 0] step:8901/10000 train_time:391799ms step_avg:44.02ms
+[2025-09-04 11:30:39] [Rank 0] step:8901/10000 train_time:391799ms step_avg:44.02ms
+[2025-09-04 11:30:40] [Rank 0] step:8921/10000 train_time:392554ms step_avg:44.00ms
+[2025-09-04 11:30:40] [Rank 0] step:8921/10000 train_time:392554ms step_avg:44.00ms
+[2025-09-04 11:30:41] [Rank 0] step:8941/10000 train_time:393310ms step_avg:43.99ms
+[2025-09-04 11:30:41] [Rank 0] step:8941/10000 train_time:393310ms step_avg:43.99ms
+[2025-09-04 11:30:42] [Rank 0] step:8961/10000 train_time:394065ms step_avg:43.98ms
+[2025-09-04 11:30:42] [Rank 0] step:8961/10000 train_time:394065ms step_avg:43.98ms
+[2025-09-04 11:30:42] [Rank 0] step:8981/10000 train_time:394820ms step_avg:43.96ms
+[2025-09-04 11:30:42] [Rank 0] step:8981/10000 train_time:394820ms step_avg:43.96ms
+[2025-09-04 11:30:43] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:30:43] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:30:44] [Rank 0] PRINT: step:9000/10000 train_loss:0.6143 val_loss:0.6081 train_time:395580ms step_avg:43.95ms
+[2025-09-04 11:30:44] [Rank 0] PRINT: step:9000/10000 train_loss:0.6143 val_loss:0.6081 train_time:395580ms step_avg:43.95ms
+[2025-09-04 11:30:44] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:30:44] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:30:44] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:30:44] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:32:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:32:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:32:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:32:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:32:24] [Rank 0]   Total Loss: 4.9750
+[2025-09-04 11:32:24] [Rank 0]   Total Loss: 4.9750
+[2025-09-04 11:32:24] [Rank 0]   Total FTA (Unweighted): 0.9875
+[2025-09-04 11:32:24] [Rank 0]   Total FTA (Unweighted): 0.9875
+[2025-09-04 11:32:24] [Rank 0]   Total FTA (Weighted):   0.9875
+[2025-09-04 11:32:24] [Rank 0]   Total FTA (Weighted):   0.9875
+[2025-09-04 11:32:24] [Rank 0]   Group 0 Loss: 4.9236
+[2025-09-04 11:32:24] [Rank 0]   Group 0 Loss: 4.9236
+[2025-09-04 11:32:24] [Rank 0]   Group 1 Loss: 4.4806
+[2025-09-04 11:32:24] [Rank 0]   Group 1 Loss: 4.4806
+[2025-09-04 11:32:24] [Rank 0]   Group 2 Loss: 4.4929
+[2025-09-04 11:32:24] [Rank 0]   Group 2 Loss: 4.4929
+[2025-09-04 11:32:24] [Rank 0]   Group 3 Loss: 4.8899
+[2025-09-04 11:32:24] [Rank 0]   Group 3 Loss: 4.8899
+[2025-09-04 11:32:24] [Rank 0]   Group 4 Loss: 4.8241
+[2025-09-04 11:32:24] [Rank 0]   Group 4 Loss: 4.8241
+[2025-09-04 11:32:24] [Rank 0]   Group 5 Loss: 4.8805
+[2025-09-04 11:32:24] [Rank 0]   Group 5 Loss: 4.8805
+[2025-09-04 11:32:24] [Rank 0]   Group 6 Loss: 4.8161
+[2025-09-04 11:32:24] [Rank 0]   Group 6 Loss: 4.8161
+[2025-09-04 11:32:24] [Rank 0]   Group 7 Loss: 4.9072
+[2025-09-04 11:32:24] [Rank 0]   Group 7 Loss: 4.9072
+[2025-09-04 11:32:24] [Rank 0]   Group 8 Loss: 5.0940
+[2025-09-04 11:32:24] [Rank 0]   Group 8 Loss: 5.0940
+[2025-09-04 11:32:24] [Rank 0]   Group 9 Loss: 5.0503
+[2025-09-04 11:32:24] [Rank 0]   Group 9 Loss: 5.0503
+[2025-09-04 11:32:24] [Rank 0]   Group 10 Loss: 5.2198
+[2025-09-04 11:32:24] [Rank 0]   Group 10 Loss: 5.2198
+[2025-09-04 11:32:24] [Rank 0]   Group 11 Loss: 5.1889
+[2025-09-04 11:32:24] [Rank 0]   Group 11 Loss: 5.1889
+[2025-09-04 11:32:24] [Rank 0]   Group 12 Loss: 5.1612
+[2025-09-04 11:32:24] [Rank 0]   Group 12 Loss: 5.1612
+[2025-09-04 11:32:24] [Rank 0]   Group 13 Loss: 5.2309
+[2025-09-04 11:32:24] [Rank 0]   Group 13 Loss: 5.2309
+[2025-09-04 11:32:24] [Rank 0]   Group 14 Loss: 5.2008
+[2025-09-04 11:32:24] [Rank 0]   Group 14 Loss: 5.2008
+[2025-09-04 11:32:24] [Rank 0]   Group 15 Loss: 5.2391
+[2025-09-04 11:32:24] [Rank 0]   Group 15 Loss: 5.2391
+[2025-09-04 11:32:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 11:32:24] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 11:32:24] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:32:24] [Rank 0]   Group 14 FTA: 0.9700
+[2025-09-04 11:32:24] [Rank 0]   Group 14 FTA: 0.9700
+[2025-09-04 11:32:24] [Rank 0]   Group 15 FTA: 0.8400
+[2025-09-04 11:32:24] [Rank 0]   Group 15 FTA: 0.8400
+[2025-09-04 11:32:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:32:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:32:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:32:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:32:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:32:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:32:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:32:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:32:25] [Rank 0] step:9001/10000 train_time:395597ms step_avg:43.95ms
+[2025-09-04 11:32:25] [Rank 0] step:9001/10000 train_time:395597ms step_avg:43.95ms
+[2025-09-04 11:32:26] [Rank 0] step:9021/10000 train_time:396354ms step_avg:43.94ms
+[2025-09-04 11:32:26] [Rank 0] step:9021/10000 train_time:396354ms step_avg:43.94ms
+[2025-09-04 11:32:27] [Rank 0] step:9041/10000 train_time:397108ms step_avg:43.92ms
+[2025-09-04 11:32:27] [Rank 0] step:9041/10000 train_time:397108ms step_avg:43.92ms
+[2025-09-04 11:32:27] [Rank 0] step:9061/10000 train_time:397862ms step_avg:43.91ms
+[2025-09-04 11:32:27] [Rank 0] step:9061/10000 train_time:397862ms step_avg:43.91ms
+[2025-09-04 11:32:28] [Rank 0] step:9081/10000 train_time:398617ms step_avg:43.90ms
+[2025-09-04 11:32:28] [Rank 0] step:9081/10000 train_time:398617ms step_avg:43.90ms
+[2025-09-04 11:32:29] [Rank 0] step:9101/10000 train_time:399371ms step_avg:43.88ms
+[2025-09-04 11:32:29] [Rank 0] step:9101/10000 train_time:399371ms step_avg:43.88ms
+[2025-09-04 11:32:30] [Rank 0] step:9121/10000 train_time:400127ms step_avg:43.87ms
+[2025-09-04 11:32:30] [Rank 0] step:9121/10000 train_time:400127ms step_avg:43.87ms
+[2025-09-04 11:32:30] [Rank 0] step:9141/10000 train_time:400881ms step_avg:43.86ms
+[2025-09-04 11:32:30] [Rank 0] step:9141/10000 train_time:400881ms step_avg:43.86ms
+[2025-09-04 11:32:31] [Rank 0] step:9161/10000 train_time:401635ms step_avg:43.84ms
+[2025-09-04 11:32:31] [Rank 0] step:9161/10000 train_time:401635ms step_avg:43.84ms
+[2025-09-04 11:32:32] [Rank 0] step:9181/10000 train_time:402390ms step_avg:43.83ms
+[2025-09-04 11:32:32] [Rank 0] step:9181/10000 train_time:402390ms step_avg:43.83ms
+[2025-09-04 11:32:33] [Rank 0] step:9201/10000 train_time:403144ms step_avg:43.82ms
+[2025-09-04 11:32:33] [Rank 0] step:9201/10000 train_time:403144ms step_avg:43.82ms
+[2025-09-04 11:32:33] [Rank 0] step:9221/10000 train_time:403898ms step_avg:43.80ms
+[2025-09-04 11:32:33] [Rank 0] step:9221/10000 train_time:403898ms step_avg:43.80ms
+[2025-09-04 11:32:34] [Rank 0] step:9241/10000 train_time:404653ms step_avg:43.79ms
+[2025-09-04 11:32:34] [Rank 0] step:9241/10000 train_time:404653ms step_avg:43.79ms
+[2025-09-04 11:32:35] [Rank 0] step:9261/10000 train_time:405407ms step_avg:43.78ms
+[2025-09-04 11:32:35] [Rank 0] step:9261/10000 train_time:405407ms step_avg:43.78ms
+[2025-09-04 11:32:36] [Rank 0] step:9281/10000 train_time:406161ms step_avg:43.76ms
+[2025-09-04 11:32:36] [Rank 0] step:9281/10000 train_time:406161ms step_avg:43.76ms
+[2025-09-04 11:32:36] [Rank 0] step:9301/10000 train_time:406916ms step_avg:43.75ms
+[2025-09-04 11:32:36] [Rank 0] step:9301/10000 train_time:406916ms step_avg:43.75ms
+[2025-09-04 11:32:37] [Rank 0] step:9321/10000 train_time:407670ms step_avg:43.74ms
+[2025-09-04 11:32:37] [Rank 0] step:9321/10000 train_time:407670ms step_avg:43.74ms
+[2025-09-04 11:32:38] [Rank 0] step:9341/10000 train_time:408424ms step_avg:43.72ms
+[2025-09-04 11:32:38] [Rank 0] step:9341/10000 train_time:408424ms step_avg:43.72ms
+[2025-09-04 11:32:39] [Rank 0] step:9361/10000 train_time:409178ms step_avg:43.71ms
+[2025-09-04 11:32:39] [Rank 0] step:9361/10000 train_time:409178ms step_avg:43.71ms
+[2025-09-04 11:32:39] [Rank 0] step:9381/10000 train_time:409933ms step_avg:43.70ms
+[2025-09-04 11:32:39] [Rank 0] step:9381/10000 train_time:409933ms step_avg:43.70ms
+[2025-09-04 11:32:40] [Rank 0] step:9401/10000 train_time:410687ms step_avg:43.69ms
+[2025-09-04 11:32:40] [Rank 0] step:9401/10000 train_time:410687ms step_avg:43.69ms
+[2025-09-04 11:32:41] [Rank 0] step:9421/10000 train_time:411442ms step_avg:43.67ms
+[2025-09-04 11:32:41] [Rank 0] step:9421/10000 train_time:411442ms step_avg:43.67ms
+[2025-09-04 11:32:42] [Rank 0] step:9441/10000 train_time:412198ms step_avg:43.66ms
+[2025-09-04 11:32:42] [Rank 0] step:9441/10000 train_time:412198ms step_avg:43.66ms
+[2025-09-04 11:32:42] [Rank 0] step:9461/10000 train_time:412952ms step_avg:43.65ms
+[2025-09-04 11:32:42] [Rank 0] step:9461/10000 train_time:412952ms step_avg:43.65ms
+[2025-09-04 11:32:43] [Rank 0] step:9481/10000 train_time:413707ms step_avg:43.64ms
+[2025-09-04 11:32:43] [Rank 0] step:9481/10000 train_time:413707ms step_avg:43.64ms
+[2025-09-04 11:32:44] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:32:44] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:32:44] [Rank 0] PRINT: step:9500/10000 train_loss:0.6115 val_loss:0.6066 train_time:414467ms step_avg:43.63ms
+[2025-09-04 11:32:44] [Rank 0] PRINT: step:9500/10000 train_loss:0.6115 val_loss:0.6066 train_time:414467ms step_avg:43.63ms
+[2025-09-04 11:32:44] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:32:44] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:32:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:32:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:34:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:34:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:34:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:34:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:34:24] [Rank 0]   Total Loss: 4.9564
+[2025-09-04 11:34:24] [Rank 0]   Total Loss: 4.9564
+[2025-09-04 11:34:24] [Rank 0]   Total FTA (Unweighted): 0.9938
+[2025-09-04 11:34:24] [Rank 0]   Total FTA (Unweighted): 0.9938
+[2025-09-04 11:34:24] [Rank 0]   Total FTA (Weighted):   0.9938
+[2025-09-04 11:34:24] [Rank 0]   Total FTA (Weighted):   0.9938
+[2025-09-04 11:34:24] [Rank 0]   Group 0 Loss: 4.8823
+[2025-09-04 11:34:24] [Rank 0]   Group 0 Loss: 4.8823
+[2025-09-04 11:34:24] [Rank 0]   Group 1 Loss: 4.4977
+[2025-09-04 11:34:24] [Rank 0]   Group 1 Loss: 4.4977
+[2025-09-04 11:34:24] [Rank 0]   Group 2 Loss: 4.5015
+[2025-09-04 11:34:24] [Rank 0]   Group 2 Loss: 4.5015
+[2025-09-04 11:34:24] [Rank 0]   Group 3 Loss: 4.8678
+[2025-09-04 11:34:24] [Rank 0]   Group 3 Loss: 4.8678
+[2025-09-04 11:34:24] [Rank 0]   Group 4 Loss: 4.7933
+[2025-09-04 11:34:24] [Rank 0]   Group 4 Loss: 4.7933
+[2025-09-04 11:34:24] [Rank 0]   Group 5 Loss: 4.8643
+[2025-09-04 11:34:24] [Rank 0]   Group 5 Loss: 4.8643
+[2025-09-04 11:34:24] [Rank 0]   Group 6 Loss: 4.7969
+[2025-09-04 11:34:24] [Rank 0]   Group 6 Loss: 4.7969
+[2025-09-04 11:34:24] [Rank 0]   Group 7 Loss: 4.8872
+[2025-09-04 11:34:24] [Rank 0]   Group 7 Loss: 4.8872
+[2025-09-04 11:34:24] [Rank 0]   Group 8 Loss: 5.0735
+[2025-09-04 11:34:24] [Rank 0]   Group 8 Loss: 5.0735
+[2025-09-04 11:34:24] [Rank 0]   Group 9 Loss: 5.0102
+[2025-09-04 11:34:24] [Rank 0]   Group 9 Loss: 5.0102
+[2025-09-04 11:34:24] [Rank 0]   Group 10 Loss: 5.2028
+[2025-09-04 11:34:24] [Rank 0]   Group 10 Loss: 5.2028
+[2025-09-04 11:34:24] [Rank 0]   Group 11 Loss: 5.1735
+[2025-09-04 11:34:24] [Rank 0]   Group 11 Loss: 5.1735
+[2025-09-04 11:34:24] [Rank 0]   Group 12 Loss: 5.1465
+[2025-09-04 11:34:24] [Rank 0]   Group 12 Loss: 5.1465
+[2025-09-04 11:34:24] [Rank 0]   Group 13 Loss: 5.2039
+[2025-09-04 11:34:24] [Rank 0]   Group 13 Loss: 5.2039
+[2025-09-04 11:34:24] [Rank 0]   Group 14 Loss: 5.1778
+[2025-09-04 11:34:24] [Rank 0]   Group 14 Loss: 5.1778
+[2025-09-04 11:34:24] [Rank 0]   Group 15 Loss: 5.2224
+[2025-09-04 11:34:24] [Rank 0]   Group 15 Loss: 5.2224
+[2025-09-04 11:34:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:34:24] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 11:34:24] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 11:34:24] [Rank 0]   Group 15 FTA: 0.9100
+[2025-09-04 11:34:24] [Rank 0]   Group 15 FTA: 0.9100
+[2025-09-04 11:34:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:34:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:34:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:34:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:34:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:34:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:34:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:34:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:34:25] [Rank 0] step:9501/10000 train_time:414483ms step_avg:43.63ms
+[2025-09-04 11:34:25] [Rank 0] step:9501/10000 train_time:414483ms step_avg:43.63ms
+[2025-09-04 11:34:26] [Rank 0] step:9521/10000 train_time:415250ms step_avg:43.61ms
+[2025-09-04 11:34:26] [Rank 0] step:9521/10000 train_time:415250ms step_avg:43.61ms
+[2025-09-04 11:34:27] [Rank 0] step:9541/10000 train_time:416005ms step_avg:43.60ms
+[2025-09-04 11:34:27] [Rank 0] step:9541/10000 train_time:416005ms step_avg:43.60ms
+[2025-09-04 11:34:28] [Rank 0] step:9561/10000 train_time:416760ms step_avg:43.59ms
+[2025-09-04 11:34:28] [Rank 0] step:9561/10000 train_time:416760ms step_avg:43.59ms
+[2025-09-04 11:34:28] [Rank 0] step:9581/10000 train_time:417514ms step_avg:43.58ms
+[2025-09-04 11:34:28] [Rank 0] step:9581/10000 train_time:417514ms step_avg:43.58ms
+[2025-09-04 11:34:29] [Rank 0] step:9601/10000 train_time:418269ms step_avg:43.57ms
+[2025-09-04 11:34:29] [Rank 0] step:9601/10000 train_time:418269ms step_avg:43.57ms
+[2025-09-04 11:34:30] [Rank 0] step:9621/10000 train_time:419024ms step_avg:43.55ms
+[2025-09-04 11:34:30] [Rank 0] step:9621/10000 train_time:419024ms step_avg:43.55ms
+[2025-09-04 11:34:31] [Rank 0] step:9641/10000 train_time:419779ms step_avg:43.54ms
+[2025-09-04 11:34:31] [Rank 0] step:9641/10000 train_time:419779ms step_avg:43.54ms
+[2025-09-04 11:34:32] [Rank 0] step:9661/10000 train_time:420814ms step_avg:43.56ms
+[2025-09-04 11:34:32] [Rank 0] step:9661/10000 train_time:420814ms step_avg:43.56ms
+[2025-09-04 11:34:33] [Rank 0] step:9681/10000 train_time:421569ms step_avg:43.55ms
+[2025-09-04 11:34:33] [Rank 0] step:9681/10000 train_time:421569ms step_avg:43.55ms
+[2025-09-04 11:34:33] [Rank 0] step:9701/10000 train_time:422324ms step_avg:43.53ms
+[2025-09-04 11:34:33] [Rank 0] step:9701/10000 train_time:422324ms step_avg:43.53ms
+[2025-09-04 11:34:34] [Rank 0] step:9721/10000 train_time:423078ms step_avg:43.52ms
+[2025-09-04 11:34:34] [Rank 0] step:9721/10000 train_time:423078ms step_avg:43.52ms
+[2025-09-04 11:34:35] [Rank 0] step:9741/10000 train_time:423833ms step_avg:43.51ms
+[2025-09-04 11:34:35] [Rank 0] step:9741/10000 train_time:423833ms step_avg:43.51ms
+[2025-09-04 11:34:36] [Rank 0] step:9761/10000 train_time:424587ms step_avg:43.50ms
+[2025-09-04 11:34:36] [Rank 0] step:9761/10000 train_time:424587ms step_avg:43.50ms
+[2025-09-04 11:34:36] [Rank 0] step:9781/10000 train_time:425342ms step_avg:43.49ms
+[2025-09-04 11:34:36] [Rank 0] step:9781/10000 train_time:425342ms step_avg:43.49ms
+[2025-09-04 11:34:37] [Rank 0] step:9801/10000 train_time:426096ms step_avg:43.47ms
+[2025-09-04 11:34:37] [Rank 0] step:9801/10000 train_time:426096ms step_avg:43.47ms
+[2025-09-04 11:34:38] [Rank 0] step:9821/10000 train_time:426851ms step_avg:43.46ms
+[2025-09-04 11:34:38] [Rank 0] step:9821/10000 train_time:426851ms step_avg:43.46ms
+[2025-09-04 11:34:39] [Rank 0] step:9841/10000 train_time:427606ms step_avg:43.45ms
+[2025-09-04 11:34:39] [Rank 0] step:9841/10000 train_time:427606ms step_avg:43.45ms
+[2025-09-04 11:34:39] [Rank 0] step:9861/10000 train_time:428360ms step_avg:43.44ms
+[2025-09-04 11:34:39] [Rank 0] step:9861/10000 train_time:428360ms step_avg:43.44ms
+[2025-09-04 11:34:40] [Rank 0] step:9881/10000 train_time:429115ms step_avg:43.43ms
+[2025-09-04 11:34:40] [Rank 0] step:9881/10000 train_time:429115ms step_avg:43.43ms
+[2025-09-04 11:34:41] [Rank 0] step:9901/10000 train_time:429870ms step_avg:43.42ms
+[2025-09-04 11:34:41] [Rank 0] step:9901/10000 train_time:429870ms step_avg:43.42ms
+[2025-09-04 11:34:42] [Rank 0] step:9921/10000 train_time:430625ms step_avg:43.41ms
+[2025-09-04 11:34:42] [Rank 0] step:9921/10000 train_time:430625ms step_avg:43.41ms
+[2025-09-04 11:34:42] [Rank 0] step:9941/10000 train_time:431381ms step_avg:43.39ms
+[2025-09-04 11:34:42] [Rank 0] step:9941/10000 train_time:431381ms step_avg:43.39ms
+[2025-09-04 11:34:43] [Rank 0] step:9961/10000 train_time:432136ms step_avg:43.38ms
+[2025-09-04 11:34:43] [Rank 0] step:9961/10000 train_time:432136ms step_avg:43.38ms
+[2025-09-04 11:34:44] [Rank 0] step:9981/10000 train_time:432891ms step_avg:43.37ms
+[2025-09-04 11:34:44] [Rank 0] step:9981/10000 train_time:432891ms step_avg:43.37ms
+[2025-09-04 11:34:45] [Rank 0] step:10000/10000 train_time:433609ms step_avg:43.36ms
+[2025-09-04 11:34:45] [Rank 0] step:10000/10000 train_time:433609ms step_avg:43.36ms
+[2025-09-04 11:34:45] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:34:45] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:34:45] [Rank 0] PRINT: step:10000/10000 train_loss:0.6093 val_loss:0.6055 train_time:433657ms step_avg:43.37ms
+[2025-09-04 11:34:45] [Rank 0] PRINT: step:10000/10000 train_loss:0.6093 val_loss:0.6055 train_time:433657ms step_avg:43.37ms
+[2025-09-04 11:34:45] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:34:45] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:34:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:34:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:36:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:36:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:36:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:36:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:36:24] [Rank 0]   Total Loss: 4.9576
+[2025-09-04 11:36:24] [Rank 0]   Total Loss: 4.9576
+[2025-09-04 11:36:24] [Rank 0]   Total FTA (Unweighted): 0.9975
+[2025-09-04 11:36:24] [Rank 0]   Total FTA (Unweighted): 0.9975
+[2025-09-04 11:36:24] [Rank 0]   Total FTA (Weighted):   0.9975
+[2025-09-04 11:36:24] [Rank 0]   Total FTA (Weighted):   0.9975
+[2025-09-04 11:36:24] [Rank 0]   Group 0 Loss: 4.8772
+[2025-09-04 11:36:24] [Rank 0]   Group 0 Loss: 4.8772
+[2025-09-04 11:36:24] [Rank 0]   Group 1 Loss: 4.5157
+[2025-09-04 11:36:24] [Rank 0]   Group 1 Loss: 4.5157
+[2025-09-04 11:36:24] [Rank 0]   Group 2 Loss: 4.5028
+[2025-09-04 11:36:24] [Rank 0]   Group 2 Loss: 4.5028
+[2025-09-04 11:36:24] [Rank 0]   Group 3 Loss: 4.8690
+[2025-09-04 11:36:24] [Rank 0]   Group 3 Loss: 4.8690
+[2025-09-04 11:36:24] [Rank 0]   Group 4 Loss: 4.7912
+[2025-09-04 11:36:24] [Rank 0]   Group 4 Loss: 4.7912
+[2025-09-04 11:36:24] [Rank 0]   Group 5 Loss: 4.8571
+[2025-09-04 11:36:24] [Rank 0]   Group 5 Loss: 4.8571
+[2025-09-04 11:36:24] [Rank 0]   Group 6 Loss: 4.7890
+[2025-09-04 11:36:24] [Rank 0]   Group 6 Loss: 4.7890
+[2025-09-04 11:36:24] [Rank 0]   Group 7 Loss: 4.8881
+[2025-09-04 11:36:24] [Rank 0]   Group 7 Loss: 4.8881
+[2025-09-04 11:36:24] [Rank 0]   Group 8 Loss: 5.0839
+[2025-09-04 11:36:24] [Rank 0]   Group 8 Loss: 5.0839
+[2025-09-04 11:36:24] [Rank 0]   Group 9 Loss: 5.0031
+[2025-09-04 11:36:24] [Rank 0]   Group 9 Loss: 5.0031
+[2025-09-04 11:36:24] [Rank 0]   Group 10 Loss: 5.2037
+[2025-09-04 11:36:24] [Rank 0]   Group 10 Loss: 5.2037
+[2025-09-04 11:36:24] [Rank 0]   Group 11 Loss: 5.1666
+[2025-09-04 11:36:24] [Rank 0]   Group 11 Loss: 5.1666
+[2025-09-04 11:36:24] [Rank 0]   Group 12 Loss: 5.1331
+[2025-09-04 11:36:24] [Rank 0]   Group 12 Loss: 5.1331
+[2025-09-04 11:36:24] [Rank 0]   Group 13 Loss: 5.2084
+[2025-09-04 11:36:24] [Rank 0]   Group 13 Loss: 5.2084
+[2025-09-04 11:36:24] [Rank 0]   Group 14 Loss: 5.1943
+[2025-09-04 11:36:24] [Rank 0]   Group 14 Loss: 5.1943
+[2025-09-04 11:36:24] [Rank 0]   Group 15 Loss: 5.2392
+[2025-09-04 11:36:24] [Rank 0]   Group 15 Loss: 5.2392
+[2025-09-04 11:36:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 11:36:24] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 11:36:24] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 11:36:24] [Rank 0]   Group 15 FTA: 0.9700
+[2025-09-04 11:36:24] [Rank 0]   Group 15 FTA: 0.9700
+[2025-09-04 11:36:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:36:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_loss_curves.png
+[2025-09-04 11:36:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:36:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/per_class_acc_curves.png
+[2025-09-04 11:36:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:36:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_loss_curve.png
+[2025-09-04 11:36:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:36:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_43/total_acc_curve.png
+[2025-09-04 11:36:25] [Rank 0] step:10001/10000 train_time:433672ms step_avg:43.36ms
+[2025-09-04 11:36:25] [Rank 0] step:10001/10000 train_time:433672ms step_avg:43.36ms
+[2025-09-04 11:36:25] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 11:36:25 2025 ---
+[2025-09-04 11:36:25] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 11:36:25 2025 ---
+[2025-09-04 11:36:25] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB
+[2025-09-04 11:36:25] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/config.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba0b3ab5c12a71c551b550207151fe71f7ed3c5c
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/config.json
@@ -0,0 +1,29 @@
+{
+    "cli_args": {
+        "unet": false,
+        "seed": 44,
+        "optimizer_mode": 10,
+        "model_parameterization": "qkvo",
+        "per_group_k": 100,
+        "muon_lr": 0.002,
+        "adam_lr": 0.002,
+        "base_dir": "logs_qa_muon/diff_modes",
+        "sgd_lr": 0.01,
+        "m_val": 15,
+        "qa_jsonl_path": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl"
+    },
+    "hyperparameters": {
+        "train_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin",
+        "val_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin",
+        "val_tokens": 491520,
+        "train_seq_len": 3072,
+        "val_seq_len": 16384,
+        "num_iterations": 10000,
+        "cooldown_frac": 0.8,
+        "vocab_size": 50257,
+        "val_loss_every": 500,
+        "save_checkpoint": false
+    },
+    "run_uuid_for_log": "d4aec5d3-43bf-44e4-bafb-5d9f46a06d8f",
+    "script_code_logged_at_start": true
+}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/fixed_eval_indices.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/fixed_eval_indices.json
new file mode 100644
index 0000000000000000000000000000000000000000..a823775225c5e592eb10700e5e0319b0491b1eb6
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/fixed_eval_indices.json
@@ -0,0 +1 @@
+{"1": [1238956, 182074, 1437575, 1061037, 383150, 1176376, 926, 823011, 832520, 1266421, 512738, 144357, 848076, 890204, 213997, 95146, 261767, 467731, 832231, 217985, 913168, 107253, 1361828, 61314, 1230420, 1133619, 146690, 429587, 419151, 58695, 1579770, 503799, 1421284, 882534, 1022637, 785343, 1154604, 67783, 1325109, 243941, 1213240, 438111, 460295, 269373, 538055, 1347006, 71775, 255496, 299906, 1227973, 815402, 190082, 1304077, 1023347, 613801, 983830, 1284420, 389321, 1625224, 717538, 1172273, 992184, 1181312, 1014039, 885952, 1538489, 158933, 1667270, 1250445, 958097, 1458224, 1306495, 62945, 733843, 1360200, 540493, 762461, 501460, 1208142, 1180559, 1333588, 690481, 355756, 618511, 733586, 650301, 799437, 165533, 1238977, 323078, 1485080, 609610, 1212241, 606952, 1253407, 1420922, 327112, 701, 777907, 1626516], "0": [1390189, 1220977, 1312259, 1201125, 1235379, 1272843, 344142, 1119560, 856330, 766456, 1145928, 550624, 684681, 1309879, 54620, 1126124, 1278077, 910610, 1532044, 73645, 460165, 462723, 457959, 476877, 320873, 1408316, 989156, 1276325, 338260, 577743, 1136972, 25115, 211215, 1296818, 968705, 436781, 318224, 385325, 1516441, 533464, 1628693, 703399, 579670, 1518978, 305154, 1625960, 1400284, 713841, 1652150, 811009, 336920, 881015, 512030, 1347531, 1005706, 412167, 173312, 1136816, 343017, 537797, 1489267, 935475, 703854, 1570584, 1504269, 1458356, 1310700, 1242505, 509976, 22096, 788120, 1616850, 1591198, 1151224, 797820, 823995, 1153364, 1060220, 727674, 580729, 148912, 334290, 686098, 1633151, 1130523, 1012440, 1374480, 953410, 1381730, 1376118, 932096, 739115, 739014, 263875, 1400308, 556177, 1341771, 348626, 575350, 254846], "10": [748592, 515936, 1645691, 329828, 1104748, 1453439, 406461, 1557507, 1146644, 1398421, 163667, 228310, 407739, 585218, 785725, 1371077, 134834, 15431, 1303784, 273099, 1530332, 1320219, 1329247, 753603, 664548, 501873, 1512467, 583734, 1355643, 600440, 428998, 1426876, 121341, 1456507, 892281, 1242339, 1304508, 968336, 1593981, 634471, 1253799, 593156, 1486078, 712425, 441285, 618217, 1024661, 1395408, 1423117, 1463918, 121752, 800727, 52402, 1185143, 1651644, 1583572, 1302889, 1040220, 1234122, 333457, 1552652, 1567291, 1405501, 222708, 1469496, 875611, 894032, 783951, 934281, 109533, 198744, 643064, 197926, 224405, 1169998, 956511, 1096712, 1213224, 828780, 204659, 564225, 257079, 754287, 1598723, 821179, 1156648, 60674, 290993, 1215838, 1556010, 1025931, 687671, 252775, 771448, 574110, 578615, 279536, 1503737, 1089787, 2338], "14": [1602125, 1388977, 709154, 666138, 1446614, 572103, 99632, 25794, 1487658, 15077, 1600205, 682155, 679850, 1462049, 970283, 81336, 378196, 1627374, 832926, 1217832, 740140, 1336941, 793290, 1441127, 297494, 303372, 339524, 1363322, 525184, 593339, 61055, 1468219, 1110310, 1061855, 1046762, 410655, 449257, 530144, 1362373, 241552, 926068, 1618007, 629406, 327589, 263223, 1525290, 621372, 606355, 534601, 126046, 934024, 1181343, 1234759, 295773, 838035, 1447427, 1395837, 1568116, 486923, 465204, 98192, 1145989, 271529, 720487, 1091877, 653313, 872526, 13890, 111971, 671086, 301528, 50698, 455373, 786098, 329199, 599178, 644737, 1404606, 33829, 279256, 161938, 309352, 1537653, 21723, 309413, 478184, 340938, 1080124, 1375458, 1301528, 331423, 212589, 1220931, 112707, 454022, 1466562, 1238127, 341561, 1497264, 247245], "3": [143966, 715765, 455823, 1414629, 814361, 278667, 700211, 1286614, 1601013, 923359, 1105960, 549680, 371512, 732263, 1236795, 432942, 1379421, 488467, 679369, 1200930, 479010, 683138, 700149, 1422714, 348019, 1263383, 1044275, 820031, 969830, 75153, 1410929, 394116, 1486865, 299414, 264214, 1416493, 428688, 107076, 753671, 674417, 1456072, 307080, 1259105, 1563344, 1198484, 456551, 1061513, 1128111, 922541, 1040277, 1360208, 1411429, 259514, 883801, 716781, 1004987, 285196, 266162, 781115, 385241, 252691, 1398963, 1440582, 1300653, 756663, 231617, 1319447, 42576, 1434337, 20658, 581284, 178420, 1626533, 694079, 1531414, 1631177, 1478512, 647215, 706568, 303323, 10540, 1010611, 1150313, 159016, 1614581, 1257722, 470491, 1333309, 94699, 756222, 1456446, 296132, 1186414, 1591257, 1155050, 789439, 1472373, 1403759, 1451276, 745193], "4": [1227524, 563788, 436964, 1380176, 1648232, 267090, 273744, 1031517, 1580586, 416846, 1309866, 114170, 1649739, 1161604, 690741, 1432068, 224570, 322835, 1559902, 479718, 709757, 410142, 1668701, 1143278, 968258, 406289, 884233, 1080971, 269480, 631196, 421463, 374502, 430292, 1309826, 1422997, 1439354, 131354, 975504, 84660, 1590390, 288064, 1419944, 695652, 320713, 260376, 222416, 937739, 1104508, 1323311, 962301, 1060820, 1117857, 280792, 692359, 656651, 358878, 828406, 598701, 127434, 1511259, 1332605, 1178325, 1281187, 203589, 792701, 684176, 543034, 980077, 982509, 640126, 1252418, 1529627, 519645, 359068, 992779, 550843, 1493488, 447632, 1512599, 361760, 626672, 89042, 604248, 574827, 87735, 1432602, 473061, 668609, 925422, 311417, 71495, 682067, 221270, 600053, 673169, 1408933, 719302, 788676, 998101, 243902], "15": [1444113, 1102795, 1149896, 1181364, 349047, 1076908, 816006, 945310, 197585, 276574, 946546, 1505399, 740223, 1154813, 1312891, 1261018, 249986, 1504592, 647427, 915615, 1245136, 1288301, 802239, 410533, 805664, 1013403, 498363, 777073, 1508176, 18111, 1147340, 880289, 215773, 924760, 1301970, 318135, 1524489, 103475, 872748, 1568512, 722849, 1364157, 838956, 1451408, 970208, 1350756, 1498244, 980461, 1098677, 1141812, 463859, 95672, 1220365, 554451, 524014, 1161048, 17478, 113548, 1027741, 1657643, 1462809, 1279104, 85588, 241936, 510707, 151817, 1191227, 574942, 735662, 1046248, 1056729, 96279, 781378, 1042711, 1602339, 1503514, 299087, 898968, 553537, 293167, 897565, 572125, 425215, 916907, 1210531, 1013275, 201532, 883447, 1641610, 1185682, 425791, 831133, 1291227, 1099064, 934983, 588125, 1284591, 1163873, 524656, 1230762], "13": [1122510, 535854, 845319, 585144, 884184, 441130, 934654, 1046629, 1512962, 1190830, 1462685, 90530, 894615, 1527783, 415323, 90852, 1556905, 1002406, 366356, 1386022, 846952, 928360, 3138, 406205, 78588, 1208032, 1210206, 257177, 137047, 711380, 693628, 105893, 392053, 1301525, 677934, 1549630, 1134520, 447564, 1194061, 672969, 740755, 1045536, 633455, 1336343, 613147, 1161931, 234095, 1246643, 643373, 834221, 1351442, 1502069, 1401555, 759927, 1320337, 666063, 291222, 956633, 400583, 963313, 683474, 534882, 1101312, 1519551, 1513194, 752124, 1028953, 1383362, 169732, 785294, 160568, 1053580, 774534, 372048, 292049, 791470, 1318537, 141966, 510573, 1440830, 930544, 1352026, 463299, 1245306, 159755, 1566413, 316835, 1135200, 266962, 112892, 253603, 1328380, 703630, 21798, 268463, 7112, 1232576, 374978, 240715, 1299198], "2": [463910, 845243, 985993, 402000, 962033, 473711, 108588, 1074932, 539045, 1162793, 1086008, 778638, 751876, 53650, 465364, 803834, 820717, 9490, 473708, 26912, 970677, 1441139, 231484, 749139, 286603, 1346834, 1641988, 699767, 816768, 1220504, 978431, 1331006, 611334, 829473, 271532, 566760, 332292, 1326050, 887731, 597530, 940973, 491159, 1150471, 122010, 225465, 382413, 1207486, 1568920, 992793, 896409, 1479386, 1539347, 759024, 293232, 1345256, 1282360, 956014, 371580, 1202087, 1638038, 892501, 38073, 1651323, 1103388, 1315239, 79005, 1045045, 1230831, 1206253, 1376976, 140037, 320204, 274836, 663525, 397448, 666367, 1501713, 570453, 836424, 694423, 157979, 279074, 1235605, 536170, 301830, 778556, 1069897, 181207, 1445408, 1082031, 969552, 1083538, 1395871, 168187, 5476, 514815, 194418, 1038889, 1288748, 1535767], "8": [258910, 518286, 80108, 811775, 937854, 1623550, 1592007, 661848, 1484674, 300, 104226, 1075215, 1190104, 8885, 1062956, 88002, 1020623, 189243, 1265153, 1472857, 301412, 21371, 441401, 1078954, 1648642, 894192, 1083738, 1182805, 1347655, 316568, 1063971, 255177, 186613, 550617, 102173, 563694, 465966, 965224, 372481, 929520, 1423441, 6575, 862309, 144894, 891435, 123628, 653301, 1378790, 1247156, 1129087, 1381991, 410876, 391895, 202136, 868149, 233421, 564464, 1174972, 1385073, 238326, 483488, 507729, 1329666, 955266, 479623, 902401, 1343890, 827390, 371073, 1480060, 1241650, 1670957, 644144, 1371709, 1324243, 670126, 23715, 1566067, 70858, 522312, 558100, 301714, 359605, 1401913, 1060057, 1027855, 987700, 720084, 324937, 562646, 390910, 1398541, 1097298, 1626769, 1461630, 1430950, 650075, 1025475, 274735, 650932], "7": [181761, 629797, 231148, 365747, 1229714, 500408, 198146, 500703, 1250728, 1578820, 753672, 84476, 1379367, 355424, 1591855, 1400949, 1194404, 912060, 884197, 320928, 1096692, 714201, 331003, 1345763, 1620998, 457231, 1542074, 1108477, 1602113, 641593, 542514, 159345, 360278, 35579, 506, 470465, 1606024, 1349783, 731258, 382125, 883800, 1652746, 382071, 99241, 20216, 1398849, 245155, 1601928, 1544592, 1632383, 1597152, 1266758, 1032571, 88287, 521379, 1423306, 579207, 1152743, 271926, 418168, 8827, 1358019, 429579, 380995, 1040420, 1067138, 283335, 1150469, 280745, 1184794, 468245, 948227, 357499, 527525, 1067827, 816862, 360563, 433490, 258517, 581805, 1089641, 1537081, 473494, 1480784, 60701, 686908, 614323, 143069, 738219, 1234467, 927103, 64088, 658420, 1353941, 1517049, 753334, 198407, 434613, 850731, 1211609], "5": [1250448, 711876, 285881, 342500, 987883, 195047, 810508, 356140, 999602, 284542, 730936, 736191, 661291, 11822, 551928, 617679, 384856, 414906, 238071, 473605, 868498, 1269087, 558407, 245083, 810169, 1456559, 1657433, 70018, 865238, 416303, 1389734, 1561961, 534163, 1098802, 1494214, 1285484, 599936, 296819, 696611, 307176, 29697, 731071, 257656, 1465441, 1668863, 166995, 1668377, 1457645, 1400335, 1525120, 33323, 437285, 459069, 620176, 231704, 846424, 1391144, 364636, 63552, 1093702, 67741, 285428, 454922, 719128, 1393391, 87651, 1418561, 1312176, 652985, 1446762, 71280, 1594253, 89848, 628629, 1117344, 882598, 1014283, 1601256, 739004, 508965, 1273914, 44245, 1267971, 1430470, 611385, 930900, 1439585, 1545682, 1300966, 1331569, 364367, 27900, 107616, 1579070, 94723, 818780, 1183546, 384575, 142627, 431847], "11": [12772, 1091320, 1136723, 560663, 715094, 1097800, 1016547, 1299126, 961155, 725788, 999768, 1277062, 650665, 1323964, 1024796, 1255124, 215111, 268904, 442949, 358645, 621958, 1666210, 1154349, 375517, 614870, 1166618, 1418680, 795635, 184985, 129300, 129366, 1343289, 1151524, 193531, 634856, 1618906, 43721, 1601844, 591772, 1322095, 1522284, 431856, 574972, 1544663, 1389257, 1113893, 1515857, 614345, 809163, 132406, 95327, 997226, 423104, 910705, 261584, 765030, 1216397, 439297, 1565563, 1595785, 1622521, 975588, 1314965, 1205327, 1501612, 1642358, 656493, 1369509, 195325, 1554913, 1661675, 790468, 511387, 1227059, 1219143, 1118001, 788451, 697651, 1575073, 1345104, 167862, 1409843, 138942, 984075, 1382364, 1299538, 214601, 1541475, 459076, 353219, 1310728, 757670, 904351, 334019, 1003259, 670104, 10386, 13556, 1097931, 1526407], "6": [295849, 644897, 1207883, 1648549, 1478750, 1486750, 1156499, 621047, 1387936, 662848, 1362507, 891859, 1188350, 1298200, 1044010, 767288, 1191230, 683760, 533545, 498960, 1421819, 366846, 54389, 1382045, 1002942, 1285004, 457209, 1074763, 823805, 1508295, 719970, 837720, 19881, 1189285, 1226048, 1078656, 1433699, 14805, 254682, 614235, 805087, 900964, 343832, 561709, 243288, 465867, 497085, 1153456, 730115, 42050, 878511, 1382508, 667820, 712856, 715963, 1499782, 1516002, 444399, 726168, 1204910, 591639, 986194, 107160, 1596752, 1239812, 591854, 536489, 1347905, 1597670, 122237, 620494, 354940, 1561646, 1352115, 74785, 163277, 590349, 1429584, 1440950, 848184, 1580723, 1631128, 979479, 758861, 84655, 1079672, 673692, 1315324, 167985, 1353485, 21815, 1175739, 1360408, 1062086, 1369999, 733863, 1371964, 61279, 1152056, 220910], "9": [780278, 1156801, 827882, 1095960, 874763, 1330006, 979412, 1154745, 1399441, 515095, 1505401, 1454104, 1563117, 1316054, 57098, 43367, 1175749, 1257252, 865038, 624410, 186727, 907758, 126168, 1353431, 167806, 1021404, 1585615, 846939, 552247, 1400471, 68525, 190495, 1556857, 181528, 47677, 136468, 189446, 269049, 262467, 503841, 1367168, 382007, 486946, 621235, 622683, 469399, 1537192, 707594, 1003091, 56021, 1354302, 799290, 1618221, 696794, 836658, 1591562, 251096, 1321886, 412122, 1352217, 946767, 1267324, 213504, 747174, 1403653, 1589967, 849743, 549269, 663800, 1315189, 451416, 1330881, 1352197, 1517840, 827169, 1228725, 1097725, 955944, 1375509, 1590213, 266039, 1558465, 1498493, 604161, 608723, 1547013, 1484549, 1530909, 1325336, 1584866, 630737, 1023780, 227549, 961233, 987591, 159267, 608424, 636687, 745397, 460687], "12": [1586709, 357745, 1274743, 202991, 1250469, 837216, 757864, 185235, 493474, 621351, 22706, 1096491, 186138, 1482718, 534525, 1157131, 1350934, 1167912, 1366603, 756644, 311945, 72417, 706444, 1464146, 1236814, 74555, 983713, 1182807, 1082263, 804850, 1028788, 1405036, 932746, 742978, 1589076, 723766, 898380, 906214, 913629, 33386, 256107, 1245317, 360574, 836359, 565572, 400905, 1269191, 1042959, 886332, 1415571, 68870, 470376, 1531, 1333957, 834102, 674685, 839119, 255119, 552181, 1239034, 490552, 468715, 963242, 1612174, 21991, 1157603, 975554, 1342832, 883818, 1475260, 1666539, 1252155, 986632, 906429, 1495805, 561464, 1219097, 567704, 1463586, 1571831, 689821, 481310, 1407520, 97442, 512704, 962995, 876670, 956570, 1292534, 914981, 1655195, 870222, 196978, 617479, 1610748, 128227, 213356, 1590232, 230140, 942835]}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..42a262817aeb99287d747c8170a5a33cf97be0a6
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14f4ff4dc1bfd42616348815ea2cc76c1651e2c4a28774f98549d2c536866149
+size 352941
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..d63636f5915976df5d628c251759047ae7464544
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34824762a7d1993b41d2b5aec13004d2cd04e432f8bd9aee88fc3c920b446fe8
+size 438101
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..fc0072c18ad4a0b93c39e8c5643868614faa1253
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:800e0930b9c1d54ad742464cf5a13146114b6615c078052bb71f6ecaeb8aa97a
+size 96854
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..79299b704c111f5dfd4c35bf2b524888d22f48df
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e1d9cfb461a1185155f681008efd883458d552287d6435bd37723a293b8e108
+size 112281
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/training_log_d4aec5d3-43bf-44e4-bafb-5d9f46a06d8f.txt b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/training_log_d4aec5d3-43bf-44e4-bafb-5d9f46a06d8f.txt
new file mode 100644
index 0000000000000000000000000000000000000000..bc22c25eedfe18687935f05a8faf9c6a701d1216
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/training_log_d4aec5d3-43bf-44e4-bafb-5d9f46a06d8f.txt
@@ -0,0 +1,5236 @@
+[2025-09-04 11:36:46] [Rank 0] PRINT: --- Script Start: Thu Sep  4 11:36:46 2025 ---
+[2025-09-04 11:36:46] [Rank 0] PRINT: --- Script Start: Thu Sep  4 11:36:46 2025 ---
+[2025-09-04 11:36:46] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=44, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 11:36:46] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=44, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 11:36:46] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 11:36:46] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 11:36:46] [Rank 0] PRINT: Using fixed seed: 44
+[2025-09-04 11:36:46] [Rank 0] PRINT: Using fixed seed: 44
+[2025-09-04 11:36:46] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44
+[2025-09-04 11:36:46] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44
+[2025-09-04 11:36:46] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 11:36:46] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 11:36:46] [Rank 0] PRINT: Constructing model...
+[2025-09-04 11:36:46] [Rank 0] PRINT: Constructing model...
+[2025-09-04 11:36:48] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 11:36:48] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 11:36:48] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 11:36:48] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 11:36:48] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 11:36:48] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 11:36:51] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 11:36:51] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 11:36:51] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 11:36:51] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 11:36:51] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 11:36:51] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 11:36:51] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 11:36:51] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 11:36:51] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 11:36:51] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 11:36:51] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 11:36:51] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 11:36:51] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 11:36:51] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 11:36:51] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 11:36:51] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 11:36:51] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 11:36:51] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 11:36:51] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 11:36:51] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 11:36:51] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 11:36:51] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 11:36:55] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 11:36:55] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 11:36:55] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 11:36:55] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 11:39:05] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 11:39:05] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 11:39:05] [Rank 0] PRINT: Starting training...
+[2025-09-04 11:39:05] [Rank 0] PRINT: Starting training...
+[2025-09-04 11:39:12] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/fixed_eval_indices.json
+[2025-09-04 11:39:12] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/fixed_eval_indices.json
+[2025-09-04 11:39:12] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:39:12] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:39:16] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 11:39:16] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 11:39:53] [Rank 0] step:21/10000 train_time:37115ms step_avg:1767.36ms
+[2025-09-04 11:39:53] [Rank 0] step:21/10000 train_time:37115ms step_avg:1767.36ms
+[2025-09-04 11:39:54] [Rank 0] step:41/10000 train_time:37855ms step_avg:923.30ms
+[2025-09-04 11:39:54] [Rank 0] step:41/10000 train_time:37855ms step_avg:923.30ms
+[2025-09-04 11:39:55] [Rank 0] step:61/10000 train_time:38595ms step_avg:632.71ms
+[2025-09-04 11:39:55] [Rank 0] step:61/10000 train_time:38595ms step_avg:632.71ms
+[2025-09-04 11:39:55] [Rank 0] step:81/10000 train_time:39336ms step_avg:485.63ms
+[2025-09-04 11:39:55] [Rank 0] step:81/10000 train_time:39336ms step_avg:485.63ms
+[2025-09-04 11:39:56] [Rank 0] step:101/10000 train_time:40077ms step_avg:396.80ms
+[2025-09-04 11:39:56] [Rank 0] step:101/10000 train_time:40077ms step_avg:396.80ms
+[2025-09-04 11:39:57] [Rank 0] step:121/10000 train_time:40817ms step_avg:337.33ms
+[2025-09-04 11:39:57] [Rank 0] step:121/10000 train_time:40817ms step_avg:337.33ms
+[2025-09-04 11:39:58] [Rank 0] step:141/10000 train_time:41559ms step_avg:294.75ms
+[2025-09-04 11:39:58] [Rank 0] step:141/10000 train_time:41559ms step_avg:294.75ms
+[2025-09-04 11:39:58] [Rank 0] step:161/10000 train_time:42299ms step_avg:262.73ms
+[2025-09-04 11:39:58] [Rank 0] step:161/10000 train_time:42299ms step_avg:262.73ms
+[2025-09-04 11:39:59] [Rank 0] step:181/10000 train_time:43039ms step_avg:237.78ms
+[2025-09-04 11:39:59] [Rank 0] step:181/10000 train_time:43039ms step_avg:237.78ms
+[2025-09-04 11:40:00] [Rank 0] step:201/10000 train_time:43780ms step_avg:217.81ms
+[2025-09-04 11:40:00] [Rank 0] step:201/10000 train_time:43780ms step_avg:217.81ms
+[2025-09-04 11:40:01] [Rank 0] step:221/10000 train_time:44521ms step_avg:201.45ms
+[2025-09-04 11:40:01] [Rank 0] step:221/10000 train_time:44521ms step_avg:201.45ms
+[2025-09-04 11:40:01] [Rank 0] step:241/10000 train_time:45306ms step_avg:187.99ms
+[2025-09-04 11:40:01] [Rank 0] step:241/10000 train_time:45306ms step_avg:187.99ms
+[2025-09-04 11:40:02] [Rank 0] step:261/10000 train_time:46091ms step_avg:176.59ms
+[2025-09-04 11:40:02] [Rank 0] step:261/10000 train_time:46091ms step_avg:176.59ms
+[2025-09-04 11:40:03] [Rank 0] step:281/10000 train_time:46831ms step_avg:166.66ms
+[2025-09-04 11:40:03] [Rank 0] step:281/10000 train_time:46831ms step_avg:166.66ms
+[2025-09-04 11:40:04] [Rank 0] step:301/10000 train_time:47572ms step_avg:158.05ms
+[2025-09-04 11:40:04] [Rank 0] step:301/10000 train_time:47572ms step_avg:158.05ms
+[2025-09-04 11:40:04] [Rank 0] step:321/10000 train_time:48313ms step_avg:150.51ms
+[2025-09-04 11:40:04] [Rank 0] step:321/10000 train_time:48313ms step_avg:150.51ms
+[2025-09-04 11:40:05] [Rank 0] step:341/10000 train_time:49054ms step_avg:143.85ms
+[2025-09-04 11:40:05] [Rank 0] step:341/10000 train_time:49054ms step_avg:143.85ms
+[2025-09-04 11:40:06] [Rank 0] step:361/10000 train_time:49794ms step_avg:137.93ms
+[2025-09-04 11:40:06] [Rank 0] step:361/10000 train_time:49794ms step_avg:137.93ms
+[2025-09-04 11:40:07] [Rank 0] step:381/10000 train_time:50536ms step_avg:132.64ms
+[2025-09-04 11:40:07] [Rank 0] step:381/10000 train_time:50536ms step_avg:132.64ms
+[2025-09-04 11:40:07] [Rank 0] step:401/10000 train_time:51277ms step_avg:127.87ms
+[2025-09-04 11:40:07] [Rank 0] step:401/10000 train_time:51277ms step_avg:127.87ms
+[2025-09-04 11:40:08] [Rank 0] step:421/10000 train_time:52018ms step_avg:123.56ms
+[2025-09-04 11:40:08] [Rank 0] step:421/10000 train_time:52018ms step_avg:123.56ms
+[2025-09-04 11:40:09] [Rank 0] step:441/10000 train_time:52759ms step_avg:119.63ms
+[2025-09-04 11:40:09] [Rank 0] step:441/10000 train_time:52759ms step_avg:119.63ms
+[2025-09-04 11:40:10] [Rank 0] step:461/10000 train_time:53499ms step_avg:116.05ms
+[2025-09-04 11:40:10] [Rank 0] step:461/10000 train_time:53499ms step_avg:116.05ms
+[2025-09-04 11:40:10] [Rank 0] step:481/10000 train_time:54240ms step_avg:112.76ms
+[2025-09-04 11:40:10] [Rank 0] step:481/10000 train_time:54240ms step_avg:112.76ms
+[2025-09-04 11:40:11] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:40:11] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:40:12] [Rank 0] PRINT: step:500/10000 train_loss:3.1280 val_loss:1.1143 train_time:54986ms step_avg:109.97ms
+[2025-09-04 11:40:12] [Rank 0] PRINT: step:500/10000 train_loss:3.1280 val_loss:1.1143 train_time:54986ms step_avg:109.97ms
+[2025-09-04 11:40:12] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:40:12] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:40:12] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:40:12] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:41:49] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:41:49] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:41:49] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:41:49] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:41:49] [Rank 0]   Total Loss: 3.8480
+[2025-09-04 11:41:49] [Rank 0]   Total Loss: 3.8480
+[2025-09-04 11:41:49] [Rank 0]   Total FTA (Unweighted): 0.4762
+[2025-09-04 11:41:49] [Rank 0]   Total FTA (Unweighted): 0.4762
+[2025-09-04 11:41:49] [Rank 0]   Total FTA (Weighted):   0.4763
+[2025-09-04 11:41:49] [Rank 0]   Total FTA (Weighted):   0.4763
+[2025-09-04 11:41:49] [Rank 0]   Group 0 Loss: 3.3748
+[2025-09-04 11:41:49] [Rank 0]   Group 0 Loss: 3.3748
+[2025-09-04 11:41:49] [Rank 0]   Group 1 Loss: 3.1688
+[2025-09-04 11:41:49] [Rank 0]   Group 1 Loss: 3.1688
+[2025-09-04 11:41:49] [Rank 0]   Group 2 Loss: 3.0756
+[2025-09-04 11:41:49] [Rank 0]   Group 2 Loss: 3.0756
+[2025-09-04 11:41:49] [Rank 0]   Group 3 Loss: 3.3522
+[2025-09-04 11:41:49] [Rank 0]   Group 3 Loss: 3.3522
+[2025-09-04 11:41:49] [Rank 0]   Group 4 Loss: 3.4471
+[2025-09-04 11:41:49] [Rank 0]   Group 4 Loss: 3.4471
+[2025-09-04 11:41:49] [Rank 0]   Group 5 Loss: 3.5337
+[2025-09-04 11:41:49] [Rank 0]   Group 5 Loss: 3.5337
+[2025-09-04 11:41:49] [Rank 0]   Group 6 Loss: 3.5510
+[2025-09-04 11:41:49] [Rank 0]   Group 6 Loss: 3.5510
+[2025-09-04 11:41:49] [Rank 0]   Group 7 Loss: 3.7014
+[2025-09-04 11:41:49] [Rank 0]   Group 7 Loss: 3.7014
+[2025-09-04 11:41:49] [Rank 0]   Group 8 Loss: 3.9809
+[2025-09-04 11:41:49] [Rank 0]   Group 8 Loss: 3.9809
+[2025-09-04 11:41:49] [Rank 0]   Group 9 Loss: 4.0437
+[2025-09-04 11:41:49] [Rank 0]   Group 9 Loss: 4.0437
+[2025-09-04 11:41:49] [Rank 0]   Group 10 Loss: 4.2605
+[2025-09-04 11:41:49] [Rank 0]   Group 10 Loss: 4.2605
+[2025-09-04 11:41:49] [Rank 0]   Group 11 Loss: 4.2839
+[2025-09-04 11:41:49] [Rank 0]   Group 11 Loss: 4.2839
+[2025-09-04 11:41:49] [Rank 0]   Group 12 Loss: 4.3567
+[2025-09-04 11:41:49] [Rank 0]   Group 12 Loss: 4.3567
+[2025-09-04 11:41:49] [Rank 0]   Group 13 Loss: 4.4905
+[2025-09-04 11:41:49] [Rank 0]   Group 13 Loss: 4.4905
+[2025-09-04 11:41:49] [Rank 0]   Group 14 Loss: 4.4653
+[2025-09-04 11:41:49] [Rank 0]   Group 14 Loss: 4.4653
+[2025-09-04 11:41:49] [Rank 0]   Group 15 Loss: 4.4811
+[2025-09-04 11:41:49] [Rank 0]   Group 15 Loss: 4.4811
+[2025-09-04 11:41:49] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:41:49] [Rank 0]   Group 4 FTA: 0.9200
+[2025-09-04 11:41:49] [Rank 0]   Group 4 FTA: 0.9200
+[2025-09-04 11:41:49] [Rank 0]   Group 5 FTA: 0.5900
+[2025-09-04 11:41:49] [Rank 0]   Group 5 FTA: 0.5900
+[2025-09-04 11:41:49] [Rank 0]   Group 6 FTA: 0.5000
+[2025-09-04 11:41:49] [Rank 0]   Group 6 FTA: 0.5000
+[2025-09-04 11:41:49] [Rank 0]   Group 7 FTA: 0.4100
+[2025-09-04 11:41:49] [Rank 0]   Group 7 FTA: 0.4100
+[2025-09-04 11:41:49] [Rank 0]   Group 8 FTA: 0.3500
+[2025-09-04 11:41:49] [Rank 0]   Group 8 FTA: 0.3500
+[2025-09-04 11:41:49] [Rank 0]   Group 9 FTA: 0.2200
+[2025-09-04 11:41:49] [Rank 0]   Group 9 FTA: 0.2200
+[2025-09-04 11:41:49] [Rank 0]   Group 10 FTA: 0.1300
+[2025-09-04 11:41:49] [Rank 0]   Group 10 FTA: 0.1300
+[2025-09-04 11:41:49] [Rank 0]   Group 11 FTA: 0.0800
+[2025-09-04 11:41:49] [Rank 0]   Group 11 FTA: 0.0800
+[2025-09-04 11:41:49] [Rank 0]   Group 12 FTA: 0.0900
+[2025-09-04 11:41:49] [Rank 0]   Group 12 FTA: 0.0900
+[2025-09-04 11:41:49] [Rank 0]   Group 13 FTA: 0.1300
+[2025-09-04 11:41:49] [Rank 0]   Group 13 FTA: 0.1300
+[2025-09-04 11:41:49] [Rank 0]   Group 14 FTA: 0.1100
+[2025-09-04 11:41:49] [Rank 0]   Group 14 FTA: 0.1100
+[2025-09-04 11:41:49] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 11:41:49] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 11:41:50] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:41:50] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:41:50] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:41:50] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:41:50] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:41:50] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:41:50] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:41:50] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:41:50] [Rank 0] step:501/10000 train_time:55002ms step_avg:109.78ms
+[2025-09-04 11:41:50] [Rank 0] step:501/10000 train_time:55002ms step_avg:109.78ms
+[2025-09-04 11:41:51] [Rank 0] step:521/10000 train_time:55739ms step_avg:106.98ms
+[2025-09-04 11:41:51] [Rank 0] step:521/10000 train_time:55739ms step_avg:106.98ms
+[2025-09-04 11:41:52] [Rank 0] step:541/10000 train_time:56480ms step_avg:104.40ms
+[2025-09-04 11:41:52] [Rank 0] step:541/10000 train_time:56480ms step_avg:104.40ms
+[2025-09-04 11:41:53] [Rank 0] step:561/10000 train_time:57221ms step_avg:102.00ms
+[2025-09-04 11:41:53] [Rank 0] step:561/10000 train_time:57221ms step_avg:102.00ms
+[2025-09-04 11:41:53] [Rank 0] step:581/10000 train_time:57961ms step_avg:99.76ms
+[2025-09-04 11:41:53] [Rank 0] step:581/10000 train_time:57961ms step_avg:99.76ms
+[2025-09-04 11:41:54] [Rank 0] step:601/10000 train_time:58702ms step_avg:97.67ms
+[2025-09-04 11:41:54] [Rank 0] step:601/10000 train_time:58702ms step_avg:97.67ms
+[2025-09-04 11:41:55] [Rank 0] step:621/10000 train_time:59443ms step_avg:95.72ms
+[2025-09-04 11:41:55] [Rank 0] step:621/10000 train_time:59443ms step_avg:95.72ms
+[2025-09-04 11:41:56] [Rank 0] step:641/10000 train_time:60183ms step_avg:93.89ms
+[2025-09-04 11:41:56] [Rank 0] step:641/10000 train_time:60183ms step_avg:93.89ms
+[2025-09-04 11:41:56] [Rank 0] step:661/10000 train_time:60924ms step_avg:92.17ms
+[2025-09-04 11:41:56] [Rank 0] step:661/10000 train_time:60924ms step_avg:92.17ms
+[2025-09-04 11:41:57] [Rank 0] step:681/10000 train_time:61665ms step_avg:90.55ms
+[2025-09-04 11:41:57] [Rank 0] step:681/10000 train_time:61665ms step_avg:90.55ms
+[2025-09-04 11:41:58] [Rank 0] step:701/10000 train_time:62407ms step_avg:89.03ms
+[2025-09-04 11:41:58] [Rank 0] step:701/10000 train_time:62407ms step_avg:89.03ms
+[2025-09-04 11:41:59] [Rank 0] step:721/10000 train_time:63149ms step_avg:87.59ms
+[2025-09-04 11:41:59] [Rank 0] step:721/10000 train_time:63149ms step_avg:87.59ms
+[2025-09-04 11:41:59] [Rank 0] step:741/10000 train_time:63891ms step_avg:86.22ms
+[2025-09-04 11:41:59] [Rank 0] step:741/10000 train_time:63891ms step_avg:86.22ms
+[2025-09-04 11:42:00] [Rank 0] step:761/10000 train_time:64636ms step_avg:84.94ms
+[2025-09-04 11:42:00] [Rank 0] step:761/10000 train_time:64636ms step_avg:84.94ms
+[2025-09-04 11:42:01] [Rank 0] step:781/10000 train_time:65382ms step_avg:83.72ms
+[2025-09-04 11:42:01] [Rank 0] step:781/10000 train_time:65382ms step_avg:83.72ms
+[2025-09-04 11:42:02] [Rank 0] step:801/10000 train_time:66131ms step_avg:82.56ms
+[2025-09-04 11:42:02] [Rank 0] step:801/10000 train_time:66131ms step_avg:82.56ms
+[2025-09-04 11:42:03] [Rank 0] step:821/10000 train_time:67149ms step_avg:81.79ms
+[2025-09-04 11:42:03] [Rank 0] step:821/10000 train_time:67149ms step_avg:81.79ms
+[2025-09-04 11:42:03] [Rank 0] step:841/10000 train_time:67895ms step_avg:80.73ms
+[2025-09-04 11:42:03] [Rank 0] step:841/10000 train_time:67895ms step_avg:80.73ms
+[2025-09-04 11:42:04] [Rank 0] step:861/10000 train_time:68641ms step_avg:79.72ms
+[2025-09-04 11:42:04] [Rank 0] step:861/10000 train_time:68641ms step_avg:79.72ms
+[2025-09-04 11:42:05] [Rank 0] step:881/10000 train_time:69386ms step_avg:78.76ms
+[2025-09-04 11:42:05] [Rank 0] step:881/10000 train_time:69386ms step_avg:78.76ms
+[2025-09-04 11:42:06] [Rank 0] step:901/10000 train_time:70131ms step_avg:77.84ms
+[2025-09-04 11:42:06] [Rank 0] step:901/10000 train_time:70131ms step_avg:77.84ms
+[2025-09-04 11:42:06] [Rank 0] step:921/10000 train_time:70877ms step_avg:76.96ms
+[2025-09-04 11:42:06] [Rank 0] step:921/10000 train_time:70877ms step_avg:76.96ms
+[2025-09-04 11:42:07] [Rank 0] step:941/10000 train_time:71623ms step_avg:76.11ms
+[2025-09-04 11:42:07] [Rank 0] step:941/10000 train_time:71623ms step_avg:76.11ms
+[2025-09-04 11:42:08] [Rank 0] step:961/10000 train_time:72369ms step_avg:75.31ms
+[2025-09-04 11:42:08] [Rank 0] step:961/10000 train_time:72369ms step_avg:75.31ms
+[2025-09-04 11:42:09] [Rank 0] step:981/10000 train_time:73113ms step_avg:74.53ms
+[2025-09-04 11:42:09] [Rank 0] step:981/10000 train_time:73113ms step_avg:74.53ms
+[2025-09-04 11:42:09] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:42:09] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:42:10] [Rank 0] PRINT: step:1000/10000 train_loss:0.9677 val_loss:0.8626 train_time:73864ms step_avg:73.86ms
+[2025-09-04 11:42:10] [Rank 0] PRINT: step:1000/10000 train_loss:0.9677 val_loss:0.8626 train_time:73864ms step_avg:73.86ms
+[2025-09-04 11:42:10] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:42:10] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:42:10] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:42:10] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:43:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:43:47] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:43:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:43:47] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:43:47] [Rank 0]   Total Loss: 4.1814
+[2025-09-04 11:43:47] [Rank 0]   Total Loss: 4.1814
+[2025-09-04 11:43:47] [Rank 0]   Total FTA (Unweighted): 0.6744
+[2025-09-04 11:43:47] [Rank 0]   Total FTA (Unweighted): 0.6744
+[2025-09-04 11:43:47] [Rank 0]   Total FTA (Weighted):   0.6744
+[2025-09-04 11:43:47] [Rank 0]   Total FTA (Weighted):   0.6744
+[2025-09-04 11:43:47] [Rank 0]   Group 0 Loss: 4.0174
+[2025-09-04 11:43:47] [Rank 0]   Group 0 Loss: 4.0174
+[2025-09-04 11:43:47] [Rank 0]   Group 1 Loss: 3.6165
+[2025-09-04 11:43:47] [Rank 0]   Group 1 Loss: 3.6165
+[2025-09-04 11:43:47] [Rank 0]   Group 2 Loss: 3.4722
+[2025-09-04 11:43:47] [Rank 0]   Group 2 Loss: 3.4722
+[2025-09-04 11:43:47] [Rank 0]   Group 3 Loss: 3.9216
+[2025-09-04 11:43:47] [Rank 0]   Group 3 Loss: 3.9216
+[2025-09-04 11:43:47] [Rank 0]   Group 4 Loss: 3.9542
+[2025-09-04 11:43:47] [Rank 0]   Group 4 Loss: 3.9542
+[2025-09-04 11:43:47] [Rank 0]   Group 5 Loss: 3.9531
+[2025-09-04 11:43:47] [Rank 0]   Group 5 Loss: 3.9531
+[2025-09-04 11:43:47] [Rank 0]   Group 6 Loss: 3.8600
+[2025-09-04 11:43:47] [Rank 0]   Group 6 Loss: 3.8600
+[2025-09-04 11:43:47] [Rank 0]   Group 7 Loss: 3.9801
+[2025-09-04 11:43:47] [Rank 0]   Group 7 Loss: 3.9801
+[2025-09-04 11:43:47] [Rank 0]   Group 8 Loss: 4.1406
+[2025-09-04 11:43:47] [Rank 0]   Group 8 Loss: 4.1406
+[2025-09-04 11:43:47] [Rank 0]   Group 9 Loss: 4.1840
+[2025-09-04 11:43:47] [Rank 0]   Group 9 Loss: 4.1840
+[2025-09-04 11:43:47] [Rank 0]   Group 10 Loss: 4.3925
+[2025-09-04 11:43:47] [Rank 0]   Group 10 Loss: 4.3925
+[2025-09-04 11:43:47] [Rank 0]   Group 11 Loss: 4.4993
+[2025-09-04 11:43:47] [Rank 0]   Group 11 Loss: 4.4993
+[2025-09-04 11:43:47] [Rank 0]   Group 12 Loss: 4.5561
+[2025-09-04 11:43:47] [Rank 0]   Group 12 Loss: 4.5561
+[2025-09-04 11:43:47] [Rank 0]   Group 13 Loss: 4.7301
+[2025-09-04 11:43:47] [Rank 0]   Group 13 Loss: 4.7301
+[2025-09-04 11:43:47] [Rank 0]   Group 14 Loss: 4.7795
+[2025-09-04 11:43:47] [Rank 0]   Group 14 Loss: 4.7795
+[2025-09-04 11:43:47] [Rank 0]   Group 15 Loss: 4.8458
+[2025-09-04 11:43:47] [Rank 0]   Group 15 Loss: 4.8458
+[2025-09-04 11:43:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:43:47] [Rank 0]   Group 6 FTA: 0.9900
+[2025-09-04 11:43:47] [Rank 0]   Group 6 FTA: 0.9900
+[2025-09-04 11:43:47] [Rank 0]   Group 7 FTA: 0.9000
+[2025-09-04 11:43:47] [Rank 0]   Group 7 FTA: 0.9000
+[2025-09-04 11:43:47] [Rank 0]   Group 8 FTA: 0.7700
+[2025-09-04 11:43:47] [Rank 0]   Group 8 FTA: 0.7700
+[2025-09-04 11:43:47] [Rank 0]   Group 9 FTA: 0.6200
+[2025-09-04 11:43:47] [Rank 0]   Group 9 FTA: 0.6200
+[2025-09-04 11:43:47] [Rank 0]   Group 10 FTA: 0.6300
+[2025-09-04 11:43:47] [Rank 0]   Group 10 FTA: 0.6300
+[2025-09-04 11:43:47] [Rank 0]   Group 11 FTA: 0.3200
+[2025-09-04 11:43:47] [Rank 0]   Group 11 FTA: 0.3200
+[2025-09-04 11:43:47] [Rank 0]   Group 12 FTA: 0.1900
+[2025-09-04 11:43:47] [Rank 0]   Group 12 FTA: 0.1900
+[2025-09-04 11:43:47] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 11:43:47] [Rank 0]   Group 13 FTA: 0.1100
+[2025-09-04 11:43:47] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 11:43:47] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 11:43:47] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 11:43:47] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 11:43:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:43:48] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:43:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:43:48] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:43:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:43:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:43:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:43:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:43:49] [Rank 0] step:1001/10000 train_time:73882ms step_avg:73.81ms
+[2025-09-04 11:43:49] [Rank 0] step:1001/10000 train_time:73882ms step_avg:73.81ms
+[2025-09-04 11:43:49] [Rank 0] step:1021/10000 train_time:74645ms step_avg:73.11ms
+[2025-09-04 11:43:49] [Rank 0] step:1021/10000 train_time:74645ms step_avg:73.11ms
+[2025-09-04 11:43:50] [Rank 0] step:1041/10000 train_time:75392ms step_avg:72.42ms
+[2025-09-04 11:43:50] [Rank 0] step:1041/10000 train_time:75392ms step_avg:72.42ms
+[2025-09-04 11:43:51] [Rank 0] step:1061/10000 train_time:76136ms step_avg:71.76ms
+[2025-09-04 11:43:51] [Rank 0] step:1061/10000 train_time:76136ms step_avg:71.76ms
+[2025-09-04 11:43:52] [Rank 0] step:1081/10000 train_time:76881ms step_avg:71.12ms
+[2025-09-04 11:43:52] [Rank 0] step:1081/10000 train_time:76881ms step_avg:71.12ms
+[2025-09-04 11:43:52] [Rank 0] step:1101/10000 train_time:77626ms step_avg:70.50ms
+[2025-09-04 11:43:52] [Rank 0] step:1101/10000 train_time:77626ms step_avg:70.50ms
+[2025-09-04 11:43:53] [Rank 0] step:1121/10000 train_time:78371ms step_avg:69.91ms
+[2025-09-04 11:43:53] [Rank 0] step:1121/10000 train_time:78371ms step_avg:69.91ms
+[2025-09-04 11:43:54] [Rank 0] step:1141/10000 train_time:79115ms step_avg:69.34ms
+[2025-09-04 11:43:54] [Rank 0] step:1141/10000 train_time:79115ms step_avg:69.34ms
+[2025-09-04 11:43:55] [Rank 0] step:1161/10000 train_time:79859ms step_avg:68.78ms
+[2025-09-04 11:43:55] [Rank 0] step:1161/10000 train_time:79859ms step_avg:68.78ms
+[2025-09-04 11:43:55] [Rank 0] step:1181/10000 train_time:80604ms step_avg:68.25ms
+[2025-09-04 11:43:55] [Rank 0] step:1181/10000 train_time:80604ms step_avg:68.25ms
+[2025-09-04 11:43:56] [Rank 0] step:1201/10000 train_time:81348ms step_avg:67.73ms
+[2025-09-04 11:43:56] [Rank 0] step:1201/10000 train_time:81348ms step_avg:67.73ms
+[2025-09-04 11:43:57] [Rank 0] step:1221/10000 train_time:82092ms step_avg:67.23ms
+[2025-09-04 11:43:57] [Rank 0] step:1221/10000 train_time:82092ms step_avg:67.23ms
+[2025-09-04 11:43:57] [Rank 0] step:1241/10000 train_time:82836ms step_avg:66.75ms
+[2025-09-04 11:43:57] [Rank 0] step:1241/10000 train_time:82836ms step_avg:66.75ms
+[2025-09-04 11:43:58] [Rank 0] step:1261/10000 train_time:83580ms step_avg:66.28ms
+[2025-09-04 11:43:58] [Rank 0] step:1261/10000 train_time:83580ms step_avg:66.28ms
+[2025-09-04 11:43:59] [Rank 0] step:1281/10000 train_time:84324ms step_avg:65.83ms
+[2025-09-04 11:43:59] [Rank 0] step:1281/10000 train_time:84324ms step_avg:65.83ms
+[2025-09-04 11:44:00] [Rank 0] step:1301/10000 train_time:85069ms step_avg:65.39ms
+[2025-09-04 11:44:00] [Rank 0] step:1301/10000 train_time:85069ms step_avg:65.39ms
+[2025-09-04 11:44:00] [Rank 0] step:1321/10000 train_time:85813ms step_avg:64.96ms
+[2025-09-04 11:44:00] [Rank 0] step:1321/10000 train_time:85813ms step_avg:64.96ms
+[2025-09-04 11:44:01] [Rank 0] step:1341/10000 train_time:86557ms step_avg:64.55ms
+[2025-09-04 11:44:01] [Rank 0] step:1341/10000 train_time:86557ms step_avg:64.55ms
+[2025-09-04 11:44:02] [Rank 0] step:1361/10000 train_time:87305ms step_avg:64.15ms
+[2025-09-04 11:44:02] [Rank 0] step:1361/10000 train_time:87305ms step_avg:64.15ms
+[2025-09-04 11:44:03] [Rank 0] step:1381/10000 train_time:88050ms step_avg:63.76ms
+[2025-09-04 11:44:03] [Rank 0] step:1381/10000 train_time:88050ms step_avg:63.76ms
+[2025-09-04 11:44:03] [Rank 0] step:1401/10000 train_time:88794ms step_avg:63.38ms
+[2025-09-04 11:44:03] [Rank 0] step:1401/10000 train_time:88794ms step_avg:63.38ms
+[2025-09-04 11:44:04] [Rank 0] step:1421/10000 train_time:89539ms step_avg:63.01ms
+[2025-09-04 11:44:04] [Rank 0] step:1421/10000 train_time:89539ms step_avg:63.01ms
+[2025-09-04 11:44:05] [Rank 0] step:1441/10000 train_time:90283ms step_avg:62.65ms
+[2025-09-04 11:44:05] [Rank 0] step:1441/10000 train_time:90283ms step_avg:62.65ms
+[2025-09-04 11:44:06] [Rank 0] step:1461/10000 train_time:91036ms step_avg:62.31ms
+[2025-09-04 11:44:06] [Rank 0] step:1461/10000 train_time:91036ms step_avg:62.31ms
+[2025-09-04 11:44:06] [Rank 0] step:1481/10000 train_time:91781ms step_avg:61.97ms
+[2025-09-04 11:44:06] [Rank 0] step:1481/10000 train_time:91781ms step_avg:61.97ms
+[2025-09-04 11:44:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:44:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:44:08] [Rank 0] PRINT: step:1500/10000 train_loss:0.8275 val_loss:0.7814 train_time:92531ms step_avg:61.69ms
+[2025-09-04 11:44:08] [Rank 0] PRINT: step:1500/10000 train_loss:0.8275 val_loss:0.7814 train_time:92531ms step_avg:61.69ms
+[2025-09-04 11:44:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:44:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:44:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:44:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:45:45] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:45:45] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:45:45] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:45:45] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:45:45] [Rank 0]   Total Loss: 4.5425
+[2025-09-04 11:45:45] [Rank 0]   Total Loss: 4.5425
+[2025-09-04 11:45:45] [Rank 0]   Total FTA (Unweighted): 0.7500
+[2025-09-04 11:45:45] [Rank 0]   Total FTA (Unweighted): 0.7500
+[2025-09-04 11:45:45] [Rank 0]   Total FTA (Weighted):   0.7500
+[2025-09-04 11:45:45] [Rank 0]   Total FTA (Weighted):   0.7500
+[2025-09-04 11:45:45] [Rank 0]   Group 0 Loss: 4.5182
+[2025-09-04 11:45:45] [Rank 0]   Group 0 Loss: 4.5182
+[2025-09-04 11:45:45] [Rank 0]   Group 1 Loss: 4.0307
+[2025-09-04 11:45:45] [Rank 0]   Group 1 Loss: 4.0307
+[2025-09-04 11:45:45] [Rank 0]   Group 2 Loss: 3.9916
+[2025-09-04 11:45:45] [Rank 0]   Group 2 Loss: 3.9916
+[2025-09-04 11:45:45] [Rank 0]   Group 3 Loss: 4.3436
+[2025-09-04 11:45:45] [Rank 0]   Group 3 Loss: 4.3436
+[2025-09-04 11:45:45] [Rank 0]   Group 4 Loss: 4.3680
+[2025-09-04 11:45:45] [Rank 0]   Group 4 Loss: 4.3680
+[2025-09-04 11:45:45] [Rank 0]   Group 5 Loss: 4.3957
+[2025-09-04 11:45:45] [Rank 0]   Group 5 Loss: 4.3957
+[2025-09-04 11:45:45] [Rank 0]   Group 6 Loss: 4.2637
+[2025-09-04 11:45:45] [Rank 0]   Group 6 Loss: 4.2637
+[2025-09-04 11:45:45] [Rank 0]   Group 7 Loss: 4.3445
+[2025-09-04 11:45:45] [Rank 0]   Group 7 Loss: 4.3445
+[2025-09-04 11:45:45] [Rank 0]   Group 8 Loss: 4.4741
+[2025-09-04 11:45:45] [Rank 0]   Group 8 Loss: 4.4741
+[2025-09-04 11:45:45] [Rank 0]   Group 9 Loss: 4.4808
+[2025-09-04 11:45:45] [Rank 0]   Group 9 Loss: 4.4808
+[2025-09-04 11:45:45] [Rank 0]   Group 10 Loss: 4.6841
+[2025-09-04 11:45:45] [Rank 0]   Group 10 Loss: 4.6841
+[2025-09-04 11:45:45] [Rank 0]   Group 11 Loss: 4.7724
+[2025-09-04 11:45:45] [Rank 0]   Group 11 Loss: 4.7724
+[2025-09-04 11:45:45] [Rank 0]   Group 12 Loss: 4.8095
+[2025-09-04 11:45:45] [Rank 0]   Group 12 Loss: 4.8095
+[2025-09-04 11:45:45] [Rank 0]   Group 13 Loss: 4.9766
+[2025-09-04 11:45:45] [Rank 0]   Group 13 Loss: 4.9766
+[2025-09-04 11:45:45] [Rank 0]   Group 14 Loss: 5.0605
+[2025-09-04 11:45:45] [Rank 0]   Group 14 Loss: 5.0605
+[2025-09-04 11:45:45] [Rank 0]   Group 15 Loss: 5.1660
+[2025-09-04 11:45:45] [Rank 0]   Group 15 Loss: 5.1660
+[2025-09-04 11:45:46] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:45:46] [Rank 0]   Group 8 FTA: 0.9300
+[2025-09-04 11:45:46] [Rank 0]   Group 8 FTA: 0.9300
+[2025-09-04 11:45:46] [Rank 0]   Group 9 FTA: 0.8000
+[2025-09-04 11:45:46] [Rank 0]   Group 9 FTA: 0.8000
+[2025-09-04 11:45:46] [Rank 0]   Group 10 FTA: 0.8300
+[2025-09-04 11:45:46] [Rank 0]   Group 10 FTA: 0.8300
+[2025-09-04 11:45:46] [Rank 0]   Group 11 FTA: 0.6300
+[2025-09-04 11:45:46] [Rank 0]   Group 11 FTA: 0.6300
+[2025-09-04 11:45:46] [Rank 0]   Group 12 FTA: 0.3800
+[2025-09-04 11:45:46] [Rank 0]   Group 12 FTA: 0.3800
+[2025-09-04 11:45:46] [Rank 0]   Group 13 FTA: 0.1800
+[2025-09-04 11:45:46] [Rank 0]   Group 13 FTA: 0.1800
+[2025-09-04 11:45:46] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 11:45:46] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 11:45:46] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 11:45:46] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 11:45:46] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:45:46] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:45:46] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:45:46] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:45:47] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:45:47] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:45:47] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:45:47] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:45:47] [Rank 0] step:1501/10000 train_time:92546ms step_avg:61.66ms
+[2025-09-04 11:45:47] [Rank 0] step:1501/10000 train_time:92546ms step_avg:61.66ms
+[2025-09-04 11:45:48] [Rank 0] step:1521/10000 train_time:93308ms step_avg:61.35ms
+[2025-09-04 11:45:48] [Rank 0] step:1521/10000 train_time:93308ms step_avg:61.35ms
+[2025-09-04 11:45:48] [Rank 0] step:1541/10000 train_time:94053ms step_avg:61.03ms
+[2025-09-04 11:45:48] [Rank 0] step:1541/10000 train_time:94053ms step_avg:61.03ms
+[2025-09-04 11:45:49] [Rank 0] step:1561/10000 train_time:94799ms step_avg:60.73ms
+[2025-09-04 11:45:49] [Rank 0] step:1561/10000 train_time:94799ms step_avg:60.73ms
+[2025-09-04 11:45:50] [Rank 0] step:1581/10000 train_time:95545ms step_avg:60.43ms
+[2025-09-04 11:45:50] [Rank 0] step:1581/10000 train_time:95545ms step_avg:60.43ms
+[2025-09-04 11:45:51] [Rank 0] step:1601/10000 train_time:96290ms step_avg:60.14ms
+[2025-09-04 11:45:51] [Rank 0] step:1601/10000 train_time:96290ms step_avg:60.14ms
+[2025-09-04 11:45:51] [Rank 0] step:1621/10000 train_time:97036ms step_avg:59.86ms
+[2025-09-04 11:45:51] [Rank 0] step:1621/10000 train_time:97036ms step_avg:59.86ms
+[2025-09-04 11:45:52] [Rank 0] step:1641/10000 train_time:98057ms step_avg:59.75ms
+[2025-09-04 11:45:52] [Rank 0] step:1641/10000 train_time:98057ms step_avg:59.75ms
+[2025-09-04 11:45:53] [Rank 0] step:1661/10000 train_time:98803ms step_avg:59.48ms
+[2025-09-04 11:45:53] [Rank 0] step:1661/10000 train_time:98803ms step_avg:59.48ms
+[2025-09-04 11:45:54] [Rank 0] step:1681/10000 train_time:99548ms step_avg:59.22ms
+[2025-09-04 11:45:54] [Rank 0] step:1681/10000 train_time:99548ms step_avg:59.22ms
+[2025-09-04 11:45:55] [Rank 0] step:1701/10000 train_time:100294ms step_avg:58.96ms
+[2025-09-04 11:45:55] [Rank 0] step:1701/10000 train_time:100294ms step_avg:58.96ms
+[2025-09-04 11:45:55] [Rank 0] step:1721/10000 train_time:101042ms step_avg:58.71ms
+[2025-09-04 11:45:55] [Rank 0] step:1721/10000 train_time:101042ms step_avg:58.71ms
+[2025-09-04 11:45:56] [Rank 0] step:1741/10000 train_time:101787ms step_avg:58.46ms
+[2025-09-04 11:45:56] [Rank 0] step:1741/10000 train_time:101787ms step_avg:58.46ms
+[2025-09-04 11:45:57] [Rank 0] step:1761/10000 train_time:102533ms step_avg:58.22ms
+[2025-09-04 11:45:57] [Rank 0] step:1761/10000 train_time:102533ms step_avg:58.22ms
+[2025-09-04 11:45:58] [Rank 0] step:1781/10000 train_time:103278ms step_avg:57.99ms
+[2025-09-04 11:45:58] [Rank 0] step:1781/10000 train_time:103278ms step_avg:57.99ms
+[2025-09-04 11:45:58] [Rank 0] step:1801/10000 train_time:104024ms step_avg:57.76ms
+[2025-09-04 11:45:58] [Rank 0] step:1801/10000 train_time:104024ms step_avg:57.76ms
+[2025-09-04 11:45:59] [Rank 0] step:1821/10000 train_time:104769ms step_avg:57.53ms
+[2025-09-04 11:45:59] [Rank 0] step:1821/10000 train_time:104769ms step_avg:57.53ms
+[2025-09-04 11:46:00] [Rank 0] step:1841/10000 train_time:105514ms step_avg:57.31ms
+[2025-09-04 11:46:00] [Rank 0] step:1841/10000 train_time:105514ms step_avg:57.31ms
+[2025-09-04 11:46:01] [Rank 0] step:1861/10000 train_time:106259ms step_avg:57.10ms
+[2025-09-04 11:46:01] [Rank 0] step:1861/10000 train_time:106259ms step_avg:57.10ms
+[2025-09-04 11:46:01] [Rank 0] step:1881/10000 train_time:107003ms step_avg:56.89ms
+[2025-09-04 11:46:01] [Rank 0] step:1881/10000 train_time:107003ms step_avg:56.89ms
+[2025-09-04 11:46:02] [Rank 0] step:1901/10000 train_time:107749ms step_avg:56.68ms
+[2025-09-04 11:46:02] [Rank 0] step:1901/10000 train_time:107749ms step_avg:56.68ms
+[2025-09-04 11:46:03] [Rank 0] step:1921/10000 train_time:108494ms step_avg:56.48ms
+[2025-09-04 11:46:03] [Rank 0] step:1921/10000 train_time:108494ms step_avg:56.48ms
+[2025-09-04 11:46:04] [Rank 0] step:1941/10000 train_time:109240ms step_avg:56.28ms
+[2025-09-04 11:46:04] [Rank 0] step:1941/10000 train_time:109240ms step_avg:56.28ms
+[2025-09-04 11:46:04] [Rank 0] step:1961/10000 train_time:109985ms step_avg:56.09ms
+[2025-09-04 11:46:04] [Rank 0] step:1961/10000 train_time:109985ms step_avg:56.09ms
+[2025-09-04 11:46:05] [Rank 0] step:1981/10000 train_time:110730ms step_avg:55.90ms
+[2025-09-04 11:46:05] [Rank 0] step:1981/10000 train_time:110730ms step_avg:55.90ms
+[2025-09-04 11:46:06] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:46:06] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:46:06] [Rank 0] PRINT: step:2000/10000 train_loss:0.7692 val_loss:0.7368 train_time:111480ms step_avg:55.74ms
+[2025-09-04 11:46:06] [Rank 0] PRINT: step:2000/10000 train_loss:0.7692 val_loss:0.7368 train_time:111480ms step_avg:55.74ms
+[2025-09-04 11:46:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:46:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:46:07] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:46:07] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:47:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:47:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:47:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:47:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:47:44] [Rank 0]   Total Loss: 4.5987
+[2025-09-04 11:47:44] [Rank 0]   Total Loss: 4.5987
+[2025-09-04 11:47:44] [Rank 0]   Total FTA (Unweighted): 0.7944
+[2025-09-04 11:47:44] [Rank 0]   Total FTA (Unweighted): 0.7944
+[2025-09-04 11:47:44] [Rank 0]   Total FTA (Weighted):   0.7944
+[2025-09-04 11:47:44] [Rank 0]   Total FTA (Weighted):   0.7944
+[2025-09-04 11:47:44] [Rank 0]   Group 0 Loss: 4.6839
+[2025-09-04 11:47:44] [Rank 0]   Group 0 Loss: 4.6839
+[2025-09-04 11:47:44] [Rank 0]   Group 1 Loss: 4.1141
+[2025-09-04 11:47:44] [Rank 0]   Group 1 Loss: 4.1141
+[2025-09-04 11:47:45] [Rank 0]   Group 2 Loss: 4.0547
+[2025-09-04 11:47:45] [Rank 0]   Group 2 Loss: 4.0547
+[2025-09-04 11:47:45] [Rank 0]   Group 3 Loss: 4.4259
+[2025-09-04 11:47:45] [Rank 0]   Group 3 Loss: 4.4259
+[2025-09-04 11:47:45] [Rank 0]   Group 4 Loss: 4.4417
+[2025-09-04 11:47:45] [Rank 0]   Group 4 Loss: 4.4417
+[2025-09-04 11:47:45] [Rank 0]   Group 5 Loss: 4.4384
+[2025-09-04 11:47:45] [Rank 0]   Group 5 Loss: 4.4384
+[2025-09-04 11:47:45] [Rank 0]   Group 6 Loss: 4.3528
+[2025-09-04 11:47:45] [Rank 0]   Group 6 Loss: 4.3528
+[2025-09-04 11:47:45] [Rank 0]   Group 7 Loss: 4.4732
+[2025-09-04 11:47:45] [Rank 0]   Group 7 Loss: 4.4732
+[2025-09-04 11:47:45] [Rank 0]   Group 8 Loss: 4.5989
+[2025-09-04 11:47:45] [Rank 0]   Group 8 Loss: 4.5989
+[2025-09-04 11:47:45] [Rank 0]   Group 9 Loss: 4.5696
+[2025-09-04 11:47:45] [Rank 0]   Group 9 Loss: 4.5696
+[2025-09-04 11:47:45] [Rank 0]   Group 10 Loss: 4.7084
+[2025-09-04 11:47:45] [Rank 0]   Group 10 Loss: 4.7084
+[2025-09-04 11:47:45] [Rank 0]   Group 11 Loss: 4.7653
+[2025-09-04 11:47:45] [Rank 0]   Group 11 Loss: 4.7653
+[2025-09-04 11:47:45] [Rank 0]   Group 12 Loss: 4.8392
+[2025-09-04 11:47:45] [Rank 0]   Group 12 Loss: 4.8392
+[2025-09-04 11:47:45] [Rank 0]   Group 13 Loss: 4.9978
+[2025-09-04 11:47:45] [Rank 0]   Group 13 Loss: 4.9978
+[2025-09-04 11:47:45] [Rank 0]   Group 14 Loss: 4.9769
+[2025-09-04 11:47:45] [Rank 0]   Group 14 Loss: 4.9769
+[2025-09-04 11:47:45] [Rank 0]   Group 15 Loss: 5.1380
+[2025-09-04 11:47:45] [Rank 0]   Group 15 Loss: 5.1380
+[2025-09-04 11:47:45] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:47:45] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 11:47:45] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 11:47:45] [Rank 0]   Group 9 FTA: 0.9000
+[2025-09-04 11:47:45] [Rank 0]   Group 9 FTA: 0.9000
+[2025-09-04 11:47:45] [Rank 0]   Group 10 FTA: 0.9200
+[2025-09-04 11:47:45] [Rank 0]   Group 10 FTA: 0.9200
+[2025-09-04 11:47:45] [Rank 0]   Group 11 FTA: 0.7800
+[2025-09-04 11:47:45] [Rank 0]   Group 11 FTA: 0.7800
+[2025-09-04 11:47:45] [Rank 0]   Group 12 FTA: 0.5200
+[2025-09-04 11:47:45] [Rank 0]   Group 12 FTA: 0.5200
+[2025-09-04 11:47:45] [Rank 0]   Group 13 FTA: 0.2700
+[2025-09-04 11:47:45] [Rank 0]   Group 13 FTA: 0.2700
+[2025-09-04 11:47:45] [Rank 0]   Group 14 FTA: 0.1800
+[2025-09-04 11:47:45] [Rank 0]   Group 14 FTA: 0.1800
+[2025-09-04 11:47:45] [Rank 0]   Group 15 FTA: 0.1500
+[2025-09-04 11:47:45] [Rank 0]   Group 15 FTA: 0.1500
+[2025-09-04 11:47:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:47:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:47:46] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:47:46] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:47:46] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:47:46] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:47:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:47:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:47:46] [Rank 0] step:2001/10000 train_time:111496ms step_avg:55.72ms
+[2025-09-04 11:47:46] [Rank 0] step:2001/10000 train_time:111496ms step_avg:55.72ms
+[2025-09-04 11:47:47] [Rank 0] step:2021/10000 train_time:112508ms step_avg:55.67ms
+[2025-09-04 11:47:47] [Rank 0] step:2021/10000 train_time:112508ms step_avg:55.67ms
+[2025-09-04 11:47:48] [Rank 0] step:2041/10000 train_time:113254ms step_avg:55.49ms
+[2025-09-04 11:47:48] [Rank 0] step:2041/10000 train_time:113254ms step_avg:55.49ms
+[2025-09-04 11:47:49] [Rank 0] step:2061/10000 train_time:113998ms step_avg:55.31ms
+[2025-09-04 11:47:49] [Rank 0] step:2061/10000 train_time:113998ms step_avg:55.31ms
+[2025-09-04 11:47:49] [Rank 0] step:2081/10000 train_time:114743ms step_avg:55.14ms
+[2025-09-04 11:47:49] [Rank 0] step:2081/10000 train_time:114743ms step_avg:55.14ms
+[2025-09-04 11:47:50] [Rank 0] step:2101/10000 train_time:115487ms step_avg:54.97ms
+[2025-09-04 11:47:50] [Rank 0] step:2101/10000 train_time:115487ms step_avg:54.97ms
+[2025-09-04 11:47:51] [Rank 0] step:2121/10000 train_time:116231ms step_avg:54.80ms
+[2025-09-04 11:47:51] [Rank 0] step:2121/10000 train_time:116231ms step_avg:54.80ms
+[2025-09-04 11:47:52] [Rank 0] step:2141/10000 train_time:116976ms step_avg:54.64ms
+[2025-09-04 11:47:52] [Rank 0] step:2141/10000 train_time:116976ms step_avg:54.64ms
+[2025-09-04 11:47:52] [Rank 0] step:2161/10000 train_time:117721ms step_avg:54.48ms
+[2025-09-04 11:47:52] [Rank 0] step:2161/10000 train_time:117721ms step_avg:54.48ms
+[2025-09-04 11:47:53] [Rank 0] step:2181/10000 train_time:118467ms step_avg:54.32ms
+[2025-09-04 11:47:53] [Rank 0] step:2181/10000 train_time:118467ms step_avg:54.32ms
+[2025-09-04 11:47:54] [Rank 0] step:2201/10000 train_time:119212ms step_avg:54.16ms
+[2025-09-04 11:47:54] [Rank 0] step:2201/10000 train_time:119212ms step_avg:54.16ms
+[2025-09-04 11:47:55] [Rank 0] step:2221/10000 train_time:119958ms step_avg:54.01ms
+[2025-09-04 11:47:55] [Rank 0] step:2221/10000 train_time:119958ms step_avg:54.01ms
+[2025-09-04 11:47:55] [Rank 0] step:2241/10000 train_time:120712ms step_avg:53.87ms
+[2025-09-04 11:47:55] [Rank 0] step:2241/10000 train_time:120712ms step_avg:53.87ms
+[2025-09-04 11:47:56] [Rank 0] step:2261/10000 train_time:121467ms step_avg:53.72ms
+[2025-09-04 11:47:56] [Rank 0] step:2261/10000 train_time:121467ms step_avg:53.72ms
+[2025-09-04 11:47:57] [Rank 0] step:2281/10000 train_time:122223ms step_avg:53.58ms
+[2025-09-04 11:47:57] [Rank 0] step:2281/10000 train_time:122223ms step_avg:53.58ms
+[2025-09-04 11:47:58] [Rank 0] step:2301/10000 train_time:122978ms step_avg:53.45ms
+[2025-09-04 11:47:58] [Rank 0] step:2301/10000 train_time:122978ms step_avg:53.45ms
+[2025-09-04 11:47:58] [Rank 0] step:2321/10000 train_time:123734ms step_avg:53.31ms
+[2025-09-04 11:47:58] [Rank 0] step:2321/10000 train_time:123734ms step_avg:53.31ms
+[2025-09-04 11:47:59] [Rank 0] step:2341/10000 train_time:124489ms step_avg:53.18ms
+[2025-09-04 11:47:59] [Rank 0] step:2341/10000 train_time:124489ms step_avg:53.18ms
+[2025-09-04 11:48:00] [Rank 0] step:2361/10000 train_time:125243ms step_avg:53.05ms
+[2025-09-04 11:48:00] [Rank 0] step:2361/10000 train_time:125243ms step_avg:53.05ms
+[2025-09-04 11:48:01] [Rank 0] step:2381/10000 train_time:125998ms step_avg:52.92ms
+[2025-09-04 11:48:01] [Rank 0] step:2381/10000 train_time:125998ms step_avg:52.92ms
+[2025-09-04 11:48:01] [Rank 0] step:2401/10000 train_time:126754ms step_avg:52.79ms
+[2025-09-04 11:48:01] [Rank 0] step:2401/10000 train_time:126754ms step_avg:52.79ms
+[2025-09-04 11:48:02] [Rank 0] step:2421/10000 train_time:127510ms step_avg:52.67ms
+[2025-09-04 11:48:02] [Rank 0] step:2421/10000 train_time:127510ms step_avg:52.67ms
+[2025-09-04 11:48:03] [Rank 0] step:2441/10000 train_time:128265ms step_avg:52.55ms
+[2025-09-04 11:48:03] [Rank 0] step:2441/10000 train_time:128265ms step_avg:52.55ms
+[2025-09-04 11:48:04] [Rank 0] step:2461/10000 train_time:129020ms step_avg:52.43ms
+[2025-09-04 11:48:04] [Rank 0] step:2461/10000 train_time:129020ms step_avg:52.43ms
+[2025-09-04 11:48:04] [Rank 0] step:2481/10000 train_time:129775ms step_avg:52.31ms
+[2025-09-04 11:48:04] [Rank 0] step:2481/10000 train_time:129775ms step_avg:52.31ms
+[2025-09-04 11:48:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:48:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:48:06] [Rank 0] PRINT: step:2500/10000 train_loss:0.7323 val_loss:0.7057 train_time:130535ms step_avg:52.21ms
+[2025-09-04 11:48:06] [Rank 0] PRINT: step:2500/10000 train_loss:0.7323 val_loss:0.7057 train_time:130535ms step_avg:52.21ms
+[2025-09-04 11:48:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:48:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:48:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:48:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:49:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:49:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:49:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:49:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:49:44] [Rank 0]   Total Loss: 4.6496
+[2025-09-04 11:49:44] [Rank 0]   Total Loss: 4.6496
+[2025-09-04 11:49:44] [Rank 0]   Total FTA (Unweighted): 0.8244
+[2025-09-04 11:49:44] [Rank 0]   Total FTA (Unweighted): 0.8244
+[2025-09-04 11:49:44] [Rank 0]   Total FTA (Weighted):   0.8244
+[2025-09-04 11:49:44] [Rank 0]   Total FTA (Weighted):   0.8244
+[2025-09-04 11:49:44] [Rank 0]   Group 0 Loss: 4.6411
+[2025-09-04 11:49:44] [Rank 0]   Group 0 Loss: 4.6411
+[2025-09-04 11:49:44] [Rank 0]   Group 1 Loss: 4.2704
+[2025-09-04 11:49:44] [Rank 0]   Group 1 Loss: 4.2704
+[2025-09-04 11:49:44] [Rank 0]   Group 2 Loss: 4.1301
+[2025-09-04 11:49:44] [Rank 0]   Group 2 Loss: 4.1301
+[2025-09-04 11:49:44] [Rank 0]   Group 3 Loss: 4.5404
+[2025-09-04 11:49:44] [Rank 0]   Group 3 Loss: 4.5404
+[2025-09-04 11:49:44] [Rank 0]   Group 4 Loss: 4.4868
+[2025-09-04 11:49:44] [Rank 0]   Group 4 Loss: 4.4868
+[2025-09-04 11:49:44] [Rank 0]   Group 5 Loss: 4.5473
+[2025-09-04 11:49:44] [Rank 0]   Group 5 Loss: 4.5473
+[2025-09-04 11:49:44] [Rank 0]   Group 6 Loss: 4.4430
+[2025-09-04 11:49:44] [Rank 0]   Group 6 Loss: 4.4430
+[2025-09-04 11:49:44] [Rank 0]   Group 7 Loss: 4.5416
+[2025-09-04 11:49:44] [Rank 0]   Group 7 Loss: 4.5416
+[2025-09-04 11:49:44] [Rank 0]   Group 8 Loss: 4.6853
+[2025-09-04 11:49:44] [Rank 0]   Group 8 Loss: 4.6853
+[2025-09-04 11:49:44] [Rank 0]   Group 9 Loss: 4.6543
+[2025-09-04 11:49:44] [Rank 0]   Group 9 Loss: 4.6543
+[2025-09-04 11:49:44] [Rank 0]   Group 10 Loss: 4.7502
+[2025-09-04 11:49:44] [Rank 0]   Group 10 Loss: 4.7502
+[2025-09-04 11:49:44] [Rank 0]   Group 11 Loss: 4.8231
+[2025-09-04 11:49:44] [Rank 0]   Group 11 Loss: 4.8231
+[2025-09-04 11:49:44] [Rank 0]   Group 12 Loss: 4.8314
+[2025-09-04 11:49:44] [Rank 0]   Group 12 Loss: 4.8314
+[2025-09-04 11:49:44] [Rank 0]   Group 13 Loss: 4.9766
+[2025-09-04 11:49:44] [Rank 0]   Group 13 Loss: 4.9766
+[2025-09-04 11:49:44] [Rank 0]   Group 14 Loss: 4.9609
+[2025-09-04 11:49:44] [Rank 0]   Group 14 Loss: 4.9609
+[2025-09-04 11:49:44] [Rank 0]   Group 15 Loss: 5.1120
+[2025-09-04 11:49:44] [Rank 0]   Group 15 Loss: 5.1120
+[2025-09-04 11:49:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 6 FTA: 0.9900
+[2025-09-04 11:49:44] [Rank 0]   Group 6 FTA: 0.9900
+[2025-09-04 11:49:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:49:44] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 11:49:44] [Rank 0]   Group 9 FTA: 0.9900
+[2025-09-04 11:49:44] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 11:49:44] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 11:49:44] [Rank 0]   Group 11 FTA: 0.8800
+[2025-09-04 11:49:44] [Rank 0]   Group 11 FTA: 0.8800
+[2025-09-04 11:49:44] [Rank 0]   Group 12 FTA: 0.7400
+[2025-09-04 11:49:44] [Rank 0]   Group 12 FTA: 0.7400
+[2025-09-04 11:49:44] [Rank 0]   Group 13 FTA: 0.3900
+[2025-09-04 11:49:44] [Rank 0]   Group 13 FTA: 0.3900
+[2025-09-04 11:49:44] [Rank 0]   Group 14 FTA: 0.1400
+[2025-09-04 11:49:44] [Rank 0]   Group 14 FTA: 0.1400
+[2025-09-04 11:49:44] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 11:49:44] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 11:49:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:49:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:49:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:49:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:49:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:49:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:49:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:49:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:49:46] [Rank 0] step:2501/10000 train_time:130550ms step_avg:52.20ms
+[2025-09-04 11:49:46] [Rank 0] step:2501/10000 train_time:130550ms step_avg:52.20ms
+[2025-09-04 11:49:46] [Rank 0] step:2521/10000 train_time:131469ms step_avg:52.15ms
+[2025-09-04 11:49:46] [Rank 0] step:2521/10000 train_time:131469ms step_avg:52.15ms
+[2025-09-04 11:49:47] [Rank 0] step:2541/10000 train_time:132281ms step_avg:52.06ms
+[2025-09-04 11:49:47] [Rank 0] step:2541/10000 train_time:132281ms step_avg:52.06ms
+[2025-09-04 11:49:48] [Rank 0] step:2561/10000 train_time:133038ms step_avg:51.95ms
+[2025-09-04 11:49:48] [Rank 0] step:2561/10000 train_time:133038ms step_avg:51.95ms
+[2025-09-04 11:49:49] [Rank 0] step:2581/10000 train_time:133950ms step_avg:51.90ms
+[2025-09-04 11:49:49] [Rank 0] step:2581/10000 train_time:133950ms step_avg:51.90ms
+[2025-09-04 11:49:50] [Rank 0] step:2601/10000 train_time:134813ms step_avg:51.83ms
+[2025-09-04 11:49:50] [Rank 0] step:2601/10000 train_time:134813ms step_avg:51.83ms
+[2025-09-04 11:49:51] [Rank 0] step:2621/10000 train_time:135568ms step_avg:51.72ms
+[2025-09-04 11:49:51] [Rank 0] step:2621/10000 train_time:135568ms step_avg:51.72ms
+[2025-09-04 11:49:51] [Rank 0] step:2641/10000 train_time:136323ms step_avg:51.62ms
+[2025-09-04 11:49:51] [Rank 0] step:2641/10000 train_time:136323ms step_avg:51.62ms
+[2025-09-04 11:49:52] [Rank 0] step:2661/10000 train_time:137079ms step_avg:51.51ms
+[2025-09-04 11:49:52] [Rank 0] step:2661/10000 train_time:137079ms step_avg:51.51ms
+[2025-09-04 11:49:53] [Rank 0] step:2681/10000 train_time:137835ms step_avg:51.41ms
+[2025-09-04 11:49:53] [Rank 0] step:2681/10000 train_time:137835ms step_avg:51.41ms
+[2025-09-04 11:49:54] [Rank 0] step:2701/10000 train_time:138590ms step_avg:51.31ms
+[2025-09-04 11:49:54] [Rank 0] step:2701/10000 train_time:138590ms step_avg:51.31ms
+[2025-09-04 11:49:54] [Rank 0] step:2721/10000 train_time:139345ms step_avg:51.21ms
+[2025-09-04 11:49:54] [Rank 0] step:2721/10000 train_time:139345ms step_avg:51.21ms
+[2025-09-04 11:49:55] [Rank 0] step:2741/10000 train_time:140101ms step_avg:51.11ms
+[2025-09-04 11:49:55] [Rank 0] step:2741/10000 train_time:140101ms step_avg:51.11ms
+[2025-09-04 11:49:56] [Rank 0] step:2761/10000 train_time:140857ms step_avg:51.02ms
+[2025-09-04 11:49:56] [Rank 0] step:2761/10000 train_time:140857ms step_avg:51.02ms
+[2025-09-04 11:49:57] [Rank 0] step:2781/10000 train_time:141612ms step_avg:50.92ms
+[2025-09-04 11:49:57] [Rank 0] step:2781/10000 train_time:141612ms step_avg:50.92ms
+[2025-09-04 11:49:57] [Rank 0] step:2801/10000 train_time:142367ms step_avg:50.83ms
+[2025-09-04 11:49:57] [Rank 0] step:2801/10000 train_time:142367ms step_avg:50.83ms
+[2025-09-04 11:49:58] [Rank 0] step:2821/10000 train_time:143406ms step_avg:50.84ms
+[2025-09-04 11:49:58] [Rank 0] step:2821/10000 train_time:143406ms step_avg:50.84ms
+[2025-09-04 11:49:59] [Rank 0] step:2841/10000 train_time:144162ms step_avg:50.74ms
+[2025-09-04 11:49:59] [Rank 0] step:2841/10000 train_time:144162ms step_avg:50.74ms
+[2025-09-04 11:50:00] [Rank 0] step:2861/10000 train_time:144917ms step_avg:50.65ms
+[2025-09-04 11:50:00] [Rank 0] step:2861/10000 train_time:144917ms step_avg:50.65ms
+[2025-09-04 11:50:01] [Rank 0] step:2881/10000 train_time:145674ms step_avg:50.56ms
+[2025-09-04 11:50:01] [Rank 0] step:2881/10000 train_time:145674ms step_avg:50.56ms
+[2025-09-04 11:50:02] [Rank 0] step:2901/10000 train_time:146478ms step_avg:50.49ms
+[2025-09-04 11:50:02] [Rank 0] step:2901/10000 train_time:146478ms step_avg:50.49ms
+[2025-09-04 11:50:02] [Rank 0] step:2921/10000 train_time:147275ms step_avg:50.42ms
+[2025-09-04 11:50:02] [Rank 0] step:2921/10000 train_time:147275ms step_avg:50.42ms
+[2025-09-04 11:50:03] [Rank 0] step:2941/10000 train_time:148031ms step_avg:50.33ms
+[2025-09-04 11:50:03] [Rank 0] step:2941/10000 train_time:148031ms step_avg:50.33ms
+[2025-09-04 11:50:04] [Rank 0] step:2961/10000 train_time:148786ms step_avg:50.25ms
+[2025-09-04 11:50:04] [Rank 0] step:2961/10000 train_time:148786ms step_avg:50.25ms
+[2025-09-04 11:50:05] [Rank 0] step:2981/10000 train_time:149541ms step_avg:50.16ms
+[2025-09-04 11:50:05] [Rank 0] step:2981/10000 train_time:149541ms step_avg:50.16ms
+[2025-09-04 11:50:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:50:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:50:06] [Rank 0] PRINT: step:3000/10000 train_loss:0.7063 val_loss:0.6851 train_time:150302ms step_avg:50.10ms
+[2025-09-04 11:50:06] [Rank 0] PRINT: step:3000/10000 train_loss:0.7063 val_loss:0.6851 train_time:150302ms step_avg:50.10ms
+[2025-09-04 11:50:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:50:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:50:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:50:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:51:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:51:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:51:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:51:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:51:44] [Rank 0]   Total Loss: 4.7941
+[2025-09-04 11:51:44] [Rank 0]   Total Loss: 4.7941
+[2025-09-04 11:51:44] [Rank 0]   Total FTA (Unweighted): 0.8525
+[2025-09-04 11:51:44] [Rank 0]   Total FTA (Unweighted): 0.8525
+[2025-09-04 11:51:44] [Rank 0]   Total FTA (Weighted):   0.8525
+[2025-09-04 11:51:44] [Rank 0]   Total FTA (Weighted):   0.8525
+[2025-09-04 11:51:44] [Rank 0]   Group 0 Loss: 4.7979
+[2025-09-04 11:51:44] [Rank 0]   Group 0 Loss: 4.7979
+[2025-09-04 11:51:44] [Rank 0]   Group 1 Loss: 4.4207
+[2025-09-04 11:51:44] [Rank 0]   Group 1 Loss: 4.4207
+[2025-09-04 11:51:44] [Rank 0]   Group 2 Loss: 4.2756
+[2025-09-04 11:51:44] [Rank 0]   Group 2 Loss: 4.2756
+[2025-09-04 11:51:44] [Rank 0]   Group 3 Loss: 4.7167
+[2025-09-04 11:51:44] [Rank 0]   Group 3 Loss: 4.7167
+[2025-09-04 11:51:44] [Rank 0]   Group 4 Loss: 4.6872
+[2025-09-04 11:51:44] [Rank 0]   Group 4 Loss: 4.6872
+[2025-09-04 11:51:44] [Rank 0]   Group 5 Loss: 4.6971
+[2025-09-04 11:51:44] [Rank 0]   Group 5 Loss: 4.6971
+[2025-09-04 11:51:44] [Rank 0]   Group 6 Loss: 4.6230
+[2025-09-04 11:51:44] [Rank 0]   Group 6 Loss: 4.6230
+[2025-09-04 11:51:44] [Rank 0]   Group 7 Loss: 4.7190
+[2025-09-04 11:51:44] [Rank 0]   Group 7 Loss: 4.7190
+[2025-09-04 11:51:44] [Rank 0]   Group 8 Loss: 4.8281
+[2025-09-04 11:51:44] [Rank 0]   Group 8 Loss: 4.8281
+[2025-09-04 11:51:44] [Rank 0]   Group 9 Loss: 4.8036
+[2025-09-04 11:51:44] [Rank 0]   Group 9 Loss: 4.8036
+[2025-09-04 11:51:44] [Rank 0]   Group 10 Loss: 4.9445
+[2025-09-04 11:51:44] [Rank 0]   Group 10 Loss: 4.9445
+[2025-09-04 11:51:44] [Rank 0]   Group 11 Loss: 4.9653
+[2025-09-04 11:51:44] [Rank 0]   Group 11 Loss: 4.9653
+[2025-09-04 11:51:44] [Rank 0]   Group 12 Loss: 4.9630
+[2025-09-04 11:51:44] [Rank 0]   Group 12 Loss: 4.9630
+[2025-09-04 11:51:44] [Rank 0]   Group 13 Loss: 5.0791
+[2025-09-04 11:51:44] [Rank 0]   Group 13 Loss: 5.0791
+[2025-09-04 11:51:44] [Rank 0]   Group 14 Loss: 5.0275
+[2025-09-04 11:51:44] [Rank 0]   Group 14 Loss: 5.0275
+[2025-09-04 11:51:44] [Rank 0]   Group 15 Loss: 5.1578
+[2025-09-04 11:51:44] [Rank 0]   Group 15 Loss: 5.1578
+[2025-09-04 11:51:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:51:44] [Rank 0]   Group 10 FTA: 0.9700
+[2025-09-04 11:51:44] [Rank 0]   Group 10 FTA: 0.9700
+[2025-09-04 11:51:44] [Rank 0]   Group 11 FTA: 0.9600
+[2025-09-04 11:51:44] [Rank 0]   Group 11 FTA: 0.9600
+[2025-09-04 11:51:44] [Rank 0]   Group 12 FTA: 0.9000
+[2025-09-04 11:51:44] [Rank 0]   Group 12 FTA: 0.9000
+[2025-09-04 11:51:44] [Rank 0]   Group 13 FTA: 0.4700
+[2025-09-04 11:51:44] [Rank 0]   Group 13 FTA: 0.4700
+[2025-09-04 11:51:44] [Rank 0]   Group 14 FTA: 0.2200
+[2025-09-04 11:51:44] [Rank 0]   Group 14 FTA: 0.2200
+[2025-09-04 11:51:44] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:51:44] [Rank 0]   Group 15 FTA: 0.1200
+[2025-09-04 11:51:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:51:45] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:51:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:51:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:51:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:51:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:51:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:51:46] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:51:46] [Rank 0] step:3001/10000 train_time:150317ms step_avg:50.09ms
+[2025-09-04 11:51:46] [Rank 0] step:3001/10000 train_time:150317ms step_avg:50.09ms
+[2025-09-04 11:51:47] [Rank 0] step:3021/10000 train_time:151087ms step_avg:50.01ms
+[2025-09-04 11:51:47] [Rank 0] step:3021/10000 train_time:151087ms step_avg:50.01ms
+[2025-09-04 11:51:47] [Rank 0] step:3041/10000 train_time:151844ms step_avg:49.93ms
+[2025-09-04 11:51:47] [Rank 0] step:3041/10000 train_time:151844ms step_avg:49.93ms
+[2025-09-04 11:51:48] [Rank 0] step:3061/10000 train_time:152597ms step_avg:49.85ms
+[2025-09-04 11:51:48] [Rank 0] step:3061/10000 train_time:152597ms step_avg:49.85ms
+[2025-09-04 11:51:49] [Rank 0] step:3081/10000 train_time:153352ms step_avg:49.77ms
+[2025-09-04 11:51:49] [Rank 0] step:3081/10000 train_time:153352ms step_avg:49.77ms
+[2025-09-04 11:51:50] [Rank 0] step:3101/10000 train_time:154107ms step_avg:49.70ms
+[2025-09-04 11:51:50] [Rank 0] step:3101/10000 train_time:154107ms step_avg:49.70ms
+[2025-09-04 11:51:50] [Rank 0] step:3121/10000 train_time:154862ms step_avg:49.62ms
+[2025-09-04 11:51:50] [Rank 0] step:3121/10000 train_time:154862ms step_avg:49.62ms
+[2025-09-04 11:51:51] [Rank 0] step:3141/10000 train_time:155617ms step_avg:49.54ms
+[2025-09-04 11:51:51] [Rank 0] step:3141/10000 train_time:155617ms step_avg:49.54ms
+[2025-09-04 11:51:52] [Rank 0] step:3161/10000 train_time:156372ms step_avg:49.47ms
+[2025-09-04 11:51:52] [Rank 0] step:3161/10000 train_time:156372ms step_avg:49.47ms
+[2025-09-04 11:51:53] [Rank 0] step:3181/10000 train_time:157127ms step_avg:49.40ms
+[2025-09-04 11:51:53] [Rank 0] step:3181/10000 train_time:157127ms step_avg:49.40ms
+[2025-09-04 11:51:54] [Rank 0] step:3201/10000 train_time:158095ms step_avg:49.39ms
+[2025-09-04 11:51:54] [Rank 0] step:3201/10000 train_time:158095ms step_avg:49.39ms
+[2025-09-04 11:51:54] [Rank 0] step:3221/10000 train_time:158849ms step_avg:49.32ms
+[2025-09-04 11:51:54] [Rank 0] step:3221/10000 train_time:158849ms step_avg:49.32ms
+[2025-09-04 11:51:55] [Rank 0] step:3241/10000 train_time:159604ms step_avg:49.25ms
+[2025-09-04 11:51:55] [Rank 0] step:3241/10000 train_time:159604ms step_avg:49.25ms
+[2025-09-04 11:51:56] [Rank 0] step:3261/10000 train_time:160592ms step_avg:49.25ms
+[2025-09-04 11:51:56] [Rank 0] step:3261/10000 train_time:160592ms step_avg:49.25ms
+[2025-09-04 11:51:57] [Rank 0] step:3281/10000 train_time:161347ms step_avg:49.18ms
+[2025-09-04 11:51:57] [Rank 0] step:3281/10000 train_time:161347ms step_avg:49.18ms
+[2025-09-04 11:51:58] [Rank 0] step:3301/10000 train_time:162101ms step_avg:49.11ms
+[2025-09-04 11:51:58] [Rank 0] step:3301/10000 train_time:162101ms step_avg:49.11ms
+[2025-09-04 11:51:58] [Rank 0] step:3321/10000 train_time:162855ms step_avg:49.04ms
+[2025-09-04 11:51:58] [Rank 0] step:3321/10000 train_time:162855ms step_avg:49.04ms
+[2025-09-04 11:51:59] [Rank 0] step:3341/10000 train_time:163610ms step_avg:48.97ms
+[2025-09-04 11:51:59] [Rank 0] step:3341/10000 train_time:163610ms step_avg:48.97ms
+[2025-09-04 11:52:00] [Rank 0] step:3361/10000 train_time:164364ms step_avg:48.90ms
+[2025-09-04 11:52:00] [Rank 0] step:3361/10000 train_time:164364ms step_avg:48.90ms
+[2025-09-04 11:52:01] [Rank 0] step:3381/10000 train_time:165118ms step_avg:48.84ms
+[2025-09-04 11:52:01] [Rank 0] step:3381/10000 train_time:165118ms step_avg:48.84ms
+[2025-09-04 11:52:01] [Rank 0] step:3401/10000 train_time:165873ms step_avg:48.77ms
+[2025-09-04 11:52:01] [Rank 0] step:3401/10000 train_time:165873ms step_avg:48.77ms
+[2025-09-04 11:52:02] [Rank 0] step:3421/10000 train_time:166628ms step_avg:48.71ms
+[2025-09-04 11:52:02] [Rank 0] step:3421/10000 train_time:166628ms step_avg:48.71ms
+[2025-09-04 11:52:03] [Rank 0] step:3441/10000 train_time:167383ms step_avg:48.64ms
+[2025-09-04 11:52:03] [Rank 0] step:3441/10000 train_time:167383ms step_avg:48.64ms
+[2025-09-04 11:52:04] [Rank 0] step:3461/10000 train_time:168138ms step_avg:48.58ms
+[2025-09-04 11:52:04] [Rank 0] step:3461/10000 train_time:168138ms step_avg:48.58ms
+[2025-09-04 11:52:04] [Rank 0] step:3481/10000 train_time:168893ms step_avg:48.52ms
+[2025-09-04 11:52:04] [Rank 0] step:3481/10000 train_time:168893ms step_avg:48.52ms
+[2025-09-04 11:52:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:52:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:52:06] [Rank 0] PRINT: step:3500/10000 train_loss:0.6897 val_loss:0.6716 train_time:169654ms step_avg:48.47ms
+[2025-09-04 11:52:06] [Rank 0] PRINT: step:3500/10000 train_loss:0.6897 val_loss:0.6716 train_time:169654ms step_avg:48.47ms
+[2025-09-04 11:52:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:52:06] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:52:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:52:06] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:53:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:53:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:53:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:53:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:53:43] [Rank 0]   Total Loss: 4.9282
+[2025-09-04 11:53:43] [Rank 0]   Total Loss: 4.9282
+[2025-09-04 11:53:43] [Rank 0]   Total FTA (Unweighted): 0.8756
+[2025-09-04 11:53:43] [Rank 0]   Total FTA (Unweighted): 0.8756
+[2025-09-04 11:53:43] [Rank 0]   Total FTA (Weighted):   0.8756
+[2025-09-04 11:53:43] [Rank 0]   Total FTA (Weighted):   0.8756
+[2025-09-04 11:53:43] [Rank 0]   Group 0 Loss: 4.9104
+[2025-09-04 11:53:43] [Rank 0]   Group 0 Loss: 4.9104
+[2025-09-04 11:53:43] [Rank 0]   Group 1 Loss: 4.5990
+[2025-09-04 11:53:43] [Rank 0]   Group 1 Loss: 4.5990
+[2025-09-04 11:53:43] [Rank 0]   Group 2 Loss: 4.3547
+[2025-09-04 11:53:43] [Rank 0]   Group 2 Loss: 4.3547
+[2025-09-04 11:53:43] [Rank 0]   Group 3 Loss: 4.8247
+[2025-09-04 11:53:43] [Rank 0]   Group 3 Loss: 4.8247
+[2025-09-04 11:53:43] [Rank 0]   Group 4 Loss: 4.8254
+[2025-09-04 11:53:43] [Rank 0]   Group 4 Loss: 4.8254
+[2025-09-04 11:53:43] [Rank 0]   Group 5 Loss: 4.8073
+[2025-09-04 11:53:43] [Rank 0]   Group 5 Loss: 4.8073
+[2025-09-04 11:53:43] [Rank 0]   Group 6 Loss: 4.7444
+[2025-09-04 11:53:43] [Rank 0]   Group 6 Loss: 4.7444
+[2025-09-04 11:53:43] [Rank 0]   Group 7 Loss: 4.8367
+[2025-09-04 11:53:43] [Rank 0]   Group 7 Loss: 4.8367
+[2025-09-04 11:53:43] [Rank 0]   Group 8 Loss: 4.9477
+[2025-09-04 11:53:43] [Rank 0]   Group 8 Loss: 4.9477
+[2025-09-04 11:53:43] [Rank 0]   Group 9 Loss: 4.9296
+[2025-09-04 11:53:43] [Rank 0]   Group 9 Loss: 4.9296
+[2025-09-04 11:53:43] [Rank 0]   Group 10 Loss: 5.1331
+[2025-09-04 11:53:43] [Rank 0]   Group 10 Loss: 5.1331
+[2025-09-04 11:53:43] [Rank 0]   Group 11 Loss: 5.2010
+[2025-09-04 11:53:43] [Rank 0]   Group 11 Loss: 5.2010
+[2025-09-04 11:53:43] [Rank 0]   Group 12 Loss: 5.0865
+[2025-09-04 11:53:43] [Rank 0]   Group 12 Loss: 5.0865
+[2025-09-04 11:53:43] [Rank 0]   Group 13 Loss: 5.2426
+[2025-09-04 11:53:43] [Rank 0]   Group 13 Loss: 5.2426
+[2025-09-04 11:53:43] [Rank 0]   Group 14 Loss: 5.1682
+[2025-09-04 11:53:43] [Rank 0]   Group 14 Loss: 5.1682
+[2025-09-04 11:53:43] [Rank 0]   Group 15 Loss: 5.2394
+[2025-09-04 11:53:43] [Rank 0]   Group 15 Loss: 5.2394
+[2025-09-04 11:53:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:53:43] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 11:53:43] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 11:53:43] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 11:53:43] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 11:53:43] [Rank 0]   Group 12 FTA: 0.9600
+[2025-09-04 11:53:43] [Rank 0]   Group 12 FTA: 0.9600
+[2025-09-04 11:53:43] [Rank 0]   Group 13 FTA: 0.6300
+[2025-09-04 11:53:43] [Rank 0]   Group 13 FTA: 0.6300
+[2025-09-04 11:53:43] [Rank 0]   Group 14 FTA: 0.3200
+[2025-09-04 11:53:43] [Rank 0]   Group 14 FTA: 0.3200
+[2025-09-04 11:53:43] [Rank 0]   Group 15 FTA: 0.1300
+[2025-09-04 11:53:43] [Rank 0]   Group 15 FTA: 0.1300
+[2025-09-04 11:53:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:53:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:53:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:53:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:53:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:53:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:53:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:53:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:53:45] [Rank 0] step:3501/10000 train_time:169670ms step_avg:48.46ms
+[2025-09-04 11:53:45] [Rank 0] step:3501/10000 train_time:169670ms step_avg:48.46ms
+[2025-09-04 11:53:45] [Rank 0] step:3521/10000 train_time:170452ms step_avg:48.41ms
+[2025-09-04 11:53:45] [Rank 0] step:3521/10000 train_time:170452ms step_avg:48.41ms
+[2025-09-04 11:53:46] [Rank 0] step:3541/10000 train_time:171208ms step_avg:48.35ms
+[2025-09-04 11:53:46] [Rank 0] step:3541/10000 train_time:171208ms step_avg:48.35ms
+[2025-09-04 11:53:47] [Rank 0] step:3561/10000 train_time:171962ms step_avg:48.29ms
+[2025-09-04 11:53:47] [Rank 0] step:3561/10000 train_time:171962ms step_avg:48.29ms
+[2025-09-04 11:53:48] [Rank 0] step:3581/10000 train_time:172717ms step_avg:48.23ms
+[2025-09-04 11:53:48] [Rank 0] step:3581/10000 train_time:172717ms step_avg:48.23ms
+[2025-09-04 11:53:48] [Rank 0] step:3601/10000 train_time:173472ms step_avg:48.17ms
+[2025-09-04 11:53:48] [Rank 0] step:3601/10000 train_time:173472ms step_avg:48.17ms
+[2025-09-04 11:53:49] [Rank 0] step:3621/10000 train_time:174227ms step_avg:48.12ms
+[2025-09-04 11:53:49] [Rank 0] step:3621/10000 train_time:174227ms step_avg:48.12ms
+[2025-09-04 11:53:51] [Rank 0] step:3641/10000 train_time:175679ms step_avg:48.25ms
+[2025-09-04 11:53:51] [Rank 0] step:3641/10000 train_time:175679ms step_avg:48.25ms
+[2025-09-04 11:53:51] [Rank 0] step:3661/10000 train_time:176435ms step_avg:48.19ms
+[2025-09-04 11:53:51] [Rank 0] step:3661/10000 train_time:176435ms step_avg:48.19ms
+[2025-09-04 11:53:52] [Rank 0] step:3681/10000 train_time:177191ms step_avg:48.14ms
+[2025-09-04 11:53:52] [Rank 0] step:3681/10000 train_time:177191ms step_avg:48.14ms
+[2025-09-04 11:53:53] [Rank 0] step:3701/10000 train_time:177947ms step_avg:48.08ms
+[2025-09-04 11:53:53] [Rank 0] step:3701/10000 train_time:177947ms step_avg:48.08ms
+[2025-09-04 11:53:54] [Rank 0] step:3721/10000 train_time:178703ms step_avg:48.03ms
+[2025-09-04 11:53:54] [Rank 0] step:3721/10000 train_time:178703ms step_avg:48.03ms
+[2025-09-04 11:53:54] [Rank 0] step:3741/10000 train_time:179459ms step_avg:47.97ms
+[2025-09-04 11:53:54] [Rank 0] step:3741/10000 train_time:179459ms step_avg:47.97ms
+[2025-09-04 11:53:55] [Rank 0] step:3761/10000 train_time:180214ms step_avg:47.92ms
+[2025-09-04 11:53:55] [Rank 0] step:3761/10000 train_time:180214ms step_avg:47.92ms
+[2025-09-04 11:53:56] [Rank 0] step:3781/10000 train_time:180970ms step_avg:47.86ms
+[2025-09-04 11:53:56] [Rank 0] step:3781/10000 train_time:180970ms step_avg:47.86ms
+[2025-09-04 11:53:57] [Rank 0] step:3801/10000 train_time:181725ms step_avg:47.81ms
+[2025-09-04 11:53:57] [Rank 0] step:3801/10000 train_time:181725ms step_avg:47.81ms
+[2025-09-04 11:53:57] [Rank 0] step:3821/10000 train_time:182481ms step_avg:47.76ms
+[2025-09-04 11:53:57] [Rank 0] step:3821/10000 train_time:182481ms step_avg:47.76ms
+[2025-09-04 11:53:58] [Rank 0] step:3841/10000 train_time:183236ms step_avg:47.71ms
+[2025-09-04 11:53:58] [Rank 0] step:3841/10000 train_time:183236ms step_avg:47.71ms
+[2025-09-04 11:53:59] [Rank 0] step:3861/10000 train_time:183991ms step_avg:47.65ms
+[2025-09-04 11:53:59] [Rank 0] step:3861/10000 train_time:183991ms step_avg:47.65ms
+[2025-09-04 11:54:00] [Rank 0] step:3881/10000 train_time:185049ms step_avg:47.68ms
+[2025-09-04 11:54:00] [Rank 0] step:3881/10000 train_time:185049ms step_avg:47.68ms
+[2025-09-04 11:54:01] [Rank 0] step:3901/10000 train_time:185804ms step_avg:47.63ms
+[2025-09-04 11:54:01] [Rank 0] step:3901/10000 train_time:185804ms step_avg:47.63ms
+[2025-09-04 11:54:02] [Rank 0] step:3921/10000 train_time:186561ms step_avg:47.58ms
+[2025-09-04 11:54:02] [Rank 0] step:3921/10000 train_time:186561ms step_avg:47.58ms
+[2025-09-04 11:54:03] [Rank 0] step:3941/10000 train_time:187618ms step_avg:47.61ms
+[2025-09-04 11:54:03] [Rank 0] step:3941/10000 train_time:187618ms step_avg:47.61ms
+[2025-09-04 11:54:03] [Rank 0] step:3961/10000 train_time:188373ms step_avg:47.56ms
+[2025-09-04 11:54:03] [Rank 0] step:3961/10000 train_time:188373ms step_avg:47.56ms
+[2025-09-04 11:54:04] [Rank 0] step:3981/10000 train_time:189127ms step_avg:47.51ms
+[2025-09-04 11:54:04] [Rank 0] step:3981/10000 train_time:189127ms step_avg:47.51ms
+[2025-09-04 11:54:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:54:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:54:05] [Rank 0] PRINT: step:4000/10000 train_loss:0.6776 val_loss:0.6604 train_time:189888ms step_avg:47.47ms
+[2025-09-04 11:54:05] [Rank 0] PRINT: step:4000/10000 train_loss:0.6776 val_loss:0.6604 train_time:189888ms step_avg:47.47ms
+[2025-09-04 11:54:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:54:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:54:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:54:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:55:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:55:43] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:55:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:55:43] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:55:43] [Rank 0]   Total Loss: 4.9519
+[2025-09-04 11:55:43] [Rank 0]   Total Loss: 4.9519
+[2025-09-04 11:55:43] [Rank 0]   Total FTA (Unweighted): 0.8969
+[2025-09-04 11:55:43] [Rank 0]   Total FTA (Unweighted): 0.8969
+[2025-09-04 11:55:43] [Rank 0]   Total FTA (Weighted):   0.8969
+[2025-09-04 11:55:43] [Rank 0]   Total FTA (Weighted):   0.8969
+[2025-09-04 11:55:43] [Rank 0]   Group 0 Loss: 4.9616
+[2025-09-04 11:55:43] [Rank 0]   Group 0 Loss: 4.9616
+[2025-09-04 11:55:43] [Rank 0]   Group 1 Loss: 4.5627
+[2025-09-04 11:55:43] [Rank 0]   Group 1 Loss: 4.5627
+[2025-09-04 11:55:43] [Rank 0]   Group 2 Loss: 4.4310
+[2025-09-04 11:55:43] [Rank 0]   Group 2 Loss: 4.4310
+[2025-09-04 11:55:43] [Rank 0]   Group 3 Loss: 4.9132
+[2025-09-04 11:55:43] [Rank 0]   Group 3 Loss: 4.9132
+[2025-09-04 11:55:43] [Rank 0]   Group 4 Loss: 4.8345
+[2025-09-04 11:55:43] [Rank 0]   Group 4 Loss: 4.8345
+[2025-09-04 11:55:43] [Rank 0]   Group 5 Loss: 4.8672
+[2025-09-04 11:55:43] [Rank 0]   Group 5 Loss: 4.8672
+[2025-09-04 11:55:43] [Rank 0]   Group 6 Loss: 4.7458
+[2025-09-04 11:55:43] [Rank 0]   Group 6 Loss: 4.7458
+[2025-09-04 11:55:43] [Rank 0]   Group 7 Loss: 4.9154
+[2025-09-04 11:55:43] [Rank 0]   Group 7 Loss: 4.9154
+[2025-09-04 11:55:43] [Rank 0]   Group 8 Loss: 5.0011
+[2025-09-04 11:55:43] [Rank 0]   Group 8 Loss: 5.0011
+[2025-09-04 11:55:43] [Rank 0]   Group 9 Loss: 4.9809
+[2025-09-04 11:55:43] [Rank 0]   Group 9 Loss: 4.9809
+[2025-09-04 11:55:43] [Rank 0]   Group 10 Loss: 5.1277
+[2025-09-04 11:55:43] [Rank 0]   Group 10 Loss: 5.1277
+[2025-09-04 11:55:43] [Rank 0]   Group 11 Loss: 5.1631
+[2025-09-04 11:55:43] [Rank 0]   Group 11 Loss: 5.1631
+[2025-09-04 11:55:43] [Rank 0]   Group 12 Loss: 5.1047
+[2025-09-04 11:55:43] [Rank 0]   Group 12 Loss: 5.1047
+[2025-09-04 11:55:43] [Rank 0]   Group 13 Loss: 5.2132
+[2025-09-04 11:55:43] [Rank 0]   Group 13 Loss: 5.2132
+[2025-09-04 11:55:43] [Rank 0]   Group 14 Loss: 5.1544
+[2025-09-04 11:55:43] [Rank 0]   Group 14 Loss: 5.1544
+[2025-09-04 11:55:43] [Rank 0]   Group 15 Loss: 5.2534
+[2025-09-04 11:55:43] [Rank 0]   Group 15 Loss: 5.2534
+[2025-09-04 11:55:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 11:55:43] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 11:55:43] [Rank 0]   Group 9 FTA: 0.9800
+[2025-09-04 11:55:43] [Rank 0]   Group 9 FTA: 0.9800
+[2025-09-04 11:55:43] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 11:55:43] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 11:55:43] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 11:55:43] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 11:55:43] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:55:43] [Rank 0]   Group 13 FTA: 0.8600
+[2025-09-04 11:55:43] [Rank 0]   Group 13 FTA: 0.8600
+[2025-09-04 11:55:43] [Rank 0]   Group 14 FTA: 0.3400
+[2025-09-04 11:55:43] [Rank 0]   Group 14 FTA: 0.3400
+[2025-09-04 11:55:43] [Rank 0]   Group 15 FTA: 0.2100
+[2025-09-04 11:55:43] [Rank 0]   Group 15 FTA: 0.2100
+[2025-09-04 11:55:43] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:55:43] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:55:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:55:44] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:55:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:55:44] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:55:44] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:55:44] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:55:44] [Rank 0] step:4001/10000 train_time:189903ms step_avg:47.46ms
+[2025-09-04 11:55:44] [Rank 0] step:4001/10000 train_time:189903ms step_avg:47.46ms
+[2025-09-04 11:55:45] [Rank 0] step:4021/10000 train_time:190949ms step_avg:47.49ms
+[2025-09-04 11:55:45] [Rank 0] step:4021/10000 train_time:190949ms step_avg:47.49ms
+[2025-09-04 11:55:46] [Rank 0] step:4041/10000 train_time:191703ms step_avg:47.44ms
+[2025-09-04 11:55:46] [Rank 0] step:4041/10000 train_time:191703ms step_avg:47.44ms
+[2025-09-04 11:55:47] [Rank 0] step:4061/10000 train_time:192458ms step_avg:47.39ms
+[2025-09-04 11:55:47] [Rank 0] step:4061/10000 train_time:192458ms step_avg:47.39ms
+[2025-09-04 11:55:48] [Rank 0] step:4081/10000 train_time:193212ms step_avg:47.34ms
+[2025-09-04 11:55:48] [Rank 0] step:4081/10000 train_time:193212ms step_avg:47.34ms
+[2025-09-04 11:55:48] [Rank 0] step:4101/10000 train_time:193966ms step_avg:47.30ms
+[2025-09-04 11:55:48] [Rank 0] step:4101/10000 train_time:193966ms step_avg:47.30ms
+[2025-09-04 11:55:49] [Rank 0] step:4121/10000 train_time:194721ms step_avg:47.25ms
+[2025-09-04 11:55:49] [Rank 0] step:4121/10000 train_time:194721ms step_avg:47.25ms
+[2025-09-04 11:55:50] [Rank 0] step:4141/10000 train_time:195476ms step_avg:47.20ms
+[2025-09-04 11:55:50] [Rank 0] step:4141/10000 train_time:195476ms step_avg:47.20ms
+[2025-09-04 11:55:51] [Rank 0] step:4161/10000 train_time:196231ms step_avg:47.16ms
+[2025-09-04 11:55:51] [Rank 0] step:4161/10000 train_time:196231ms step_avg:47.16ms
+[2025-09-04 11:55:51] [Rank 0] step:4181/10000 train_time:196985ms step_avg:47.11ms
+[2025-09-04 11:55:51] [Rank 0] step:4181/10000 train_time:196985ms step_avg:47.11ms
+[2025-09-04 11:55:52] [Rank 0] step:4201/10000 train_time:197740ms step_avg:47.07ms
+[2025-09-04 11:55:52] [Rank 0] step:4201/10000 train_time:197740ms step_avg:47.07ms
+[2025-09-04 11:55:53] [Rank 0] step:4221/10000 train_time:198494ms step_avg:47.03ms
+[2025-09-04 11:55:53] [Rank 0] step:4221/10000 train_time:198494ms step_avg:47.03ms
+[2025-09-04 11:55:54] [Rank 0] step:4241/10000 train_time:199248ms step_avg:46.98ms
+[2025-09-04 11:55:54] [Rank 0] step:4241/10000 train_time:199248ms step_avg:46.98ms
+[2025-09-04 11:55:54] [Rank 0] step:4261/10000 train_time:200003ms step_avg:46.94ms
+[2025-09-04 11:55:54] [Rank 0] step:4261/10000 train_time:200003ms step_avg:46.94ms
+[2025-09-04 11:55:55] [Rank 0] step:4281/10000 train_time:200758ms step_avg:46.90ms
+[2025-09-04 11:55:55] [Rank 0] step:4281/10000 train_time:200758ms step_avg:46.90ms
+[2025-09-04 11:55:56] [Rank 0] step:4301/10000 train_time:201513ms step_avg:46.85ms
+[2025-09-04 11:55:56] [Rank 0] step:4301/10000 train_time:201513ms step_avg:46.85ms
+[2025-09-04 11:55:57] [Rank 0] step:4321/10000 train_time:202268ms step_avg:46.81ms
+[2025-09-04 11:55:57] [Rank 0] step:4321/10000 train_time:202268ms step_avg:46.81ms
+[2025-09-04 11:55:57] [Rank 0] step:4341/10000 train_time:203023ms step_avg:46.77ms
+[2025-09-04 11:55:57] [Rank 0] step:4341/10000 train_time:203023ms step_avg:46.77ms
+[2025-09-04 11:55:58] [Rank 0] step:4361/10000 train_time:203777ms step_avg:46.73ms
+[2025-09-04 11:55:58] [Rank 0] step:4361/10000 train_time:203777ms step_avg:46.73ms
+[2025-09-04 11:55:59] [Rank 0] step:4381/10000 train_time:204532ms step_avg:46.69ms
+[2025-09-04 11:55:59] [Rank 0] step:4381/10000 train_time:204532ms step_avg:46.69ms
+[2025-09-04 11:56:00] [Rank 0] step:4401/10000 train_time:205286ms step_avg:46.65ms
+[2025-09-04 11:56:00] [Rank 0] step:4401/10000 train_time:205286ms step_avg:46.65ms
+[2025-09-04 11:56:00] [Rank 0] step:4421/10000 train_time:206041ms step_avg:46.61ms
+[2025-09-04 11:56:00] [Rank 0] step:4421/10000 train_time:206041ms step_avg:46.61ms
+[2025-09-04 11:56:01] [Rank 0] step:4441/10000 train_time:206795ms step_avg:46.57ms
+[2025-09-04 11:56:01] [Rank 0] step:4441/10000 train_time:206795ms step_avg:46.57ms
+[2025-09-04 11:56:02] [Rank 0] step:4461/10000 train_time:207550ms step_avg:46.53ms
+[2025-09-04 11:56:02] [Rank 0] step:4461/10000 train_time:207550ms step_avg:46.53ms
+[2025-09-04 11:56:03] [Rank 0] step:4481/10000 train_time:208305ms step_avg:46.49ms
+[2025-09-04 11:56:03] [Rank 0] step:4481/10000 train_time:208305ms step_avg:46.49ms
+[2025-09-04 11:56:03] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:56:03] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:56:04] [Rank 0] PRINT: step:4500/10000 train_loss:0.6675 val_loss:0.6511 train_time:209065ms step_avg:46.46ms
+[2025-09-04 11:56:04] [Rank 0] PRINT: step:4500/10000 train_loss:0.6675 val_loss:0.6511 train_time:209065ms step_avg:46.46ms
+[2025-09-04 11:56:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:56:04] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:56:04] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:56:04] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:57:41] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:57:41] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:57:41] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:57:41] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:57:41] [Rank 0]   Total Loss: 4.9464
+[2025-09-04 11:57:41] [Rank 0]   Total Loss: 4.9464
+[2025-09-04 11:57:41] [Rank 0]   Total FTA (Unweighted): 0.9106
+[2025-09-04 11:57:41] [Rank 0]   Total FTA (Unweighted): 0.9106
+[2025-09-04 11:57:41] [Rank 0]   Total FTA (Weighted):   0.9106
+[2025-09-04 11:57:41] [Rank 0]   Total FTA (Weighted):   0.9106
+[2025-09-04 11:57:41] [Rank 0]   Group 0 Loss: 4.9430
+[2025-09-04 11:57:41] [Rank 0]   Group 0 Loss: 4.9430
+[2025-09-04 11:57:41] [Rank 0]   Group 1 Loss: 4.4836
+[2025-09-04 11:57:41] [Rank 0]   Group 1 Loss: 4.4836
+[2025-09-04 11:57:41] [Rank 0]   Group 2 Loss: 4.4086
+[2025-09-04 11:57:41] [Rank 0]   Group 2 Loss: 4.4086
+[2025-09-04 11:57:41] [Rank 0]   Group 3 Loss: 4.8200
+[2025-09-04 11:57:41] [Rank 0]   Group 3 Loss: 4.8200
+[2025-09-04 11:57:41] [Rank 0]   Group 4 Loss: 4.8660
+[2025-09-04 11:57:41] [Rank 0]   Group 4 Loss: 4.8660
+[2025-09-04 11:57:41] [Rank 0]   Group 5 Loss: 4.8975
+[2025-09-04 11:57:41] [Rank 0]   Group 5 Loss: 4.8975
+[2025-09-04 11:57:41] [Rank 0]   Group 6 Loss: 4.7839
+[2025-09-04 11:57:41] [Rank 0]   Group 6 Loss: 4.7839
+[2025-09-04 11:57:41] [Rank 0]   Group 7 Loss: 4.8861
+[2025-09-04 11:57:41] [Rank 0]   Group 7 Loss: 4.8861
+[2025-09-04 11:57:41] [Rank 0]   Group 8 Loss: 5.0613
+[2025-09-04 11:57:41] [Rank 0]   Group 8 Loss: 5.0613
+[2025-09-04 11:57:41] [Rank 0]   Group 9 Loss: 4.9886
+[2025-09-04 11:57:41] [Rank 0]   Group 9 Loss: 4.9886
+[2025-09-04 11:57:41] [Rank 0]   Group 10 Loss: 5.1241
+[2025-09-04 11:57:41] [Rank 0]   Group 10 Loss: 5.1241
+[2025-09-04 11:57:41] [Rank 0]   Group 11 Loss: 5.1864
+[2025-09-04 11:57:41] [Rank 0]   Group 11 Loss: 5.1864
+[2025-09-04 11:57:41] [Rank 0]   Group 12 Loss: 5.1006
+[2025-09-04 11:57:41] [Rank 0]   Group 12 Loss: 5.1006
+[2025-09-04 11:57:41] [Rank 0]   Group 13 Loss: 5.2008
+[2025-09-04 11:57:41] [Rank 0]   Group 13 Loss: 5.2008
+[2025-09-04 11:57:41] [Rank 0]   Group 14 Loss: 5.1568
+[2025-09-04 11:57:41] [Rank 0]   Group 14 Loss: 5.1568
+[2025-09-04 11:57:41] [Rank 0]   Group 15 Loss: 5.2357
+[2025-09-04 11:57:41] [Rank 0]   Group 15 Loss: 5.2357
+[2025-09-04 11:57:41] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:57:41] [Rank 0]   Group 13 FTA: 0.9100
+[2025-09-04 11:57:41] [Rank 0]   Group 13 FTA: 0.9100
+[2025-09-04 11:57:41] [Rank 0]   Group 14 FTA: 0.4300
+[2025-09-04 11:57:41] [Rank 0]   Group 14 FTA: 0.4300
+[2025-09-04 11:57:41] [Rank 0]   Group 15 FTA: 0.2300
+[2025-09-04 11:57:41] [Rank 0]   Group 15 FTA: 0.2300
+[2025-09-04 11:57:41] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:57:41] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:57:42] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:57:42] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:57:42] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:57:42] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:57:42] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:57:42] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:57:42] [Rank 0] step:4501/10000 train_time:209080ms step_avg:46.45ms
+[2025-09-04 11:57:42] [Rank 0] step:4501/10000 train_time:209080ms step_avg:46.45ms
+[2025-09-04 11:57:43] [Rank 0] step:4521/10000 train_time:209840ms step_avg:46.41ms
+[2025-09-04 11:57:43] [Rank 0] step:4521/10000 train_time:209840ms step_avg:46.41ms
+[2025-09-04 11:57:44] [Rank 0] step:4541/10000 train_time:210596ms step_avg:46.38ms
+[2025-09-04 11:57:44] [Rank 0] step:4541/10000 train_time:210596ms step_avg:46.38ms
+[2025-09-04 11:57:44] [Rank 0] step:4561/10000 train_time:211352ms step_avg:46.34ms
+[2025-09-04 11:57:44] [Rank 0] step:4561/10000 train_time:211352ms step_avg:46.34ms
+[2025-09-04 11:57:45] [Rank 0] step:4581/10000 train_time:212107ms step_avg:46.30ms
+[2025-09-04 11:57:45] [Rank 0] step:4581/10000 train_time:212107ms step_avg:46.30ms
+[2025-09-04 11:57:46] [Rank 0] step:4601/10000 train_time:212863ms step_avg:46.26ms
+[2025-09-04 11:57:46] [Rank 0] step:4601/10000 train_time:212863ms step_avg:46.26ms
+[2025-09-04 11:57:47] [Rank 0] step:4621/10000 train_time:213618ms step_avg:46.23ms
+[2025-09-04 11:57:47] [Rank 0] step:4621/10000 train_time:213618ms step_avg:46.23ms
+[2025-09-04 11:57:48] [Rank 0] step:4641/10000 train_time:214374ms step_avg:46.19ms
+[2025-09-04 11:57:48] [Rank 0] step:4641/10000 train_time:214374ms step_avg:46.19ms
+[2025-09-04 11:57:48] [Rank 0] step:4661/10000 train_time:215130ms step_avg:46.16ms
+[2025-09-04 11:57:48] [Rank 0] step:4661/10000 train_time:215130ms step_avg:46.16ms
+[2025-09-04 11:57:49] [Rank 0] step:4681/10000 train_time:215886ms step_avg:46.12ms
+[2025-09-04 11:57:49] [Rank 0] step:4681/10000 train_time:215886ms step_avg:46.12ms
+[2025-09-04 11:57:50] [Rank 0] step:4701/10000 train_time:216641ms step_avg:46.08ms
+[2025-09-04 11:57:50] [Rank 0] step:4701/10000 train_time:216641ms step_avg:46.08ms
+[2025-09-04 11:57:51] [Rank 0] step:4721/10000 train_time:217396ms step_avg:46.05ms
+[2025-09-04 11:57:51] [Rank 0] step:4721/10000 train_time:217396ms step_avg:46.05ms
+[2025-09-04 11:57:51] [Rank 0] step:4741/10000 train_time:218152ms step_avg:46.01ms
+[2025-09-04 11:57:51] [Rank 0] step:4741/10000 train_time:218152ms step_avg:46.01ms
+[2025-09-04 11:57:52] [Rank 0] step:4761/10000 train_time:218907ms step_avg:45.98ms
+[2025-09-04 11:57:52] [Rank 0] step:4761/10000 train_time:218907ms step_avg:45.98ms
+[2025-09-04 11:57:53] [Rank 0] step:4781/10000 train_time:219663ms step_avg:45.94ms
+[2025-09-04 11:57:53] [Rank 0] step:4781/10000 train_time:219663ms step_avg:45.94ms
+[2025-09-04 11:57:54] [Rank 0] step:4801/10000 train_time:220418ms step_avg:45.91ms
+[2025-09-04 11:57:54] [Rank 0] step:4801/10000 train_time:220418ms step_avg:45.91ms
+[2025-09-04 11:57:54] [Rank 0] step:4821/10000 train_time:221173ms step_avg:45.88ms
+[2025-09-04 11:57:54] [Rank 0] step:4821/10000 train_time:221173ms step_avg:45.88ms
+[2025-09-04 11:57:55] [Rank 0] step:4841/10000 train_time:222235ms step_avg:45.91ms
+[2025-09-04 11:57:55] [Rank 0] step:4841/10000 train_time:222235ms step_avg:45.91ms
+[2025-09-04 11:57:56] [Rank 0] step:4861/10000 train_time:222990ms step_avg:45.87ms
+[2025-09-04 11:57:56] [Rank 0] step:4861/10000 train_time:222990ms step_avg:45.87ms
+[2025-09-04 11:57:57] [Rank 0] step:4881/10000 train_time:223745ms step_avg:45.84ms
+[2025-09-04 11:57:57] [Rank 0] step:4881/10000 train_time:223745ms step_avg:45.84ms
+[2025-09-04 11:57:58] [Rank 0] step:4901/10000 train_time:224500ms step_avg:45.81ms
+[2025-09-04 11:57:58] [Rank 0] step:4901/10000 train_time:224500ms step_avg:45.81ms
+[2025-09-04 11:57:58] [Rank 0] step:4921/10000 train_time:225255ms step_avg:45.77ms
+[2025-09-04 11:57:58] [Rank 0] step:4921/10000 train_time:225255ms step_avg:45.77ms
+[2025-09-04 11:57:59] [Rank 0] step:4941/10000 train_time:226010ms step_avg:45.74ms
+[2025-09-04 11:57:59] [Rank 0] step:4941/10000 train_time:226010ms step_avg:45.74ms
+[2025-09-04 11:58:00] [Rank 0] step:4961/10000 train_time:226765ms step_avg:45.71ms
+[2025-09-04 11:58:00] [Rank 0] step:4961/10000 train_time:226765ms step_avg:45.71ms
+[2025-09-04 11:58:01] [Rank 0] step:4981/10000 train_time:227521ms step_avg:45.68ms
+[2025-09-04 11:58:01] [Rank 0] step:4981/10000 train_time:227521ms step_avg:45.68ms
+[2025-09-04 11:58:01] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:58:01] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 11:58:02] [Rank 0] PRINT: step:5000/10000 train_loss:0.6587 val_loss:0.6435 train_time:228281ms step_avg:45.66ms
+[2025-09-04 11:58:02] [Rank 0] PRINT: step:5000/10000 train_loss:0.6587 val_loss:0.6435 train_time:228281ms step_avg:45.66ms
+[2025-09-04 11:58:02] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:58:02] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 11:58:02] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:58:02] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 11:59:39] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:59:39] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 11:59:39] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:59:39] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 11:59:39] [Rank 0]   Total Loss: 5.0902
+[2025-09-04 11:59:39] [Rank 0]   Total Loss: 5.0902
+[2025-09-04 11:59:39] [Rank 0]   Total FTA (Unweighted): 0.9263
+[2025-09-04 11:59:39] [Rank 0]   Total FTA (Unweighted): 0.9263
+[2025-09-04 11:59:39] [Rank 0]   Total FTA (Weighted):   0.9263
+[2025-09-04 11:59:39] [Rank 0]   Total FTA (Weighted):   0.9263
+[2025-09-04 11:59:39] [Rank 0]   Group 0 Loss: 4.9925
+[2025-09-04 11:59:39] [Rank 0]   Group 0 Loss: 4.9925
+[2025-09-04 11:59:39] [Rank 0]   Group 1 Loss: 4.6083
+[2025-09-04 11:59:39] [Rank 0]   Group 1 Loss: 4.6083
+[2025-09-04 11:59:39] [Rank 0]   Group 2 Loss: 4.5883
+[2025-09-04 11:59:39] [Rank 0]   Group 2 Loss: 4.5883
+[2025-09-04 11:59:39] [Rank 0]   Group 3 Loss: 4.9427
+[2025-09-04 11:59:39] [Rank 0]   Group 3 Loss: 4.9427
+[2025-09-04 11:59:39] [Rank 0]   Group 4 Loss: 4.9477
+[2025-09-04 11:59:39] [Rank 0]   Group 4 Loss: 4.9477
+[2025-09-04 11:59:39] [Rank 0]   Group 5 Loss: 5.0506
+[2025-09-04 11:59:39] [Rank 0]   Group 5 Loss: 5.0506
+[2025-09-04 11:59:39] [Rank 0]   Group 6 Loss: 4.9171
+[2025-09-04 11:59:39] [Rank 0]   Group 6 Loss: 4.9171
+[2025-09-04 11:59:39] [Rank 0]   Group 7 Loss: 5.0150
+[2025-09-04 11:59:39] [Rank 0]   Group 7 Loss: 5.0150
+[2025-09-04 11:59:39] [Rank 0]   Group 8 Loss: 5.1901
+[2025-09-04 11:59:39] [Rank 0]   Group 8 Loss: 5.1901
+[2025-09-04 11:59:39] [Rank 0]   Group 9 Loss: 5.1820
+[2025-09-04 11:59:39] [Rank 0]   Group 9 Loss: 5.1820
+[2025-09-04 11:59:39] [Rank 0]   Group 10 Loss: 5.3363
+[2025-09-04 11:59:39] [Rank 0]   Group 10 Loss: 5.3363
+[2025-09-04 11:59:39] [Rank 0]   Group 11 Loss: 5.3615
+[2025-09-04 11:59:39] [Rank 0]   Group 11 Loss: 5.3615
+[2025-09-04 11:59:39] [Rank 0]   Group 12 Loss: 5.2410
+[2025-09-04 11:59:39] [Rank 0]   Group 12 Loss: 5.2410
+[2025-09-04 11:59:39] [Rank 0]   Group 13 Loss: 5.3990
+[2025-09-04 11:59:39] [Rank 0]   Group 13 Loss: 5.3990
+[2025-09-04 11:59:39] [Rank 0]   Group 14 Loss: 5.3549
+[2025-09-04 11:59:39] [Rank 0]   Group 14 Loss: 5.3549
+[2025-09-04 11:59:39] [Rank 0]   Group 15 Loss: 5.3164
+[2025-09-04 11:59:39] [Rank 0]   Group 15 Loss: 5.3164
+[2025-09-04 11:59:39] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:59:39] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 11:59:40] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:59:40] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 11:59:40] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:59:40] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 11:59:40] [Rank 0]   Group 13 FTA: 0.9500
+[2025-09-04 11:59:40] [Rank 0]   Group 13 FTA: 0.9500
+[2025-09-04 11:59:40] [Rank 0]   Group 14 FTA: 0.5700
+[2025-09-04 11:59:40] [Rank 0]   Group 14 FTA: 0.5700
+[2025-09-04 11:59:40] [Rank 0]   Group 15 FTA: 0.3000
+[2025-09-04 11:59:40] [Rank 0]   Group 15 FTA: 0.3000
+[2025-09-04 11:59:40] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:59:40] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 11:59:40] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:59:40] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 11:59:41] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:59:41] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 11:59:41] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:59:41] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 11:59:41] [Rank 0] step:5001/10000 train_time:228297ms step_avg:45.65ms
+[2025-09-04 11:59:41] [Rank 0] step:5001/10000 train_time:228297ms step_avg:45.65ms
+[2025-09-04 11:59:42] [Rank 0] step:5021/10000 train_time:229071ms step_avg:45.62ms
+[2025-09-04 11:59:42] [Rank 0] step:5021/10000 train_time:229071ms step_avg:45.62ms
+[2025-09-04 11:59:43] [Rank 0] step:5041/10000 train_time:229827ms step_avg:45.59ms
+[2025-09-04 11:59:43] [Rank 0] step:5041/10000 train_time:229827ms step_avg:45.59ms
+[2025-09-04 11:59:43] [Rank 0] step:5061/10000 train_time:230582ms step_avg:45.56ms
+[2025-09-04 11:59:43] [Rank 0] step:5061/10000 train_time:230582ms step_avg:45.56ms
+[2025-09-04 11:59:44] [Rank 0] step:5081/10000 train_time:231338ms step_avg:45.53ms
+[2025-09-04 11:59:44] [Rank 0] step:5081/10000 train_time:231338ms step_avg:45.53ms
+[2025-09-04 11:59:45] [Rank 0] step:5101/10000 train_time:232093ms step_avg:45.50ms
+[2025-09-04 11:59:45] [Rank 0] step:5101/10000 train_time:232093ms step_avg:45.50ms
+[2025-09-04 11:59:46] [Rank 0] step:5121/10000 train_time:232848ms step_avg:45.47ms
+[2025-09-04 11:59:46] [Rank 0] step:5121/10000 train_time:232848ms step_avg:45.47ms
+[2025-09-04 11:59:46] [Rank 0] step:5141/10000 train_time:233603ms step_avg:45.44ms
+[2025-09-04 11:59:46] [Rank 0] step:5141/10000 train_time:233603ms step_avg:45.44ms
+[2025-09-04 11:59:47] [Rank 0] step:5161/10000 train_time:234359ms step_avg:45.41ms
+[2025-09-04 11:59:47] [Rank 0] step:5161/10000 train_time:234359ms step_avg:45.41ms
+[2025-09-04 11:59:48] [Rank 0] step:5181/10000 train_time:235114ms step_avg:45.38ms
+[2025-09-04 11:59:48] [Rank 0] step:5181/10000 train_time:235114ms step_avg:45.38ms
+[2025-09-04 11:59:49] [Rank 0] step:5201/10000 train_time:235870ms step_avg:45.35ms
+[2025-09-04 11:59:49] [Rank 0] step:5201/10000 train_time:235870ms step_avg:45.35ms
+[2025-09-04 11:59:49] [Rank 0] step:5221/10000 train_time:236624ms step_avg:45.32ms
+[2025-09-04 11:59:49] [Rank 0] step:5221/10000 train_time:236624ms step_avg:45.32ms
+[2025-09-04 11:59:50] [Rank 0] step:5241/10000 train_time:237380ms step_avg:45.29ms
+[2025-09-04 11:59:50] [Rank 0] step:5241/10000 train_time:237380ms step_avg:45.29ms
+[2025-09-04 11:59:51] [Rank 0] step:5261/10000 train_time:238136ms step_avg:45.26ms
+[2025-09-04 11:59:51] [Rank 0] step:5261/10000 train_time:238136ms step_avg:45.26ms
+[2025-09-04 11:59:52] [Rank 0] step:5281/10000 train_time:238891ms step_avg:45.24ms
+[2025-09-04 11:59:52] [Rank 0] step:5281/10000 train_time:238891ms step_avg:45.24ms
+[2025-09-04 11:59:52] [Rank 0] step:5301/10000 train_time:239647ms step_avg:45.21ms
+[2025-09-04 11:59:52] [Rank 0] step:5301/10000 train_time:239647ms step_avg:45.21ms
+[2025-09-04 11:59:53] [Rank 0] step:5321/10000 train_time:240401ms step_avg:45.18ms
+[2025-09-04 11:59:53] [Rank 0] step:5321/10000 train_time:240401ms step_avg:45.18ms
+[2025-09-04 11:59:54] [Rank 0] step:5341/10000 train_time:241157ms step_avg:45.15ms
+[2025-09-04 11:59:54] [Rank 0] step:5341/10000 train_time:241157ms step_avg:45.15ms
+[2025-09-04 11:59:55] [Rank 0] step:5361/10000 train_time:241912ms step_avg:45.12ms
+[2025-09-04 11:59:55] [Rank 0] step:5361/10000 train_time:241912ms step_avg:45.12ms
+[2025-09-04 11:59:55] [Rank 0] step:5381/10000 train_time:242667ms step_avg:45.10ms
+[2025-09-04 11:59:55] [Rank 0] step:5381/10000 train_time:242667ms step_avg:45.10ms
+[2025-09-04 11:59:56] [Rank 0] step:5401/10000 train_time:243424ms step_avg:45.07ms
+[2025-09-04 11:59:56] [Rank 0] step:5401/10000 train_time:243424ms step_avg:45.07ms
+[2025-09-04 11:59:57] [Rank 0] step:5421/10000 train_time:244178ms step_avg:45.04ms
+[2025-09-04 11:59:57] [Rank 0] step:5421/10000 train_time:244178ms step_avg:45.04ms
+[2025-09-04 11:59:58] [Rank 0] step:5441/10000 train_time:244934ms step_avg:45.02ms
+[2025-09-04 11:59:58] [Rank 0] step:5441/10000 train_time:244934ms step_avg:45.02ms
+[2025-09-04 11:59:58] [Rank 0] step:5461/10000 train_time:245689ms step_avg:44.99ms
+[2025-09-04 11:59:58] [Rank 0] step:5461/10000 train_time:245689ms step_avg:44.99ms
+[2025-09-04 11:59:59] [Rank 0] step:5481/10000 train_time:246445ms step_avg:44.96ms
+[2025-09-04 11:59:59] [Rank 0] step:5481/10000 train_time:246445ms step_avg:44.96ms
+[2025-09-04 12:00:00] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:00:00] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:00:00] [Rank 0] PRINT: step:5500/10000 train_loss:0.6511 val_loss:0.6365 train_time:247205ms step_avg:44.95ms
+[2025-09-04 12:00:00] [Rank 0] PRINT: step:5500/10000 train_loss:0.6511 val_loss:0.6365 train_time:247205ms step_avg:44.95ms
+[2025-09-04 12:00:00] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:00:00] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:00:01] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:00:01] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:01:38] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:01:38] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:01:38] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:01:38] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:01:38] [Rank 0]   Total Loss: 5.1819
+[2025-09-04 12:01:38] [Rank 0]   Total Loss: 5.1819
+[2025-09-04 12:01:38] [Rank 0]   Total FTA (Unweighted): 0.9313
+[2025-09-04 12:01:38] [Rank 0]   Total FTA (Unweighted): 0.9313
+[2025-09-04 12:01:38] [Rank 0]   Total FTA (Weighted):   0.9313
+[2025-09-04 12:01:38] [Rank 0]   Total FTA (Weighted):   0.9313
+[2025-09-04 12:01:38] [Rank 0]   Group 0 Loss: 5.0885
+[2025-09-04 12:01:38] [Rank 0]   Group 0 Loss: 5.0885
+[2025-09-04 12:01:38] [Rank 0]   Group 1 Loss: 4.8430
+[2025-09-04 12:01:38] [Rank 0]   Group 1 Loss: 4.8430
+[2025-09-04 12:01:38] [Rank 0]   Group 2 Loss: 4.6186
+[2025-09-04 12:01:38] [Rank 0]   Group 2 Loss: 4.6186
+[2025-09-04 12:01:38] [Rank 0]   Group 3 Loss: 5.0732
+[2025-09-04 12:01:38] [Rank 0]   Group 3 Loss: 5.0732
+[2025-09-04 12:01:38] [Rank 0]   Group 4 Loss: 5.0361
+[2025-09-04 12:01:38] [Rank 0]   Group 4 Loss: 5.0361
+[2025-09-04 12:01:38] [Rank 0]   Group 5 Loss: 5.1298
+[2025-09-04 12:01:38] [Rank 0]   Group 5 Loss: 5.1298
+[2025-09-04 12:01:38] [Rank 0]   Group 6 Loss: 5.0356
+[2025-09-04 12:01:38] [Rank 0]   Group 6 Loss: 5.0356
+[2025-09-04 12:01:38] [Rank 0]   Group 7 Loss: 5.1129
+[2025-09-04 12:01:38] [Rank 0]   Group 7 Loss: 5.1129
+[2025-09-04 12:01:38] [Rank 0]   Group 8 Loss: 5.2682
+[2025-09-04 12:01:38] [Rank 0]   Group 8 Loss: 5.2682
+[2025-09-04 12:01:38] [Rank 0]   Group 9 Loss: 5.2232
+[2025-09-04 12:01:38] [Rank 0]   Group 9 Loss: 5.2232
+[2025-09-04 12:01:38] [Rank 0]   Group 10 Loss: 5.4296
+[2025-09-04 12:01:38] [Rank 0]   Group 10 Loss: 5.4296
+[2025-09-04 12:01:38] [Rank 0]   Group 11 Loss: 5.4541
+[2025-09-04 12:01:38] [Rank 0]   Group 11 Loss: 5.4541
+[2025-09-04 12:01:38] [Rank 0]   Group 12 Loss: 5.3126
+[2025-09-04 12:01:38] [Rank 0]   Group 12 Loss: 5.3126
+[2025-09-04 12:01:38] [Rank 0]   Group 13 Loss: 5.4671
+[2025-09-04 12:01:38] [Rank 0]   Group 13 Loss: 5.4671
+[2025-09-04 12:01:38] [Rank 0]   Group 14 Loss: 5.4205
+[2025-09-04 12:01:38] [Rank 0]   Group 14 Loss: 5.4205
+[2025-09-04 12:01:38] [Rank 0]   Group 15 Loss: 5.3976
+[2025-09-04 12:01:38] [Rank 0]   Group 15 Loss: 5.3976
+[2025-09-04 12:01:38] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:01:38] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 12:01:38] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 12:01:38] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:01:38] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:01:38] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:01:38] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:01:38] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 12:01:38] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 12:01:38] [Rank 0]   Group 14 FTA: 0.6500
+[2025-09-04 12:01:38] [Rank 0]   Group 14 FTA: 0.6500
+[2025-09-04 12:01:38] [Rank 0]   Group 15 FTA: 0.3000
+[2025-09-04 12:01:38] [Rank 0]   Group 15 FTA: 0.3000
+[2025-09-04 12:01:38] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:01:38] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:01:39] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:01:39] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:01:39] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:01:39] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:01:39] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:01:39] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:01:39] [Rank 0] step:5501/10000 train_time:247220ms step_avg:44.94ms
+[2025-09-04 12:01:39] [Rank 0] step:5501/10000 train_time:247220ms step_avg:44.94ms
+[2025-09-04 12:01:40] [Rank 0] step:5521/10000 train_time:247998ms step_avg:44.92ms
+[2025-09-04 12:01:40] [Rank 0] step:5521/10000 train_time:247998ms step_avg:44.92ms
+[2025-09-04 12:01:41] [Rank 0] step:5541/10000 train_time:248753ms step_avg:44.89ms
+[2025-09-04 12:01:41] [Rank 0] step:5541/10000 train_time:248753ms step_avg:44.89ms
+[2025-09-04 12:01:42] [Rank 0] step:5561/10000 train_time:249508ms step_avg:44.87ms
+[2025-09-04 12:01:42] [Rank 0] step:5561/10000 train_time:249508ms step_avg:44.87ms
+[2025-09-04 12:01:42] [Rank 0] step:5581/10000 train_time:250262ms step_avg:44.84ms
+[2025-09-04 12:01:42] [Rank 0] step:5581/10000 train_time:250262ms step_avg:44.84ms
+[2025-09-04 12:01:43] [Rank 0] step:5601/10000 train_time:251018ms step_avg:44.82ms
+[2025-09-04 12:01:43] [Rank 0] step:5601/10000 train_time:251018ms step_avg:44.82ms
+[2025-09-04 12:01:44] [Rank 0] step:5621/10000 train_time:251773ms step_avg:44.79ms
+[2025-09-04 12:01:44] [Rank 0] step:5621/10000 train_time:251773ms step_avg:44.79ms
+[2025-09-04 12:01:45] [Rank 0] step:5641/10000 train_time:252798ms step_avg:44.81ms
+[2025-09-04 12:01:45] [Rank 0] step:5641/10000 train_time:252798ms step_avg:44.81ms
+[2025-09-04 12:01:46] [Rank 0] step:5661/10000 train_time:253553ms step_avg:44.79ms
+[2025-09-04 12:01:46] [Rank 0] step:5661/10000 train_time:253553ms step_avg:44.79ms
+[2025-09-04 12:01:46] [Rank 0] step:5681/10000 train_time:254308ms step_avg:44.76ms
+[2025-09-04 12:01:46] [Rank 0] step:5681/10000 train_time:254308ms step_avg:44.76ms
+[2025-09-04 12:01:47] [Rank 0] step:5701/10000 train_time:255062ms step_avg:44.74ms
+[2025-09-04 12:01:47] [Rank 0] step:5701/10000 train_time:255062ms step_avg:44.74ms
+[2025-09-04 12:01:48] [Rank 0] step:5721/10000 train_time:255817ms step_avg:44.72ms
+[2025-09-04 12:01:48] [Rank 0] step:5721/10000 train_time:255817ms step_avg:44.72ms
+[2025-09-04 12:01:49] [Rank 0] step:5741/10000 train_time:256571ms step_avg:44.69ms
+[2025-09-04 12:01:49] [Rank 0] step:5741/10000 train_time:256571ms step_avg:44.69ms
+[2025-09-04 12:01:49] [Rank 0] step:5761/10000 train_time:257326ms step_avg:44.67ms
+[2025-09-04 12:01:49] [Rank 0] step:5761/10000 train_time:257326ms step_avg:44.67ms
+[2025-09-04 12:01:50] [Rank 0] step:5781/10000 train_time:258081ms step_avg:44.64ms
+[2025-09-04 12:01:50] [Rank 0] step:5781/10000 train_time:258081ms step_avg:44.64ms
+[2025-09-04 12:01:51] [Rank 0] step:5801/10000 train_time:258835ms step_avg:44.62ms
+[2025-09-04 12:01:51] [Rank 0] step:5801/10000 train_time:258835ms step_avg:44.62ms
+[2025-09-04 12:01:52] [Rank 0] step:5821/10000 train_time:259589ms step_avg:44.60ms
+[2025-09-04 12:01:52] [Rank 0] step:5821/10000 train_time:259589ms step_avg:44.60ms
+[2025-09-04 12:01:52] [Rank 0] step:5841/10000 train_time:260344ms step_avg:44.57ms
+[2025-09-04 12:01:52] [Rank 0] step:5841/10000 train_time:260344ms step_avg:44.57ms
+[2025-09-04 12:01:53] [Rank 0] step:5861/10000 train_time:261100ms step_avg:44.55ms
+[2025-09-04 12:01:53] [Rank 0] step:5861/10000 train_time:261100ms step_avg:44.55ms
+[2025-09-04 12:01:54] [Rank 0] step:5881/10000 train_time:261854ms step_avg:44.53ms
+[2025-09-04 12:01:54] [Rank 0] step:5881/10000 train_time:261854ms step_avg:44.53ms
+[2025-09-04 12:01:55] [Rank 0] step:5901/10000 train_time:262609ms step_avg:44.50ms
+[2025-09-04 12:01:55] [Rank 0] step:5901/10000 train_time:262609ms step_avg:44.50ms
+[2025-09-04 12:01:55] [Rank 0] step:5921/10000 train_time:263364ms step_avg:44.48ms
+[2025-09-04 12:01:55] [Rank 0] step:5921/10000 train_time:263364ms step_avg:44.48ms
+[2025-09-04 12:01:56] [Rank 0] step:5941/10000 train_time:264119ms step_avg:44.46ms
+[2025-09-04 12:01:56] [Rank 0] step:5941/10000 train_time:264119ms step_avg:44.46ms
+[2025-09-04 12:01:57] [Rank 0] step:5961/10000 train_time:264874ms step_avg:44.43ms
+[2025-09-04 12:01:57] [Rank 0] step:5961/10000 train_time:264874ms step_avg:44.43ms
+[2025-09-04 12:01:58] [Rank 0] step:5981/10000 train_time:265629ms step_avg:44.41ms
+[2025-09-04 12:01:58] [Rank 0] step:5981/10000 train_time:265629ms step_avg:44.41ms
+[2025-09-04 12:01:58] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:01:58] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:01:59] [Rank 0] PRINT: step:6000/10000 train_loss:0.6441 val_loss:0.6302 train_time:266389ms step_avg:44.40ms
+[2025-09-04 12:01:59] [Rank 0] PRINT: step:6000/10000 train_loss:0.6441 val_loss:0.6302 train_time:266389ms step_avg:44.40ms
+[2025-09-04 12:01:59] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:01:59] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:01:59] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:01:59] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:03:36] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:03:36] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:03:36] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:03:36] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:03:36] [Rank 0]   Total Loss: 5.1891
+[2025-09-04 12:03:36] [Rank 0]   Total Loss: 5.1891
+[2025-09-04 12:03:36] [Rank 0]   Total FTA (Unweighted): 0.9519
+[2025-09-04 12:03:36] [Rank 0]   Total FTA (Unweighted): 0.9519
+[2025-09-04 12:03:36] [Rank 0]   Total FTA (Weighted):   0.9519
+[2025-09-04 12:03:36] [Rank 0]   Total FTA (Weighted):   0.9519
+[2025-09-04 12:03:36] [Rank 0]   Group 0 Loss: 5.0351
+[2025-09-04 12:03:36] [Rank 0]   Group 0 Loss: 5.0351
+[2025-09-04 12:03:36] [Rank 0]   Group 1 Loss: 4.8225
+[2025-09-04 12:03:36] [Rank 0]   Group 1 Loss: 4.8225
+[2025-09-04 12:03:36] [Rank 0]   Group 2 Loss: 4.6466
+[2025-09-04 12:03:36] [Rank 0]   Group 2 Loss: 4.6466
+[2025-09-04 12:03:36] [Rank 0]   Group 3 Loss: 5.0751
+[2025-09-04 12:03:36] [Rank 0]   Group 3 Loss: 5.0751
+[2025-09-04 12:03:36] [Rank 0]   Group 4 Loss: 5.0598
+[2025-09-04 12:03:36] [Rank 0]   Group 4 Loss: 5.0598
+[2025-09-04 12:03:36] [Rank 0]   Group 5 Loss: 5.1575
+[2025-09-04 12:03:36] [Rank 0]   Group 5 Loss: 5.1575
+[2025-09-04 12:03:36] [Rank 0]   Group 6 Loss: 5.0456
+[2025-09-04 12:03:36] [Rank 0]   Group 6 Loss: 5.0456
+[2025-09-04 12:03:36] [Rank 0]   Group 7 Loss: 5.1209
+[2025-09-04 12:03:36] [Rank 0]   Group 7 Loss: 5.1209
+[2025-09-04 12:03:36] [Rank 0]   Group 8 Loss: 5.2492
+[2025-09-04 12:03:36] [Rank 0]   Group 8 Loss: 5.2492
+[2025-09-04 12:03:36] [Rank 0]   Group 9 Loss: 5.2404
+[2025-09-04 12:03:36] [Rank 0]   Group 9 Loss: 5.2404
+[2025-09-04 12:03:36] [Rank 0]   Group 10 Loss: 5.4495
+[2025-09-04 12:03:36] [Rank 0]   Group 10 Loss: 5.4495
+[2025-09-04 12:03:36] [Rank 0]   Group 11 Loss: 5.4606
+[2025-09-04 12:03:36] [Rank 0]   Group 11 Loss: 5.4606
+[2025-09-04 12:03:36] [Rank 0]   Group 12 Loss: 5.2998
+[2025-09-04 12:03:36] [Rank 0]   Group 12 Loss: 5.2998
+[2025-09-04 12:03:36] [Rank 0]   Group 13 Loss: 5.5293
+[2025-09-04 12:03:36] [Rank 0]   Group 13 Loss: 5.5293
+[2025-09-04 12:03:36] [Rank 0]   Group 14 Loss: 5.4374
+[2025-09-04 12:03:36] [Rank 0]   Group 14 Loss: 5.4374
+[2025-09-04 12:03:36] [Rank 0]   Group 15 Loss: 5.3966
+[2025-09-04 12:03:36] [Rank 0]   Group 15 Loss: 5.3966
+[2025-09-04 12:03:36] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:03:36] [Rank 0]   Group 13 FTA: 0.9700
+[2025-09-04 12:03:36] [Rank 0]   Group 13 FTA: 0.9700
+[2025-09-04 12:03:36] [Rank 0]   Group 14 FTA: 0.7800
+[2025-09-04 12:03:36] [Rank 0]   Group 14 FTA: 0.7800
+[2025-09-04 12:03:36] [Rank 0]   Group 15 FTA: 0.4800
+[2025-09-04 12:03:36] [Rank 0]   Group 15 FTA: 0.4800
+[2025-09-04 12:03:36] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:03:36] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:03:37] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:03:37] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:03:37] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:03:37] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:03:37] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:03:37] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:03:37] [Rank 0] step:6001/10000 train_time:266404ms step_avg:44.39ms
+[2025-09-04 12:03:37] [Rank 0] step:6001/10000 train_time:266404ms step_avg:44.39ms
+[2025-09-04 12:03:38] [Rank 0] step:6021/10000 train_time:267429ms step_avg:44.42ms
+[2025-09-04 12:03:38] [Rank 0] step:6021/10000 train_time:267429ms step_avg:44.42ms
+[2025-09-04 12:03:39] [Rank 0] step:6041/10000 train_time:268185ms step_avg:44.39ms
+[2025-09-04 12:03:39] [Rank 0] step:6041/10000 train_time:268185ms step_avg:44.39ms
+[2025-09-04 12:03:40] [Rank 0] step:6061/10000 train_time:268940ms step_avg:44.37ms
+[2025-09-04 12:03:40] [Rank 0] step:6061/10000 train_time:268940ms step_avg:44.37ms
+[2025-09-04 12:03:41] [Rank 0] step:6081/10000 train_time:269696ms step_avg:44.35ms
+[2025-09-04 12:03:41] [Rank 0] step:6081/10000 train_time:269696ms step_avg:44.35ms
+[2025-09-04 12:03:41] [Rank 0] step:6101/10000 train_time:270452ms step_avg:44.33ms
+[2025-09-04 12:03:41] [Rank 0] step:6101/10000 train_time:270452ms step_avg:44.33ms
+[2025-09-04 12:03:42] [Rank 0] step:6121/10000 train_time:271208ms step_avg:44.31ms
+[2025-09-04 12:03:42] [Rank 0] step:6121/10000 train_time:271208ms step_avg:44.31ms
+[2025-09-04 12:03:43] [Rank 0] step:6141/10000 train_time:271963ms step_avg:44.29ms
+[2025-09-04 12:03:43] [Rank 0] step:6141/10000 train_time:271963ms step_avg:44.29ms
+[2025-09-04 12:03:44] [Rank 0] step:6161/10000 train_time:272719ms step_avg:44.27ms
+[2025-09-04 12:03:44] [Rank 0] step:6161/10000 train_time:272719ms step_avg:44.27ms
+[2025-09-04 12:03:44] [Rank 0] step:6181/10000 train_time:273475ms step_avg:44.24ms
+[2025-09-04 12:03:44] [Rank 0] step:6181/10000 train_time:273475ms step_avg:44.24ms
+[2025-09-04 12:03:45] [Rank 0] step:6201/10000 train_time:274231ms step_avg:44.22ms
+[2025-09-04 12:03:45] [Rank 0] step:6201/10000 train_time:274231ms step_avg:44.22ms
+[2025-09-04 12:03:46] [Rank 0] step:6221/10000 train_time:274987ms step_avg:44.20ms
+[2025-09-04 12:03:46] [Rank 0] step:6221/10000 train_time:274987ms step_avg:44.20ms
+[2025-09-04 12:03:47] [Rank 0] step:6241/10000 train_time:275742ms step_avg:44.18ms
+[2025-09-04 12:03:47] [Rank 0] step:6241/10000 train_time:275742ms step_avg:44.18ms
+[2025-09-04 12:03:48] [Rank 0] step:6261/10000 train_time:276499ms step_avg:44.16ms
+[2025-09-04 12:03:48] [Rank 0] step:6261/10000 train_time:276499ms step_avg:44.16ms
+[2025-09-04 12:03:48] [Rank 0] step:6281/10000 train_time:277253ms step_avg:44.14ms
+[2025-09-04 12:03:48] [Rank 0] step:6281/10000 train_time:277253ms step_avg:44.14ms
+[2025-09-04 12:03:49] [Rank 0] step:6301/10000 train_time:278008ms step_avg:44.12ms
+[2025-09-04 12:03:49] [Rank 0] step:6301/10000 train_time:278008ms step_avg:44.12ms
+[2025-09-04 12:03:50] [Rank 0] step:6321/10000 train_time:278765ms step_avg:44.10ms
+[2025-09-04 12:03:50] [Rank 0] step:6321/10000 train_time:278765ms step_avg:44.10ms
+[2025-09-04 12:03:51] [Rank 0] step:6341/10000 train_time:279519ms step_avg:44.08ms
+[2025-09-04 12:03:51] [Rank 0] step:6341/10000 train_time:279519ms step_avg:44.08ms
+[2025-09-04 12:03:51] [Rank 0] step:6361/10000 train_time:280279ms step_avg:44.06ms
+[2025-09-04 12:03:51] [Rank 0] step:6361/10000 train_time:280279ms step_avg:44.06ms
+[2025-09-04 12:03:52] [Rank 0] step:6381/10000 train_time:281034ms step_avg:44.04ms
+[2025-09-04 12:03:52] [Rank 0] step:6381/10000 train_time:281034ms step_avg:44.04ms
+[2025-09-04 12:03:53] [Rank 0] step:6401/10000 train_time:281790ms step_avg:44.02ms
+[2025-09-04 12:03:53] [Rank 0] step:6401/10000 train_time:281790ms step_avg:44.02ms
+[2025-09-04 12:03:54] [Rank 0] step:6421/10000 train_time:282544ms step_avg:44.00ms
+[2025-09-04 12:03:54] [Rank 0] step:6421/10000 train_time:282544ms step_avg:44.00ms
+[2025-09-04 12:03:54] [Rank 0] step:6441/10000 train_time:283298ms step_avg:43.98ms
+[2025-09-04 12:03:54] [Rank 0] step:6441/10000 train_time:283298ms step_avg:43.98ms
+[2025-09-04 12:03:55] [Rank 0] step:6461/10000 train_time:284053ms step_avg:43.96ms
+[2025-09-04 12:03:55] [Rank 0] step:6461/10000 train_time:284053ms step_avg:43.96ms
+[2025-09-04 12:03:56] [Rank 0] step:6481/10000 train_time:284809ms step_avg:43.95ms
+[2025-09-04 12:03:56] [Rank 0] step:6481/10000 train_time:284809ms step_avg:43.95ms
+[2025-09-04 12:03:57] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:03:57] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:03:57] [Rank 0] PRINT: step:6500/10000 train_loss:0.6378 val_loss:0.6248 train_time:285569ms step_avg:43.93ms
+[2025-09-04 12:03:57] [Rank 0] PRINT: step:6500/10000 train_loss:0.6378 val_loss:0.6248 train_time:285569ms step_avg:43.93ms
+[2025-09-04 12:03:57] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:03:57] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:03:57] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:03:57] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:05:35] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:05:35] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:05:35] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:05:35] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:05:35] [Rank 0]   Total Loss: 5.1695
+[2025-09-04 12:05:35] [Rank 0]   Total Loss: 5.1695
+[2025-09-04 12:05:35] [Rank 0]   Total FTA (Unweighted): 0.9550
+[2025-09-04 12:05:35] [Rank 0]   Total FTA (Unweighted): 0.9550
+[2025-09-04 12:05:35] [Rank 0]   Total FTA (Weighted):   0.9550
+[2025-09-04 12:05:35] [Rank 0]   Total FTA (Weighted):   0.9550
+[2025-09-04 12:05:35] [Rank 0]   Group 0 Loss: 5.0519
+[2025-09-04 12:05:35] [Rank 0]   Group 0 Loss: 5.0519
+[2025-09-04 12:05:35] [Rank 0]   Group 1 Loss: 4.8452
+[2025-09-04 12:05:35] [Rank 0]   Group 1 Loss: 4.8452
+[2025-09-04 12:05:35] [Rank 0]   Group 2 Loss: 4.6247
+[2025-09-04 12:05:35] [Rank 0]   Group 2 Loss: 4.6247
+[2025-09-04 12:05:35] [Rank 0]   Group 3 Loss: 5.0130
+[2025-09-04 12:05:35] [Rank 0]   Group 3 Loss: 5.0130
+[2025-09-04 12:05:35] [Rank 0]   Group 4 Loss: 5.0873
+[2025-09-04 12:05:35] [Rank 0]   Group 4 Loss: 5.0873
+[2025-09-04 12:05:35] [Rank 0]   Group 5 Loss: 5.0950
+[2025-09-04 12:05:35] [Rank 0]   Group 5 Loss: 5.0950
+[2025-09-04 12:05:35] [Rank 0]   Group 6 Loss: 5.0256
+[2025-09-04 12:05:35] [Rank 0]   Group 6 Loss: 5.0256
+[2025-09-04 12:05:35] [Rank 0]   Group 7 Loss: 5.1040
+[2025-09-04 12:05:35] [Rank 0]   Group 7 Loss: 5.1040
+[2025-09-04 12:05:35] [Rank 0]   Group 8 Loss: 5.2523
+[2025-09-04 12:05:35] [Rank 0]   Group 8 Loss: 5.2523
+[2025-09-04 12:05:35] [Rank 0]   Group 9 Loss: 5.1996
+[2025-09-04 12:05:35] [Rank 0]   Group 9 Loss: 5.1996
+[2025-09-04 12:05:35] [Rank 0]   Group 10 Loss: 5.3868
+[2025-09-04 12:05:35] [Rank 0]   Group 10 Loss: 5.3868
+[2025-09-04 12:05:35] [Rank 0]   Group 11 Loss: 5.4098
+[2025-09-04 12:05:35] [Rank 0]   Group 11 Loss: 5.4098
+[2025-09-04 12:05:35] [Rank 0]   Group 12 Loss: 5.3271
+[2025-09-04 12:05:35] [Rank 0]   Group 12 Loss: 5.3271
+[2025-09-04 12:05:35] [Rank 0]   Group 13 Loss: 5.4928
+[2025-09-04 12:05:35] [Rank 0]   Group 13 Loss: 5.4928
+[2025-09-04 12:05:35] [Rank 0]   Group 14 Loss: 5.4054
+[2025-09-04 12:05:35] [Rank 0]   Group 14 Loss: 5.4054
+[2025-09-04 12:05:35] [Rank 0]   Group 15 Loss: 5.3912
+[2025-09-04 12:05:35] [Rank 0]   Group 15 Loss: 5.3912
+[2025-09-04 12:05:35] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:05:35] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 12:05:35] [Rank 0]   Group 13 FTA: 0.9800
+[2025-09-04 12:05:35] [Rank 0]   Group 14 FTA: 0.8600
+[2025-09-04 12:05:35] [Rank 0]   Group 14 FTA: 0.8600
+[2025-09-04 12:05:35] [Rank 0]   Group 15 FTA: 0.4400
+[2025-09-04 12:05:35] [Rank 0]   Group 15 FTA: 0.4400
+[2025-09-04 12:05:36] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:05:36] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:05:36] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:05:36] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:05:37] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:05:37] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:05:37] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:05:37] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:05:37] [Rank 0] step:6501/10000 train_time:285584ms step_avg:43.93ms
+[2025-09-04 12:05:37] [Rank 0] step:6501/10000 train_time:285584ms step_avg:43.93ms
+[2025-09-04 12:05:38] [Rank 0] step:6521/10000 train_time:286349ms step_avg:43.91ms
+[2025-09-04 12:05:38] [Rank 0] step:6521/10000 train_time:286349ms step_avg:43.91ms
+[2025-09-04 12:05:39] [Rank 0] step:6541/10000 train_time:287360ms step_avg:43.93ms
+[2025-09-04 12:05:39] [Rank 0] step:6541/10000 train_time:287360ms step_avg:43.93ms
+[2025-09-04 12:05:39] [Rank 0] step:6561/10000 train_time:288117ms step_avg:43.91ms
+[2025-09-04 12:05:39] [Rank 0] step:6561/10000 train_time:288117ms step_avg:43.91ms
+[2025-09-04 12:05:40] [Rank 0] step:6581/10000 train_time:288871ms step_avg:43.89ms
+[2025-09-04 12:05:40] [Rank 0] step:6581/10000 train_time:288871ms step_avg:43.89ms
+[2025-09-04 12:05:41] [Rank 0] step:6601/10000 train_time:289627ms step_avg:43.88ms
+[2025-09-04 12:05:41] [Rank 0] step:6601/10000 train_time:289627ms step_avg:43.88ms
+[2025-09-04 12:05:42] [Rank 0] step:6621/10000 train_time:290382ms step_avg:43.86ms
+[2025-09-04 12:05:42] [Rank 0] step:6621/10000 train_time:290382ms step_avg:43.86ms
+[2025-09-04 12:05:42] [Rank 0] step:6641/10000 train_time:291142ms step_avg:43.84ms
+[2025-09-04 12:05:42] [Rank 0] step:6641/10000 train_time:291142ms step_avg:43.84ms
+[2025-09-04 12:05:43] [Rank 0] step:6661/10000 train_time:291897ms step_avg:43.82ms
+[2025-09-04 12:05:43] [Rank 0] step:6661/10000 train_time:291897ms step_avg:43.82ms
+[2025-09-04 12:05:44] [Rank 0] step:6681/10000 train_time:292652ms step_avg:43.80ms
+[2025-09-04 12:05:44] [Rank 0] step:6681/10000 train_time:292652ms step_avg:43.80ms
+[2025-09-04 12:05:45] [Rank 0] step:6701/10000 train_time:293407ms step_avg:43.79ms
+[2025-09-04 12:05:45] [Rank 0] step:6701/10000 train_time:293407ms step_avg:43.79ms
+[2025-09-04 12:05:45] [Rank 0] step:6721/10000 train_time:294162ms step_avg:43.77ms
+[2025-09-04 12:05:45] [Rank 0] step:6721/10000 train_time:294162ms step_avg:43.77ms
+[2025-09-04 12:05:46] [Rank 0] step:6741/10000 train_time:294917ms step_avg:43.75ms
+[2025-09-04 12:05:46] [Rank 0] step:6741/10000 train_time:294917ms step_avg:43.75ms
+[2025-09-04 12:05:47] [Rank 0] step:6761/10000 train_time:295672ms step_avg:43.73ms
+[2025-09-04 12:05:47] [Rank 0] step:6761/10000 train_time:295672ms step_avg:43.73ms
+[2025-09-04 12:05:48] [Rank 0] step:6781/10000 train_time:296427ms step_avg:43.71ms
+[2025-09-04 12:05:48] [Rank 0] step:6781/10000 train_time:296427ms step_avg:43.71ms
+[2025-09-04 12:05:48] [Rank 0] step:6801/10000 train_time:297183ms step_avg:43.70ms
+[2025-09-04 12:05:48] [Rank 0] step:6801/10000 train_time:297183ms step_avg:43.70ms
+[2025-09-04 12:05:49] [Rank 0] step:6821/10000 train_time:297939ms step_avg:43.68ms
+[2025-09-04 12:05:49] [Rank 0] step:6821/10000 train_time:297939ms step_avg:43.68ms
+[2025-09-04 12:05:50] [Rank 0] step:6841/10000 train_time:298967ms step_avg:43.70ms
+[2025-09-04 12:05:50] [Rank 0] step:6841/10000 train_time:298967ms step_avg:43.70ms
+[2025-09-04 12:05:51] [Rank 0] step:6861/10000 train_time:299723ms step_avg:43.68ms
+[2025-09-04 12:05:51] [Rank 0] step:6861/10000 train_time:299723ms step_avg:43.68ms
+[2025-09-04 12:05:52] [Rank 0] step:6881/10000 train_time:300478ms step_avg:43.67ms
+[2025-09-04 12:05:52] [Rank 0] step:6881/10000 train_time:300478ms step_avg:43.67ms
+[2025-09-04 12:05:53] [Rank 0] step:6901/10000 train_time:301234ms step_avg:43.65ms
+[2025-09-04 12:05:53] [Rank 0] step:6901/10000 train_time:301234ms step_avg:43.65ms
+[2025-09-04 12:05:53] [Rank 0] step:6921/10000 train_time:301989ms step_avg:43.63ms
+[2025-09-04 12:05:53] [Rank 0] step:6921/10000 train_time:301989ms step_avg:43.63ms
+[2025-09-04 12:05:54] [Rank 0] step:6941/10000 train_time:302744ms step_avg:43.62ms
+[2025-09-04 12:05:54] [Rank 0] step:6941/10000 train_time:302744ms step_avg:43.62ms
+[2025-09-04 12:05:55] [Rank 0] step:6961/10000 train_time:303500ms step_avg:43.60ms
+[2025-09-04 12:05:55] [Rank 0] step:6961/10000 train_time:303500ms step_avg:43.60ms
+[2025-09-04 12:05:56] [Rank 0] step:6981/10000 train_time:304257ms step_avg:43.58ms
+[2025-09-04 12:05:56] [Rank 0] step:6981/10000 train_time:304257ms step_avg:43.58ms
+[2025-09-04 12:05:56] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:05:56] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:05:57] [Rank 0] PRINT: step:7000/10000 train_loss:0.6318 val_loss:0.6198 train_time:305017ms step_avg:43.57ms
+[2025-09-04 12:05:57] [Rank 0] PRINT: step:7000/10000 train_loss:0.6318 val_loss:0.6198 train_time:305017ms step_avg:43.57ms
+[2025-09-04 12:05:57] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:05:57] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:05:57] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:05:57] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:07:34] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:07:34] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:07:34] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:07:34] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:07:34] [Rank 0]   Total Loss: 5.1511
+[2025-09-04 12:07:34] [Rank 0]   Total Loss: 5.1511
+[2025-09-04 12:07:34] [Rank 0]   Total FTA (Unweighted): 0.9644
+[2025-09-04 12:07:34] [Rank 0]   Total FTA (Unweighted): 0.9644
+[2025-09-04 12:07:34] [Rank 0]   Total FTA (Weighted):   0.9644
+[2025-09-04 12:07:34] [Rank 0]   Total FTA (Weighted):   0.9644
+[2025-09-04 12:07:34] [Rank 0]   Group 0 Loss: 5.0487
+[2025-09-04 12:07:34] [Rank 0]   Group 0 Loss: 5.0487
+[2025-09-04 12:07:34] [Rank 0]   Group 1 Loss: 4.7026
+[2025-09-04 12:07:34] [Rank 0]   Group 1 Loss: 4.7026
+[2025-09-04 12:07:34] [Rank 0]   Group 2 Loss: 4.6117
+[2025-09-04 12:07:34] [Rank 0]   Group 2 Loss: 4.6117
+[2025-09-04 12:07:34] [Rank 0]   Group 3 Loss: 5.0489
+[2025-09-04 12:07:34] [Rank 0]   Group 3 Loss: 5.0489
+[2025-09-04 12:07:34] [Rank 0]   Group 4 Loss: 5.0654
+[2025-09-04 12:07:34] [Rank 0]   Group 4 Loss: 5.0654
+[2025-09-04 12:07:34] [Rank 0]   Group 5 Loss: 5.0629
+[2025-09-04 12:07:34] [Rank 0]   Group 5 Loss: 5.0629
+[2025-09-04 12:07:34] [Rank 0]   Group 6 Loss: 4.9948
+[2025-09-04 12:07:34] [Rank 0]   Group 6 Loss: 4.9948
+[2025-09-04 12:07:34] [Rank 0]   Group 7 Loss: 5.0512
+[2025-09-04 12:07:34] [Rank 0]   Group 7 Loss: 5.0512
+[2025-09-04 12:07:34] [Rank 0]   Group 8 Loss: 5.2514
+[2025-09-04 12:07:34] [Rank 0]   Group 8 Loss: 5.2514
+[2025-09-04 12:07:34] [Rank 0]   Group 9 Loss: 5.2311
+[2025-09-04 12:07:34] [Rank 0]   Group 9 Loss: 5.2311
+[2025-09-04 12:07:34] [Rank 0]   Group 10 Loss: 5.3582
+[2025-09-04 12:07:34] [Rank 0]   Group 10 Loss: 5.3582
+[2025-09-04 12:07:34] [Rank 0]   Group 11 Loss: 5.3866
+[2025-09-04 12:07:34] [Rank 0]   Group 11 Loss: 5.3866
+[2025-09-04 12:07:34] [Rank 0]   Group 12 Loss: 5.3027
+[2025-09-04 12:07:34] [Rank 0]   Group 12 Loss: 5.3027
+[2025-09-04 12:07:34] [Rank 0]   Group 13 Loss: 5.5173
+[2025-09-04 12:07:34] [Rank 0]   Group 13 Loss: 5.5173
+[2025-09-04 12:07:34] [Rank 0]   Group 14 Loss: 5.4184
+[2025-09-04 12:07:34] [Rank 0]   Group 14 Loss: 5.4184
+[2025-09-04 12:07:34] [Rank 0]   Group 15 Loss: 5.3652
+[2025-09-04 12:07:34] [Rank 0]   Group 15 Loss: 5.3652
+[2025-09-04 12:07:34] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:07:34] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:07:34] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:07:34] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:07:34] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:07:34] [Rank 0]   Group 14 FTA: 0.9000
+[2025-09-04 12:07:34] [Rank 0]   Group 14 FTA: 0.9000
+[2025-09-04 12:07:34] [Rank 0]   Group 15 FTA: 0.5500
+[2025-09-04 12:07:34] [Rank 0]   Group 15 FTA: 0.5500
+[2025-09-04 12:07:35] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:07:35] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:07:35] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:07:35] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:07:35] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:07:35] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:07:36] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:07:36] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:07:36] [Rank 0] step:7001/10000 train_time:305032ms step_avg:43.57ms
+[2025-09-04 12:07:36] [Rank 0] step:7001/10000 train_time:305032ms step_avg:43.57ms
+[2025-09-04 12:07:36] [Rank 0] step:7021/10000 train_time:305805ms step_avg:43.56ms
+[2025-09-04 12:07:36] [Rank 0] step:7021/10000 train_time:305805ms step_avg:43.56ms
+[2025-09-04 12:07:37] [Rank 0] step:7041/10000 train_time:306561ms step_avg:43.54ms
+[2025-09-04 12:07:37] [Rank 0] step:7041/10000 train_time:306561ms step_avg:43.54ms
+[2025-09-04 12:07:38] [Rank 0] step:7061/10000 train_time:307316ms step_avg:43.52ms
+[2025-09-04 12:07:38] [Rank 0] step:7061/10000 train_time:307316ms step_avg:43.52ms
+[2025-09-04 12:07:39] [Rank 0] step:7081/10000 train_time:308070ms step_avg:43.51ms
+[2025-09-04 12:07:39] [Rank 0] step:7081/10000 train_time:308070ms step_avg:43.51ms
+[2025-09-04 12:07:39] [Rank 0] step:7101/10000 train_time:308825ms step_avg:43.49ms
+[2025-09-04 12:07:39] [Rank 0] step:7101/10000 train_time:308825ms step_avg:43.49ms
+[2025-09-04 12:07:40] [Rank 0] step:7121/10000 train_time:309580ms step_avg:43.47ms
+[2025-09-04 12:07:40] [Rank 0] step:7121/10000 train_time:309580ms step_avg:43.47ms
+[2025-09-04 12:07:41] [Rank 0] step:7141/10000 train_time:310336ms step_avg:43.46ms
+[2025-09-04 12:07:41] [Rank 0] step:7141/10000 train_time:310336ms step_avg:43.46ms
+[2025-09-04 12:07:42] [Rank 0] step:7161/10000 train_time:311335ms step_avg:43.48ms
+[2025-09-04 12:07:42] [Rank 0] step:7161/10000 train_time:311335ms step_avg:43.48ms
+[2025-09-04 12:07:43] [Rank 0] step:7181/10000 train_time:312089ms step_avg:43.46ms
+[2025-09-04 12:07:43] [Rank 0] step:7181/10000 train_time:312089ms step_avg:43.46ms
+[2025-09-04 12:07:43] [Rank 0] step:7201/10000 train_time:312844ms step_avg:43.44ms
+[2025-09-04 12:07:43] [Rank 0] step:7201/10000 train_time:312844ms step_avg:43.44ms
+[2025-09-04 12:07:45] [Rank 0] step:7221/10000 train_time:313867ms step_avg:43.47ms
+[2025-09-04 12:07:45] [Rank 0] step:7221/10000 train_time:313867ms step_avg:43.47ms
+[2025-09-04 12:07:45] [Rank 0] step:7241/10000 train_time:314622ms step_avg:43.45ms
+[2025-09-04 12:07:45] [Rank 0] step:7241/10000 train_time:314622ms step_avg:43.45ms
+[2025-09-04 12:07:46] [Rank 0] step:7261/10000 train_time:315376ms step_avg:43.43ms
+[2025-09-04 12:07:46] [Rank 0] step:7261/10000 train_time:315376ms step_avg:43.43ms
+[2025-09-04 12:07:47] [Rank 0] step:7281/10000 train_time:316132ms step_avg:43.42ms
+[2025-09-04 12:07:47] [Rank 0] step:7281/10000 train_time:316132ms step_avg:43.42ms
+[2025-09-04 12:07:48] [Rank 0] step:7301/10000 train_time:316888ms step_avg:43.40ms
+[2025-09-04 12:07:48] [Rank 0] step:7301/10000 train_time:316888ms step_avg:43.40ms
+[2025-09-04 12:07:48] [Rank 0] step:7321/10000 train_time:317643ms step_avg:43.39ms
+[2025-09-04 12:07:48] [Rank 0] step:7321/10000 train_time:317643ms step_avg:43.39ms
+[2025-09-04 12:07:49] [Rank 0] step:7341/10000 train_time:318399ms step_avg:43.37ms
+[2025-09-04 12:07:49] [Rank 0] step:7341/10000 train_time:318399ms step_avg:43.37ms
+[2025-09-04 12:07:50] [Rank 0] step:7361/10000 train_time:319154ms step_avg:43.36ms
+[2025-09-04 12:07:50] [Rank 0] step:7361/10000 train_time:319154ms step_avg:43.36ms
+[2025-09-04 12:07:51] [Rank 0] step:7381/10000 train_time:319910ms step_avg:43.34ms
+[2025-09-04 12:07:51] [Rank 0] step:7381/10000 train_time:319910ms step_avg:43.34ms
+[2025-09-04 12:07:51] [Rank 0] step:7401/10000 train_time:320665ms step_avg:43.33ms
+[2025-09-04 12:07:51] [Rank 0] step:7401/10000 train_time:320665ms step_avg:43.33ms
+[2025-09-04 12:07:52] [Rank 0] step:7421/10000 train_time:321421ms step_avg:43.31ms
+[2025-09-04 12:07:52] [Rank 0] step:7421/10000 train_time:321421ms step_avg:43.31ms
+[2025-09-04 12:07:53] [Rank 0] step:7441/10000 train_time:322177ms step_avg:43.30ms
+[2025-09-04 12:07:53] [Rank 0] step:7441/10000 train_time:322177ms step_avg:43.30ms
+[2025-09-04 12:07:54] [Rank 0] step:7461/10000 train_time:322934ms step_avg:43.28ms
+[2025-09-04 12:07:54] [Rank 0] step:7461/10000 train_time:322934ms step_avg:43.28ms
+[2025-09-04 12:07:54] [Rank 0] step:7481/10000 train_time:323687ms step_avg:43.27ms
+[2025-09-04 12:07:54] [Rank 0] step:7481/10000 train_time:323687ms step_avg:43.27ms
+[2025-09-04 12:07:55] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:07:55] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:07:56] [Rank 0] PRINT: step:7500/10000 train_loss:0.6264 val_loss:0.6162 train_time:324447ms step_avg:43.26ms
+[2025-09-04 12:07:56] [Rank 0] PRINT: step:7500/10000 train_loss:0.6264 val_loss:0.6162 train_time:324447ms step_avg:43.26ms
+[2025-09-04 12:07:56] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:07:56] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:07:56] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:07:56] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:09:33] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:09:33] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:09:33] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:09:33] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:09:33] [Rank 0]   Total Loss: 5.1189
+[2025-09-04 12:09:33] [Rank 0]   Total Loss: 5.1189
+[2025-09-04 12:09:33] [Rank 0]   Total FTA (Unweighted): 0.9744
+[2025-09-04 12:09:33] [Rank 0]   Total FTA (Unweighted): 0.9744
+[2025-09-04 12:09:33] [Rank 0]   Total FTA (Weighted):   0.9744
+[2025-09-04 12:09:33] [Rank 0]   Total FTA (Weighted):   0.9744
+[2025-09-04 12:09:33] [Rank 0]   Group 0 Loss: 5.0087
+[2025-09-04 12:09:33] [Rank 0]   Group 0 Loss: 5.0087
+[2025-09-04 12:09:33] [Rank 0]   Group 1 Loss: 4.6728
+[2025-09-04 12:09:33] [Rank 0]   Group 1 Loss: 4.6728
+[2025-09-04 12:09:33] [Rank 0]   Group 2 Loss: 4.5960
+[2025-09-04 12:09:33] [Rank 0]   Group 2 Loss: 4.5960
+[2025-09-04 12:09:34] [Rank 0]   Group 3 Loss: 5.0044
+[2025-09-04 12:09:34] [Rank 0]   Group 3 Loss: 5.0044
+[2025-09-04 12:09:34] [Rank 0]   Group 4 Loss: 5.0298
+[2025-09-04 12:09:34] [Rank 0]   Group 4 Loss: 5.0298
+[2025-09-04 12:09:34] [Rank 0]   Group 5 Loss: 5.0438
+[2025-09-04 12:09:34] [Rank 0]   Group 5 Loss: 5.0438
+[2025-09-04 12:09:34] [Rank 0]   Group 6 Loss: 4.9598
+[2025-09-04 12:09:34] [Rank 0]   Group 6 Loss: 4.9598
+[2025-09-04 12:09:34] [Rank 0]   Group 7 Loss: 5.0462
+[2025-09-04 12:09:34] [Rank 0]   Group 7 Loss: 5.0462
+[2025-09-04 12:09:34] [Rank 0]   Group 8 Loss: 5.2187
+[2025-09-04 12:09:34] [Rank 0]   Group 8 Loss: 5.2187
+[2025-09-04 12:09:34] [Rank 0]   Group 9 Loss: 5.1776
+[2025-09-04 12:09:34] [Rank 0]   Group 9 Loss: 5.1776
+[2025-09-04 12:09:34] [Rank 0]   Group 10 Loss: 5.3405
+[2025-09-04 12:09:34] [Rank 0]   Group 10 Loss: 5.3405
+[2025-09-04 12:09:34] [Rank 0]   Group 11 Loss: 5.3420
+[2025-09-04 12:09:34] [Rank 0]   Group 11 Loss: 5.3420
+[2025-09-04 12:09:34] [Rank 0]   Group 12 Loss: 5.2961
+[2025-09-04 12:09:34] [Rank 0]   Group 12 Loss: 5.2961
+[2025-09-04 12:09:34] [Rank 0]   Group 13 Loss: 5.4831
+[2025-09-04 12:09:34] [Rank 0]   Group 13 Loss: 5.4831
+[2025-09-04 12:09:34] [Rank 0]   Group 14 Loss: 5.3452
+[2025-09-04 12:09:34] [Rank 0]   Group 14 Loss: 5.3452
+[2025-09-04 12:09:34] [Rank 0]   Group 15 Loss: 5.3373
+[2025-09-04 12:09:34] [Rank 0]   Group 15 Loss: 5.3373
+[2025-09-04 12:09:34] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:09:34] [Rank 0]   Group 14 FTA: 0.9300
+[2025-09-04 12:09:34] [Rank 0]   Group 14 FTA: 0.9300
+[2025-09-04 12:09:34] [Rank 0]   Group 15 FTA: 0.6600
+[2025-09-04 12:09:34] [Rank 0]   Group 15 FTA: 0.6600
+[2025-09-04 12:09:34] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:09:34] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:09:34] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:09:34] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:09:35] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:09:35] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:09:35] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:09:35] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:09:35] [Rank 0] step:7501/10000 train_time:324463ms step_avg:43.26ms
+[2025-09-04 12:09:35] [Rank 0] step:7501/10000 train_time:324463ms step_avg:43.26ms
+[2025-09-04 12:09:36] [Rank 0] step:7521/10000 train_time:325225ms step_avg:43.24ms
+[2025-09-04 12:09:36] [Rank 0] step:7521/10000 train_time:325225ms step_avg:43.24ms
+[2025-09-04 12:09:37] [Rank 0] step:7541/10000 train_time:325979ms step_avg:43.23ms
+[2025-09-04 12:09:37] [Rank 0] step:7541/10000 train_time:325979ms step_avg:43.23ms
+[2025-09-04 12:09:37] [Rank 0] step:7561/10000 train_time:326735ms step_avg:43.21ms
+[2025-09-04 12:09:37] [Rank 0] step:7561/10000 train_time:326735ms step_avg:43.21ms
+[2025-09-04 12:09:38] [Rank 0] step:7581/10000 train_time:327490ms step_avg:43.20ms
+[2025-09-04 12:09:38] [Rank 0] step:7581/10000 train_time:327490ms step_avg:43.20ms
+[2025-09-04 12:09:39] [Rank 0] step:7601/10000 train_time:328245ms step_avg:43.18ms
+[2025-09-04 12:09:39] [Rank 0] step:7601/10000 train_time:328245ms step_avg:43.18ms
+[2025-09-04 12:09:40] [Rank 0] step:7621/10000 train_time:329000ms step_avg:43.17ms
+[2025-09-04 12:09:40] [Rank 0] step:7621/10000 train_time:329000ms step_avg:43.17ms
+[2025-09-04 12:09:41] [Rank 0] step:7641/10000 train_time:330023ms step_avg:43.19ms
+[2025-09-04 12:09:41] [Rank 0] step:7641/10000 train_time:330023ms step_avg:43.19ms
+[2025-09-04 12:09:41] [Rank 0] step:7661/10000 train_time:330777ms step_avg:43.18ms
+[2025-09-04 12:09:41] [Rank 0] step:7661/10000 train_time:330777ms step_avg:43.18ms
+[2025-09-04 12:09:42] [Rank 0] step:7681/10000 train_time:331532ms step_avg:43.16ms
+[2025-09-04 12:09:42] [Rank 0] step:7681/10000 train_time:331532ms step_avg:43.16ms
+[2025-09-04 12:09:43] [Rank 0] step:7701/10000 train_time:332295ms step_avg:43.15ms
+[2025-09-04 12:09:43] [Rank 0] step:7701/10000 train_time:332295ms step_avg:43.15ms
+[2025-09-04 12:09:44] [Rank 0] step:7721/10000 train_time:333050ms step_avg:43.14ms
+[2025-09-04 12:09:44] [Rank 0] step:7721/10000 train_time:333050ms step_avg:43.14ms
+[2025-09-04 12:09:44] [Rank 0] step:7741/10000 train_time:333804ms step_avg:43.12ms
+[2025-09-04 12:09:44] [Rank 0] step:7741/10000 train_time:333804ms step_avg:43.12ms
+[2025-09-04 12:09:45] [Rank 0] step:7761/10000 train_time:334560ms step_avg:43.11ms
+[2025-09-04 12:09:45] [Rank 0] step:7761/10000 train_time:334560ms step_avg:43.11ms
+[2025-09-04 12:09:46] [Rank 0] step:7781/10000 train_time:335314ms step_avg:43.09ms
+[2025-09-04 12:09:46] [Rank 0] step:7781/10000 train_time:335314ms step_avg:43.09ms
+[2025-09-04 12:09:47] [Rank 0] step:7801/10000 train_time:336069ms step_avg:43.08ms
+[2025-09-04 12:09:47] [Rank 0] step:7801/10000 train_time:336069ms step_avg:43.08ms
+[2025-09-04 12:09:47] [Rank 0] step:7821/10000 train_time:336823ms step_avg:43.07ms
+[2025-09-04 12:09:47] [Rank 0] step:7821/10000 train_time:336823ms step_avg:43.07ms
+[2025-09-04 12:09:48] [Rank 0] step:7841/10000 train_time:337838ms step_avg:43.09ms
+[2025-09-04 12:09:48] [Rank 0] step:7841/10000 train_time:337838ms step_avg:43.09ms
+[2025-09-04 12:09:49] [Rank 0] step:7861/10000 train_time:338593ms step_avg:43.07ms
+[2025-09-04 12:09:49] [Rank 0] step:7861/10000 train_time:338593ms step_avg:43.07ms
+[2025-09-04 12:09:50] [Rank 0] step:7881/10000 train_time:339348ms step_avg:43.06ms
+[2025-09-04 12:09:50] [Rank 0] step:7881/10000 train_time:339348ms step_avg:43.06ms
+[2025-09-04 12:09:51] [Rank 0] step:7901/10000 train_time:340379ms step_avg:43.08ms
+[2025-09-04 12:09:51] [Rank 0] step:7901/10000 train_time:340379ms step_avg:43.08ms
+[2025-09-04 12:09:52] [Rank 0] step:7921/10000 train_time:341134ms step_avg:43.07ms
+[2025-09-04 12:09:52] [Rank 0] step:7921/10000 train_time:341134ms step_avg:43.07ms
+[2025-09-04 12:09:53] [Rank 0] step:7941/10000 train_time:341889ms step_avg:43.05ms
+[2025-09-04 12:09:53] [Rank 0] step:7941/10000 train_time:341889ms step_avg:43.05ms
+[2025-09-04 12:09:53] [Rank 0] step:7961/10000 train_time:342655ms step_avg:43.04ms
+[2025-09-04 12:09:53] [Rank 0] step:7961/10000 train_time:342655ms step_avg:43.04ms
+[2025-09-04 12:09:54] [Rank 0] step:7981/10000 train_time:343410ms step_avg:43.03ms
+[2025-09-04 12:09:54] [Rank 0] step:7981/10000 train_time:343410ms step_avg:43.03ms
+[2025-09-04 12:09:55] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:09:55] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:09:55] [Rank 0] PRINT: step:8000/10000 train_loss:0.6217 val_loss:0.6128 train_time:344169ms step_avg:43.02ms
+[2025-09-04 12:09:55] [Rank 0] PRINT: step:8000/10000 train_loss:0.6217 val_loss:0.6128 train_time:344169ms step_avg:43.02ms
+[2025-09-04 12:09:55] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:09:55] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:09:55] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:09:55] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:11:33] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:11:33] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:11:33] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:11:33] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:11:33] [Rank 0]   Total Loss: 5.0931
+[2025-09-04 12:11:33] [Rank 0]   Total Loss: 5.0931
+[2025-09-04 12:11:33] [Rank 0]   Total FTA (Unweighted): 0.9844
+[2025-09-04 12:11:33] [Rank 0]   Total FTA (Unweighted): 0.9844
+[2025-09-04 12:11:33] [Rank 0]   Total FTA (Weighted):   0.9844
+[2025-09-04 12:11:33] [Rank 0]   Total FTA (Weighted):   0.9844
+[2025-09-04 12:11:33] [Rank 0]   Group 0 Loss: 4.9177
+[2025-09-04 12:11:33] [Rank 0]   Group 0 Loss: 4.9177
+[2025-09-04 12:11:33] [Rank 0]   Group 1 Loss: 4.6542
+[2025-09-04 12:11:33] [Rank 0]   Group 1 Loss: 4.6542
+[2025-09-04 12:11:33] [Rank 0]   Group 2 Loss: 4.5469
+[2025-09-04 12:11:33] [Rank 0]   Group 2 Loss: 4.5469
+[2025-09-04 12:11:33] [Rank 0]   Group 3 Loss: 4.9732
+[2025-09-04 12:11:33] [Rank 0]   Group 3 Loss: 4.9732
+[2025-09-04 12:11:33] [Rank 0]   Group 4 Loss: 4.9954
+[2025-09-04 12:11:33] [Rank 0]   Group 4 Loss: 4.9954
+[2025-09-04 12:11:33] [Rank 0]   Group 5 Loss: 5.0380
+[2025-09-04 12:11:33] [Rank 0]   Group 5 Loss: 5.0380
+[2025-09-04 12:11:33] [Rank 0]   Group 6 Loss: 4.9305
+[2025-09-04 12:11:33] [Rank 0]   Group 6 Loss: 4.9305
+[2025-09-04 12:11:33] [Rank 0]   Group 7 Loss: 5.0104
+[2025-09-04 12:11:33] [Rank 0]   Group 7 Loss: 5.0104
+[2025-09-04 12:11:33] [Rank 0]   Group 8 Loss: 5.2015
+[2025-09-04 12:11:33] [Rank 0]   Group 8 Loss: 5.2015
+[2025-09-04 12:11:33] [Rank 0]   Group 9 Loss: 5.1531
+[2025-09-04 12:11:33] [Rank 0]   Group 9 Loss: 5.1531
+[2025-09-04 12:11:33] [Rank 0]   Group 10 Loss: 5.3466
+[2025-09-04 12:11:33] [Rank 0]   Group 10 Loss: 5.3466
+[2025-09-04 12:11:33] [Rank 0]   Group 11 Loss: 5.3369
+[2025-09-04 12:11:33] [Rank 0]   Group 11 Loss: 5.3369
+[2025-09-04 12:11:33] [Rank 0]   Group 12 Loss: 5.2684
+[2025-09-04 12:11:33] [Rank 0]   Group 12 Loss: 5.2684
+[2025-09-04 12:11:33] [Rank 0]   Group 13 Loss: 5.4553
+[2025-09-04 12:11:33] [Rank 0]   Group 13 Loss: 5.4553
+[2025-09-04 12:11:33] [Rank 0]   Group 14 Loss: 5.3452
+[2025-09-04 12:11:33] [Rank 0]   Group 14 Loss: 5.3452
+[2025-09-04 12:11:33] [Rank 0]   Group 15 Loss: 5.3170
+[2025-09-04 12:11:33] [Rank 0]   Group 15 Loss: 5.3170
+[2025-09-04 12:11:33] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:11:33] [Rank 0]   Group 14 FTA: 0.9600
+[2025-09-04 12:11:33] [Rank 0]   Group 14 FTA: 0.9600
+[2025-09-04 12:11:33] [Rank 0]   Group 15 FTA: 0.7900
+[2025-09-04 12:11:33] [Rank 0]   Group 15 FTA: 0.7900
+[2025-09-04 12:11:34] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:11:34] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:11:34] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:11:34] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:11:34] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:11:34] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:11:35] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:11:35] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:11:35] [Rank 0] step:8001/10000 train_time:344185ms step_avg:43.02ms
+[2025-09-04 12:11:35] [Rank 0] step:8001/10000 train_time:344185ms step_avg:43.02ms
+[2025-09-04 12:11:36] [Rank 0] step:8021/10000 train_time:345223ms step_avg:43.04ms
+[2025-09-04 12:11:36] [Rank 0] step:8021/10000 train_time:345223ms step_avg:43.04ms
+[2025-09-04 12:11:37] [Rank 0] step:8041/10000 train_time:345978ms step_avg:43.03ms
+[2025-09-04 12:11:37] [Rank 0] step:8041/10000 train_time:345978ms step_avg:43.03ms
+[2025-09-04 12:11:37] [Rank 0] step:8061/10000 train_time:346732ms step_avg:43.01ms
+[2025-09-04 12:11:37] [Rank 0] step:8061/10000 train_time:346732ms step_avg:43.01ms
+[2025-09-04 12:11:38] [Rank 0] step:8081/10000 train_time:347486ms step_avg:43.00ms
+[2025-09-04 12:11:38] [Rank 0] step:8081/10000 train_time:347486ms step_avg:43.00ms
+[2025-09-04 12:11:39] [Rank 0] step:8101/10000 train_time:348241ms step_avg:42.99ms
+[2025-09-04 12:11:39] [Rank 0] step:8101/10000 train_time:348241ms step_avg:42.99ms
+[2025-09-04 12:11:40] [Rank 0] step:8121/10000 train_time:348996ms step_avg:42.97ms
+[2025-09-04 12:11:40] [Rank 0] step:8121/10000 train_time:348996ms step_avg:42.97ms
+[2025-09-04 12:11:40] [Rank 0] step:8141/10000 train_time:349749ms step_avg:42.96ms
+[2025-09-04 12:11:40] [Rank 0] step:8141/10000 train_time:349749ms step_avg:42.96ms
+[2025-09-04 12:11:41] [Rank 0] step:8161/10000 train_time:350504ms step_avg:42.95ms
+[2025-09-04 12:11:41] [Rank 0] step:8161/10000 train_time:350504ms step_avg:42.95ms
+[2025-09-04 12:11:42] [Rank 0] step:8181/10000 train_time:351258ms step_avg:42.94ms
+[2025-09-04 12:11:42] [Rank 0] step:8181/10000 train_time:351258ms step_avg:42.94ms
+[2025-09-04 12:11:43] [Rank 0] step:8201/10000 train_time:352013ms step_avg:42.92ms
+[2025-09-04 12:11:43] [Rank 0] step:8201/10000 train_time:352013ms step_avg:42.92ms
+[2025-09-04 12:11:43] [Rank 0] step:8221/10000 train_time:352768ms step_avg:42.91ms
+[2025-09-04 12:11:43] [Rank 0] step:8221/10000 train_time:352768ms step_avg:42.91ms
+[2025-09-04 12:11:44] [Rank 0] step:8241/10000 train_time:353522ms step_avg:42.90ms
+[2025-09-04 12:11:44] [Rank 0] step:8241/10000 train_time:353522ms step_avg:42.90ms
+[2025-09-04 12:11:45] [Rank 0] step:8261/10000 train_time:354276ms step_avg:42.89ms
+[2025-09-04 12:11:45] [Rank 0] step:8261/10000 train_time:354276ms step_avg:42.89ms
+[2025-09-04 12:11:46] [Rank 0] step:8281/10000 train_time:355030ms step_avg:42.87ms
+[2025-09-04 12:11:46] [Rank 0] step:8281/10000 train_time:355030ms step_avg:42.87ms
+[2025-09-04 12:11:46] [Rank 0] step:8301/10000 train_time:355785ms step_avg:42.86ms
+[2025-09-04 12:11:46] [Rank 0] step:8301/10000 train_time:355785ms step_avg:42.86ms
+[2025-09-04 12:11:47] [Rank 0] step:8321/10000 train_time:356540ms step_avg:42.85ms
+[2025-09-04 12:11:47] [Rank 0] step:8321/10000 train_time:356540ms step_avg:42.85ms
+[2025-09-04 12:11:48] [Rank 0] step:8341/10000 train_time:357294ms step_avg:42.84ms
+[2025-09-04 12:11:48] [Rank 0] step:8341/10000 train_time:357294ms step_avg:42.84ms
+[2025-09-04 12:11:49] [Rank 0] step:8361/10000 train_time:358049ms step_avg:42.82ms
+[2025-09-04 12:11:49] [Rank 0] step:8361/10000 train_time:358049ms step_avg:42.82ms
+[2025-09-04 12:11:49] [Rank 0] step:8381/10000 train_time:358803ms step_avg:42.81ms
+[2025-09-04 12:11:49] [Rank 0] step:8381/10000 train_time:358803ms step_avg:42.81ms
+[2025-09-04 12:11:50] [Rank 0] step:8401/10000 train_time:359558ms step_avg:42.80ms
+[2025-09-04 12:11:50] [Rank 0] step:8401/10000 train_time:359558ms step_avg:42.80ms
+[2025-09-04 12:11:51] [Rank 0] step:8421/10000 train_time:360312ms step_avg:42.79ms
+[2025-09-04 12:11:51] [Rank 0] step:8421/10000 train_time:360312ms step_avg:42.79ms
+[2025-09-04 12:11:52] [Rank 0] step:8441/10000 train_time:361068ms step_avg:42.78ms
+[2025-09-04 12:11:52] [Rank 0] step:8441/10000 train_time:361068ms step_avg:42.78ms
+[2025-09-04 12:11:52] [Rank 0] step:8461/10000 train_time:361822ms step_avg:42.76ms
+[2025-09-04 12:11:52] [Rank 0] step:8461/10000 train_time:361822ms step_avg:42.76ms
+[2025-09-04 12:11:53] [Rank 0] step:8481/10000 train_time:362580ms step_avg:42.75ms
+[2025-09-04 12:11:53] [Rank 0] step:8481/10000 train_time:362580ms step_avg:42.75ms
+[2025-09-04 12:11:54] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:11:54] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:11:54] [Rank 0] PRINT: step:8500/10000 train_loss:0.6176 val_loss:0.6099 train_time:363339ms step_avg:42.75ms
+[2025-09-04 12:11:54] [Rank 0] PRINT: step:8500/10000 train_loss:0.6176 val_loss:0.6099 train_time:363339ms step_avg:42.75ms
+[2025-09-04 12:11:54] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:11:54] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:11:55] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:11:55] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:13:31] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:13:31] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:13:32] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:13:32] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:13:32] [Rank 0]   Total Loss: 5.1675
+[2025-09-04 12:13:32] [Rank 0]   Total Loss: 5.1675
+[2025-09-04 12:13:32] [Rank 0]   Total FTA (Unweighted): 0.9875
+[2025-09-04 12:13:32] [Rank 0]   Total FTA (Unweighted): 0.9875
+[2025-09-04 12:13:32] [Rank 0]   Total FTA (Weighted):   0.9875
+[2025-09-04 12:13:32] [Rank 0]   Total FTA (Weighted):   0.9875
+[2025-09-04 12:13:32] [Rank 0]   Group 0 Loss: 5.0650
+[2025-09-04 12:13:32] [Rank 0]   Group 0 Loss: 5.0650
+[2025-09-04 12:13:32] [Rank 0]   Group 1 Loss: 4.7586
+[2025-09-04 12:13:32] [Rank 0]   Group 1 Loss: 4.7586
+[2025-09-04 12:13:32] [Rank 0]   Group 2 Loss: 4.6202
+[2025-09-04 12:13:32] [Rank 0]   Group 2 Loss: 4.6202
+[2025-09-04 12:13:32] [Rank 0]   Group 3 Loss: 5.0470
+[2025-09-04 12:13:32] [Rank 0]   Group 3 Loss: 5.0470
+[2025-09-04 12:13:32] [Rank 0]   Group 4 Loss: 5.0797
+[2025-09-04 12:13:32] [Rank 0]   Group 4 Loss: 5.0797
+[2025-09-04 12:13:32] [Rank 0]   Group 5 Loss: 5.0814
+[2025-09-04 12:13:32] [Rank 0]   Group 5 Loss: 5.0814
+[2025-09-04 12:13:32] [Rank 0]   Group 6 Loss: 4.9797
+[2025-09-04 12:13:32] [Rank 0]   Group 6 Loss: 4.9797
+[2025-09-04 12:13:32] [Rank 0]   Group 7 Loss: 5.0875
+[2025-09-04 12:13:32] [Rank 0]   Group 7 Loss: 5.0875
+[2025-09-04 12:13:32] [Rank 0]   Group 8 Loss: 5.2643
+[2025-09-04 12:13:32] [Rank 0]   Group 8 Loss: 5.2643
+[2025-09-04 12:13:32] [Rank 0]   Group 9 Loss: 5.1868
+[2025-09-04 12:13:32] [Rank 0]   Group 9 Loss: 5.1868
+[2025-09-04 12:13:32] [Rank 0]   Group 10 Loss: 5.3966
+[2025-09-04 12:13:32] [Rank 0]   Group 10 Loss: 5.3966
+[2025-09-04 12:13:32] [Rank 0]   Group 11 Loss: 5.3880
+[2025-09-04 12:13:32] [Rank 0]   Group 11 Loss: 5.3880
+[2025-09-04 12:13:32] [Rank 0]   Group 12 Loss: 5.3398
+[2025-09-04 12:13:32] [Rank 0]   Group 12 Loss: 5.3398
+[2025-09-04 12:13:32] [Rank 0]   Group 13 Loss: 5.5519
+[2025-09-04 12:13:32] [Rank 0]   Group 13 Loss: 5.5519
+[2025-09-04 12:13:32] [Rank 0]   Group 14 Loss: 5.4263
+[2025-09-04 12:13:32] [Rank 0]   Group 14 Loss: 5.4263
+[2025-09-04 12:13:32] [Rank 0]   Group 15 Loss: 5.4079
+[2025-09-04 12:13:32] [Rank 0]   Group 15 Loss: 5.4079
+[2025-09-04 12:13:32] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:13:32] [Rank 0]   Group 15 FTA: 0.8000
+[2025-09-04 12:13:32] [Rank 0]   Group 15 FTA: 0.8000
+[2025-09-04 12:13:32] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:13:32] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:13:33] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:13:33] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:13:33] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:13:33] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:13:33] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:13:33] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:13:33] [Rank 0] step:8501/10000 train_time:363358ms step_avg:42.74ms
+[2025-09-04 12:13:33] [Rank 0] step:8501/10000 train_time:363358ms step_avg:42.74ms
+[2025-09-04 12:13:34] [Rank 0] step:8521/10000 train_time:364118ms step_avg:42.73ms
+[2025-09-04 12:13:34] [Rank 0] step:8521/10000 train_time:364118ms step_avg:42.73ms
+[2025-09-04 12:13:35] [Rank 0] step:8541/10000 train_time:364872ms step_avg:42.72ms
+[2025-09-04 12:13:35] [Rank 0] step:8541/10000 train_time:364872ms step_avg:42.72ms
+[2025-09-04 12:13:35] [Rank 0] step:8561/10000 train_time:365627ms step_avg:42.71ms
+[2025-09-04 12:13:35] [Rank 0] step:8561/10000 train_time:365627ms step_avg:42.71ms
+[2025-09-04 12:13:36] [Rank 0] step:8581/10000 train_time:366381ms step_avg:42.70ms
+[2025-09-04 12:13:36] [Rank 0] step:8581/10000 train_time:366381ms step_avg:42.70ms
+[2025-09-04 12:13:37] [Rank 0] step:8601/10000 train_time:367136ms step_avg:42.69ms
+[2025-09-04 12:13:37] [Rank 0] step:8601/10000 train_time:367136ms step_avg:42.69ms
+[2025-09-04 12:13:38] [Rank 0] step:8621/10000 train_time:367890ms step_avg:42.67ms
+[2025-09-04 12:13:38] [Rank 0] step:8621/10000 train_time:367890ms step_avg:42.67ms
+[2025-09-04 12:13:38] [Rank 0] step:8641/10000 train_time:368645ms step_avg:42.66ms
+[2025-09-04 12:13:38] [Rank 0] step:8641/10000 train_time:368645ms step_avg:42.66ms
+[2025-09-04 12:13:39] [Rank 0] step:8661/10000 train_time:369400ms step_avg:42.65ms
+[2025-09-04 12:13:39] [Rank 0] step:8661/10000 train_time:369400ms step_avg:42.65ms
+[2025-09-04 12:13:40] [Rank 0] step:8681/10000 train_time:370155ms step_avg:42.64ms
+[2025-09-04 12:13:40] [Rank 0] step:8681/10000 train_time:370155ms step_avg:42.64ms
+[2025-09-04 12:13:41] [Rank 0] step:8701/10000 train_time:370909ms step_avg:42.63ms
+[2025-09-04 12:13:41] [Rank 0] step:8701/10000 train_time:370909ms step_avg:42.63ms
+[2025-09-04 12:13:41] [Rank 0] step:8721/10000 train_time:371665ms step_avg:42.62ms
+[2025-09-04 12:13:41] [Rank 0] step:8721/10000 train_time:371665ms step_avg:42.62ms
+[2025-09-04 12:13:42] [Rank 0] step:8741/10000 train_time:372423ms step_avg:42.61ms
+[2025-09-04 12:13:42] [Rank 0] step:8741/10000 train_time:372423ms step_avg:42.61ms
+[2025-09-04 12:13:43] [Rank 0] step:8761/10000 train_time:373178ms step_avg:42.60ms
+[2025-09-04 12:13:43] [Rank 0] step:8761/10000 train_time:373178ms step_avg:42.60ms
+[2025-09-04 12:13:44] [Rank 0] step:8781/10000 train_time:373933ms step_avg:42.58ms
+[2025-09-04 12:13:44] [Rank 0] step:8781/10000 train_time:373933ms step_avg:42.58ms
+[2025-09-04 12:13:44] [Rank 0] step:8801/10000 train_time:374687ms step_avg:42.57ms
+[2025-09-04 12:13:44] [Rank 0] step:8801/10000 train_time:374687ms step_avg:42.57ms
+[2025-09-04 12:13:45] [Rank 0] step:8821/10000 train_time:375442ms step_avg:42.56ms
+[2025-09-04 12:13:45] [Rank 0] step:8821/10000 train_time:375442ms step_avg:42.56ms
+[2025-09-04 12:13:46] [Rank 0] step:8841/10000 train_time:376471ms step_avg:42.58ms
+[2025-09-04 12:13:46] [Rank 0] step:8841/10000 train_time:376471ms step_avg:42.58ms
+[2025-09-04 12:13:47] [Rank 0] step:8861/10000 train_time:377226ms step_avg:42.57ms
+[2025-09-04 12:13:47] [Rank 0] step:8861/10000 train_time:377226ms step_avg:42.57ms
+[2025-09-04 12:13:48] [Rank 0] step:8881/10000 train_time:377980ms step_avg:42.56ms
+[2025-09-04 12:13:48] [Rank 0] step:8881/10000 train_time:377980ms step_avg:42.56ms
+[2025-09-04 12:13:49] [Rank 0] step:8901/10000 train_time:378735ms step_avg:42.55ms
+[2025-09-04 12:13:49] [Rank 0] step:8901/10000 train_time:378735ms step_avg:42.55ms
+[2025-09-04 12:13:49] [Rank 0] step:8921/10000 train_time:379490ms step_avg:42.54ms
+[2025-09-04 12:13:49] [Rank 0] step:8921/10000 train_time:379490ms step_avg:42.54ms
+[2025-09-04 12:13:50] [Rank 0] step:8941/10000 train_time:380246ms step_avg:42.53ms
+[2025-09-04 12:13:50] [Rank 0] step:8941/10000 train_time:380246ms step_avg:42.53ms
+[2025-09-04 12:13:51] [Rank 0] step:8961/10000 train_time:381001ms step_avg:42.52ms
+[2025-09-04 12:13:51] [Rank 0] step:8961/10000 train_time:381001ms step_avg:42.52ms
+[2025-09-04 12:13:52] [Rank 0] step:8981/10000 train_time:381755ms step_avg:42.51ms
+[2025-09-04 12:13:52] [Rank 0] step:8981/10000 train_time:381755ms step_avg:42.51ms
+[2025-09-04 12:13:52] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:13:52] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:13:53] [Rank 0] PRINT: step:9000/10000 train_loss:0.6139 val_loss:0.6081 train_time:382515ms step_avg:42.50ms
+[2025-09-04 12:13:53] [Rank 0] PRINT: step:9000/10000 train_loss:0.6139 val_loss:0.6081 train_time:382515ms step_avg:42.50ms
+[2025-09-04 12:13:53] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:13:53] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:13:53] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:13:53] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:15:31] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:15:31] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:15:31] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:15:31] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:15:31] [Rank 0]   Total Loss: 5.1651
+[2025-09-04 12:15:31] [Rank 0]   Total Loss: 5.1651
+[2025-09-04 12:15:31] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 12:15:31] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 12:15:31] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 12:15:31] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 12:15:31] [Rank 0]   Group 0 Loss: 4.9938
+[2025-09-04 12:15:31] [Rank 0]   Group 0 Loss: 4.9938
+[2025-09-04 12:15:31] [Rank 0]   Group 1 Loss: 4.7902
+[2025-09-04 12:15:31] [Rank 0]   Group 1 Loss: 4.7902
+[2025-09-04 12:15:31] [Rank 0]   Group 2 Loss: 4.6022
+[2025-09-04 12:15:31] [Rank 0]   Group 2 Loss: 4.6022
+[2025-09-04 12:15:31] [Rank 0]   Group 3 Loss: 5.0396
+[2025-09-04 12:15:31] [Rank 0]   Group 3 Loss: 5.0396
+[2025-09-04 12:15:31] [Rank 0]   Group 4 Loss: 5.0650
+[2025-09-04 12:15:31] [Rank 0]   Group 4 Loss: 5.0650
+[2025-09-04 12:15:31] [Rank 0]   Group 5 Loss: 5.0799
+[2025-09-04 12:15:31] [Rank 0]   Group 5 Loss: 5.0799
+[2025-09-04 12:15:31] [Rank 0]   Group 6 Loss: 4.9809
+[2025-09-04 12:15:31] [Rank 0]   Group 6 Loss: 4.9809
+[2025-09-04 12:15:31] [Rank 0]   Group 7 Loss: 5.0993
+[2025-09-04 12:15:31] [Rank 0]   Group 7 Loss: 5.0993
+[2025-09-04 12:15:31] [Rank 0]   Group 8 Loss: 5.2645
+[2025-09-04 12:15:31] [Rank 0]   Group 8 Loss: 5.2645
+[2025-09-04 12:15:31] [Rank 0]   Group 9 Loss: 5.2097
+[2025-09-04 12:15:31] [Rank 0]   Group 9 Loss: 5.2097
+[2025-09-04 12:15:31] [Rank 0]   Group 10 Loss: 5.3816
+[2025-09-04 12:15:31] [Rank 0]   Group 10 Loss: 5.3816
+[2025-09-04 12:15:31] [Rank 0]   Group 11 Loss: 5.3968
+[2025-09-04 12:15:31] [Rank 0]   Group 11 Loss: 5.3968
+[2025-09-04 12:15:31] [Rank 0]   Group 12 Loss: 5.3425
+[2025-09-04 12:15:31] [Rank 0]   Group 12 Loss: 5.3425
+[2025-09-04 12:15:31] [Rank 0]   Group 13 Loss: 5.5558
+[2025-09-04 12:15:31] [Rank 0]   Group 13 Loss: 5.5558
+[2025-09-04 12:15:31] [Rank 0]   Group 14 Loss: 5.4330
+[2025-09-04 12:15:31] [Rank 0]   Group 14 Loss: 5.4330
+[2025-09-04 12:15:31] [Rank 0]   Group 15 Loss: 5.4065
+[2025-09-04 12:15:31] [Rank 0]   Group 15 Loss: 5.4065
+[2025-09-04 12:15:31] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:15:31] [Rank 0]   Group 15 FTA: 0.9300
+[2025-09-04 12:15:31] [Rank 0]   Group 15 FTA: 0.9300
+[2025-09-04 12:15:32] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:15:32] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:15:32] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:15:32] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:15:33] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:15:33] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:15:33] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:15:33] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:15:33] [Rank 0] step:9001/10000 train_time:382530ms step_avg:42.50ms
+[2025-09-04 12:15:33] [Rank 0] step:9001/10000 train_time:382530ms step_avg:42.50ms
+[2025-09-04 12:15:34] [Rank 0] step:9021/10000 train_time:383285ms step_avg:42.49ms
+[2025-09-04 12:15:34] [Rank 0] step:9021/10000 train_time:383285ms step_avg:42.49ms
+[2025-09-04 12:15:34] [Rank 0] step:9041/10000 train_time:384039ms step_avg:42.48ms
+[2025-09-04 12:15:34] [Rank 0] step:9041/10000 train_time:384039ms step_avg:42.48ms
+[2025-09-04 12:15:35] [Rank 0] step:9061/10000 train_time:384794ms step_avg:42.47ms
+[2025-09-04 12:15:35] [Rank 0] step:9061/10000 train_time:384794ms step_avg:42.47ms
+[2025-09-04 12:15:36] [Rank 0] step:9081/10000 train_time:385548ms step_avg:42.46ms
+[2025-09-04 12:15:36] [Rank 0] step:9081/10000 train_time:385548ms step_avg:42.46ms
+[2025-09-04 12:15:37] [Rank 0] step:9101/10000 train_time:386302ms step_avg:42.45ms
+[2025-09-04 12:15:37] [Rank 0] step:9101/10000 train_time:386302ms step_avg:42.45ms
+[2025-09-04 12:15:37] [Rank 0] step:9121/10000 train_time:387057ms step_avg:42.44ms
+[2025-09-04 12:15:37] [Rank 0] step:9121/10000 train_time:387057ms step_avg:42.44ms
+[2025-09-04 12:15:38] [Rank 0] step:9141/10000 train_time:387812ms step_avg:42.43ms
+[2025-09-04 12:15:38] [Rank 0] step:9141/10000 train_time:387812ms step_avg:42.43ms
+[2025-09-04 12:15:39] [Rank 0] step:9161/10000 train_time:388566ms step_avg:42.42ms
+[2025-09-04 12:15:39] [Rank 0] step:9161/10000 train_time:388566ms step_avg:42.42ms
+[2025-09-04 12:15:40] [Rank 0] step:9181/10000 train_time:389321ms step_avg:42.41ms
+[2025-09-04 12:15:40] [Rank 0] step:9181/10000 train_time:389321ms step_avg:42.41ms
+[2025-09-04 12:15:40] [Rank 0] step:9201/10000 train_time:390075ms step_avg:42.39ms
+[2025-09-04 12:15:40] [Rank 0] step:9201/10000 train_time:390075ms step_avg:42.39ms
+[2025-09-04 12:15:41] [Rank 0] step:9221/10000 train_time:390830ms step_avg:42.38ms
+[2025-09-04 12:15:41] [Rank 0] step:9221/10000 train_time:390830ms step_avg:42.38ms
+[2025-09-04 12:15:42] [Rank 0] step:9241/10000 train_time:391584ms step_avg:42.37ms
+[2025-09-04 12:15:42] [Rank 0] step:9241/10000 train_time:391584ms step_avg:42.37ms
+[2025-09-04 12:15:43] [Rank 0] step:9261/10000 train_time:392339ms step_avg:42.36ms
+[2025-09-04 12:15:43] [Rank 0] step:9261/10000 train_time:392339ms step_avg:42.36ms
+[2025-09-04 12:15:44] [Rank 0] step:9281/10000 train_time:393095ms step_avg:42.35ms
+[2025-09-04 12:15:44] [Rank 0] step:9281/10000 train_time:393095ms step_avg:42.35ms
+[2025-09-04 12:15:44] [Rank 0] step:9301/10000 train_time:393850ms step_avg:42.34ms
+[2025-09-04 12:15:44] [Rank 0] step:9301/10000 train_time:393850ms step_avg:42.34ms
+[2025-09-04 12:15:45] [Rank 0] step:9321/10000 train_time:394606ms step_avg:42.34ms
+[2025-09-04 12:15:45] [Rank 0] step:9321/10000 train_time:394606ms step_avg:42.34ms
+[2025-09-04 12:15:46] [Rank 0] step:9341/10000 train_time:395361ms step_avg:42.33ms
+[2025-09-04 12:15:46] [Rank 0] step:9341/10000 train_time:395361ms step_avg:42.33ms
+[2025-09-04 12:15:47] [Rank 0] step:9361/10000 train_time:396116ms step_avg:42.32ms
+[2025-09-04 12:15:47] [Rank 0] step:9361/10000 train_time:396116ms step_avg:42.32ms
+[2025-09-04 12:15:47] [Rank 0] step:9381/10000 train_time:396872ms step_avg:42.31ms
+[2025-09-04 12:15:47] [Rank 0] step:9381/10000 train_time:396872ms step_avg:42.31ms
+[2025-09-04 12:15:48] [Rank 0] step:9401/10000 train_time:397627ms step_avg:42.30ms
+[2025-09-04 12:15:48] [Rank 0] step:9401/10000 train_time:397627ms step_avg:42.30ms
+[2025-09-04 12:15:49] [Rank 0] step:9421/10000 train_time:398383ms step_avg:42.29ms
+[2025-09-04 12:15:49] [Rank 0] step:9421/10000 train_time:398383ms step_avg:42.29ms
+[2025-09-04 12:15:50] [Rank 0] step:9441/10000 train_time:399139ms step_avg:42.28ms
+[2025-09-04 12:15:50] [Rank 0] step:9441/10000 train_time:399139ms step_avg:42.28ms
+[2025-09-04 12:15:50] [Rank 0] step:9461/10000 train_time:399894ms step_avg:42.27ms
+[2025-09-04 12:15:50] [Rank 0] step:9461/10000 train_time:399894ms step_avg:42.27ms
+[2025-09-04 12:15:51] [Rank 0] step:9481/10000 train_time:400649ms step_avg:42.26ms
+[2025-09-04 12:15:51] [Rank 0] step:9481/10000 train_time:400649ms step_avg:42.26ms
+[2025-09-04 12:15:52] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:15:52] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:15:52] [Rank 0] PRINT: step:9500/10000 train_loss:0.6112 val_loss:0.6065 train_time:401410ms step_avg:42.25ms
+[2025-09-04 12:15:52] [Rank 0] PRINT: step:9500/10000 train_loss:0.6112 val_loss:0.6065 train_time:401410ms step_avg:42.25ms
+[2025-09-04 12:15:52] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:15:52] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:15:52] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:15:52] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:17:30] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:17:30] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:17:30] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:17:30] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:17:30] [Rank 0]   Total Loss: 5.1714
+[2025-09-04 12:17:30] [Rank 0]   Total Loss: 5.1714
+[2025-09-04 12:17:30] [Rank 0]   Total FTA (Unweighted): 0.9962
+[2025-09-04 12:17:30] [Rank 0]   Total FTA (Unweighted): 0.9962
+[2025-09-04 12:17:30] [Rank 0]   Total FTA (Weighted):   0.9962
+[2025-09-04 12:17:30] [Rank 0]   Total FTA (Weighted):   0.9962
+[2025-09-04 12:17:30] [Rank 0]   Group 0 Loss: 5.0161
+[2025-09-04 12:17:30] [Rank 0]   Group 0 Loss: 5.0161
+[2025-09-04 12:17:30] [Rank 0]   Group 1 Loss: 4.7588
+[2025-09-04 12:17:30] [Rank 0]   Group 1 Loss: 4.7588
+[2025-09-04 12:17:30] [Rank 0]   Group 2 Loss: 4.6027
+[2025-09-04 12:17:30] [Rank 0]   Group 2 Loss: 4.6027
+[2025-09-04 12:17:30] [Rank 0]   Group 3 Loss: 5.0574
+[2025-09-04 12:17:30] [Rank 0]   Group 3 Loss: 5.0574
+[2025-09-04 12:17:30] [Rank 0]   Group 4 Loss: 5.0800
+[2025-09-04 12:17:30] [Rank 0]   Group 4 Loss: 5.0800
+[2025-09-04 12:17:30] [Rank 0]   Group 5 Loss: 5.0776
+[2025-09-04 12:17:30] [Rank 0]   Group 5 Loss: 5.0776
+[2025-09-04 12:17:30] [Rank 0]   Group 6 Loss: 4.9877
+[2025-09-04 12:17:30] [Rank 0]   Group 6 Loss: 4.9877
+[2025-09-04 12:17:30] [Rank 0]   Group 7 Loss: 5.1101
+[2025-09-04 12:17:30] [Rank 0]   Group 7 Loss: 5.1101
+[2025-09-04 12:17:30] [Rank 0]   Group 8 Loss: 5.2732
+[2025-09-04 12:17:30] [Rank 0]   Group 8 Loss: 5.2732
+[2025-09-04 12:17:30] [Rank 0]   Group 9 Loss: 5.2298
+[2025-09-04 12:17:30] [Rank 0]   Group 9 Loss: 5.2298
+[2025-09-04 12:17:30] [Rank 0]   Group 10 Loss: 5.3993
+[2025-09-04 12:17:30] [Rank 0]   Group 10 Loss: 5.3993
+[2025-09-04 12:17:30] [Rank 0]   Group 11 Loss: 5.3998
+[2025-09-04 12:17:30] [Rank 0]   Group 11 Loss: 5.3998
+[2025-09-04 12:17:30] [Rank 0]   Group 12 Loss: 5.3398
+[2025-09-04 12:17:30] [Rank 0]   Group 12 Loss: 5.3398
+[2025-09-04 12:17:30] [Rank 0]   Group 13 Loss: 5.5495
+[2025-09-04 12:17:30] [Rank 0]   Group 13 Loss: 5.5495
+[2025-09-04 12:17:30] [Rank 0]   Group 14 Loss: 5.4362
+[2025-09-04 12:17:30] [Rank 0]   Group 14 Loss: 5.4362
+[2025-09-04 12:17:30] [Rank 0]   Group 15 Loss: 5.4239
+[2025-09-04 12:17:30] [Rank 0]   Group 15 Loss: 5.4239
+[2025-09-04 12:17:30] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:17:30] [Rank 0]   Group 15 FTA: 0.9400
+[2025-09-04 12:17:30] [Rank 0]   Group 15 FTA: 0.9400
+[2025-09-04 12:17:31] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:17:31] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:17:31] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:17:31] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:17:31] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:17:31] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:17:32] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:17:32] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:17:32] [Rank 0] step:9501/10000 train_time:401427ms step_avg:42.25ms
+[2025-09-04 12:17:32] [Rank 0] step:9501/10000 train_time:401427ms step_avg:42.25ms
+[2025-09-04 12:17:32] [Rank 0] step:9521/10000 train_time:402196ms step_avg:42.24ms
+[2025-09-04 12:17:32] [Rank 0] step:9521/10000 train_time:402196ms step_avg:42.24ms
+[2025-09-04 12:17:33] [Rank 0] step:9541/10000 train_time:402950ms step_avg:42.23ms
+[2025-09-04 12:17:33] [Rank 0] step:9541/10000 train_time:402950ms step_avg:42.23ms
+[2025-09-04 12:17:34] [Rank 0] step:9561/10000 train_time:403705ms step_avg:42.22ms
+[2025-09-04 12:17:34] [Rank 0] step:9561/10000 train_time:403705ms step_avg:42.22ms
+[2025-09-04 12:17:35] [Rank 0] step:9581/10000 train_time:404459ms step_avg:42.21ms
+[2025-09-04 12:17:35] [Rank 0] step:9581/10000 train_time:404459ms step_avg:42.21ms
+[2025-09-04 12:17:35] [Rank 0] step:9601/10000 train_time:405214ms step_avg:42.21ms
+[2025-09-04 12:17:35] [Rank 0] step:9601/10000 train_time:405214ms step_avg:42.21ms
+[2025-09-04 12:17:36] [Rank 0] step:9621/10000 train_time:405969ms step_avg:42.20ms
+[2025-09-04 12:17:36] [Rank 0] step:9621/10000 train_time:405969ms step_avg:42.20ms
+[2025-09-04 12:17:37] [Rank 0] step:9641/10000 train_time:406724ms step_avg:42.19ms
+[2025-09-04 12:17:37] [Rank 0] step:9641/10000 train_time:406724ms step_avg:42.19ms
+[2025-09-04 12:17:38] [Rank 0] step:9661/10000 train_time:407758ms step_avg:42.21ms
+[2025-09-04 12:17:38] [Rank 0] step:9661/10000 train_time:407758ms step_avg:42.21ms
+[2025-09-04 12:17:39] [Rank 0] step:9681/10000 train_time:408513ms step_avg:42.20ms
+[2025-09-04 12:17:39] [Rank 0] step:9681/10000 train_time:408513ms step_avg:42.20ms
+[2025-09-04 12:17:40] [Rank 0] step:9701/10000 train_time:409268ms step_avg:42.19ms
+[2025-09-04 12:17:40] [Rank 0] step:9701/10000 train_time:409268ms step_avg:42.19ms
+[2025-09-04 12:17:40] [Rank 0] step:9721/10000 train_time:410022ms step_avg:42.18ms
+[2025-09-04 12:17:40] [Rank 0] step:9721/10000 train_time:410022ms step_avg:42.18ms
+[2025-09-04 12:17:41] [Rank 0] step:9741/10000 train_time:410777ms step_avg:42.17ms
+[2025-09-04 12:17:41] [Rank 0] step:9741/10000 train_time:410777ms step_avg:42.17ms
+[2025-09-04 12:17:42] [Rank 0] step:9761/10000 train_time:411532ms step_avg:42.16ms
+[2025-09-04 12:17:42] [Rank 0] step:9761/10000 train_time:411532ms step_avg:42.16ms
+[2025-09-04 12:17:43] [Rank 0] step:9781/10000 train_time:412287ms step_avg:42.15ms
+[2025-09-04 12:17:43] [Rank 0] step:9781/10000 train_time:412287ms step_avg:42.15ms
+[2025-09-04 12:17:43] [Rank 0] step:9801/10000 train_time:413043ms step_avg:42.14ms
+[2025-09-04 12:17:43] [Rank 0] step:9801/10000 train_time:413043ms step_avg:42.14ms
+[2025-09-04 12:17:44] [Rank 0] step:9821/10000 train_time:413798ms step_avg:42.13ms
+[2025-09-04 12:17:44] [Rank 0] step:9821/10000 train_time:413798ms step_avg:42.13ms
+[2025-09-04 12:17:45] [Rank 0] step:9841/10000 train_time:414553ms step_avg:42.13ms
+[2025-09-04 12:17:45] [Rank 0] step:9841/10000 train_time:414553ms step_avg:42.13ms
+[2025-09-04 12:17:46] [Rank 0] step:9861/10000 train_time:415309ms step_avg:42.12ms
+[2025-09-04 12:17:46] [Rank 0] step:9861/10000 train_time:415309ms step_avg:42.12ms
+[2025-09-04 12:17:46] [Rank 0] step:9881/10000 train_time:416064ms step_avg:42.11ms
+[2025-09-04 12:17:46] [Rank 0] step:9881/10000 train_time:416064ms step_avg:42.11ms
+[2025-09-04 12:17:47] [Rank 0] step:9901/10000 train_time:416819ms step_avg:42.10ms
+[2025-09-04 12:17:47] [Rank 0] step:9901/10000 train_time:416819ms step_avg:42.10ms
+[2025-09-04 12:17:48] [Rank 0] step:9921/10000 train_time:417573ms step_avg:42.09ms
+[2025-09-04 12:17:48] [Rank 0] step:9921/10000 train_time:417573ms step_avg:42.09ms
+[2025-09-04 12:17:49] [Rank 0] step:9941/10000 train_time:418328ms step_avg:42.08ms
+[2025-09-04 12:17:49] [Rank 0] step:9941/10000 train_time:418328ms step_avg:42.08ms
+[2025-09-04 12:17:49] [Rank 0] step:9961/10000 train_time:419084ms step_avg:42.07ms
+[2025-09-04 12:17:49] [Rank 0] step:9961/10000 train_time:419084ms step_avg:42.07ms
+[2025-09-04 12:17:50] [Rank 0] step:9981/10000 train_time:419838ms step_avg:42.06ms
+[2025-09-04 12:17:50] [Rank 0] step:9981/10000 train_time:419838ms step_avg:42.06ms
+[2025-09-04 12:17:51] [Rank 0] step:10000/10000 train_time:420555ms step_avg:42.06ms
+[2025-09-04 12:17:51] [Rank 0] step:10000/10000 train_time:420555ms step_avg:42.06ms
+[2025-09-04 12:17:51] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:17:51] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:17:51] [Rank 0] PRINT: step:10000/10000 train_loss:0.6090 val_loss:0.6054 train_time:420605ms step_avg:42.06ms
+[2025-09-04 12:17:51] [Rank 0] PRINT: step:10000/10000 train_loss:0.6090 val_loss:0.6054 train_time:420605ms step_avg:42.06ms
+[2025-09-04 12:17:51] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:17:51] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:17:51] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:17:51] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:19:29] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:19:29] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:19:29] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:19:29] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:19:29] [Rank 0]   Total Loss: 5.2099
+[2025-09-04 12:19:29] [Rank 0]   Total Loss: 5.2099
+[2025-09-04 12:19:29] [Rank 0]   Total FTA (Unweighted): 0.9981
+[2025-09-04 12:19:29] [Rank 0]   Total FTA (Unweighted): 0.9981
+[2025-09-04 12:19:29] [Rank 0]   Total FTA (Weighted):   0.9981
+[2025-09-04 12:19:29] [Rank 0]   Total FTA (Weighted):   0.9981
+[2025-09-04 12:19:29] [Rank 0]   Group 0 Loss: 5.0379
+[2025-09-04 12:19:29] [Rank 0]   Group 0 Loss: 5.0379
+[2025-09-04 12:19:29] [Rank 0]   Group 1 Loss: 4.7919
+[2025-09-04 12:19:29] [Rank 0]   Group 1 Loss: 4.7919
+[2025-09-04 12:19:29] [Rank 0]   Group 2 Loss: 4.6145
+[2025-09-04 12:19:29] [Rank 0]   Group 2 Loss: 4.6145
+[2025-09-04 12:19:29] [Rank 0]   Group 3 Loss: 5.0761
+[2025-09-04 12:19:29] [Rank 0]   Group 3 Loss: 5.0761
+[2025-09-04 12:19:29] [Rank 0]   Group 4 Loss: 5.1445
+[2025-09-04 12:19:29] [Rank 0]   Group 4 Loss: 5.1445
+[2025-09-04 12:19:29] [Rank 0]   Group 5 Loss: 5.1112
+[2025-09-04 12:19:29] [Rank 0]   Group 5 Loss: 5.1112
+[2025-09-04 12:19:29] [Rank 0]   Group 6 Loss: 5.0264
+[2025-09-04 12:19:29] [Rank 0]   Group 6 Loss: 5.0264
+[2025-09-04 12:19:29] [Rank 0]   Group 7 Loss: 5.1369
+[2025-09-04 12:19:29] [Rank 0]   Group 7 Loss: 5.1369
+[2025-09-04 12:19:29] [Rank 0]   Group 8 Loss: 5.3079
+[2025-09-04 12:19:29] [Rank 0]   Group 8 Loss: 5.3079
+[2025-09-04 12:19:29] [Rank 0]   Group 9 Loss: 5.2522
+[2025-09-04 12:19:29] [Rank 0]   Group 9 Loss: 5.2522
+[2025-09-04 12:19:29] [Rank 0]   Group 10 Loss: 5.4655
+[2025-09-04 12:19:29] [Rank 0]   Group 10 Loss: 5.4655
+[2025-09-04 12:19:29] [Rank 0]   Group 11 Loss: 5.4594
+[2025-09-04 12:19:29] [Rank 0]   Group 11 Loss: 5.4594
+[2025-09-04 12:19:29] [Rank 0]   Group 12 Loss: 5.3660
+[2025-09-04 12:19:29] [Rank 0]   Group 12 Loss: 5.3660
+[2025-09-04 12:19:29] [Rank 0]   Group 13 Loss: 5.5968
+[2025-09-04 12:19:29] [Rank 0]   Group 13 Loss: 5.5968
+[2025-09-04 12:19:29] [Rank 0]   Group 14 Loss: 5.5036
+[2025-09-04 12:19:29] [Rank 0]   Group 14 Loss: 5.5036
+[2025-09-04 12:19:29] [Rank 0]   Group 15 Loss: 5.4669
+[2025-09-04 12:19:29] [Rank 0]   Group 15 Loss: 5.4669
+[2025-09-04 12:19:29] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:19:29] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 12:19:29] [Rank 0]   Group 14 FTA: 0.9900
+[2025-09-04 12:19:29] [Rank 0]   Group 15 FTA: 0.9800
+[2025-09-04 12:19:29] [Rank 0]   Group 15 FTA: 0.9800
+[2025-09-04 12:19:30] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:19:30] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_loss_curves.png
+[2025-09-04 12:19:30] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:19:30] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/per_class_acc_curves.png
+[2025-09-04 12:19:30] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:19:30] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_loss_curve.png
+[2025-09-04 12:19:31] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:19:31] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_44/total_acc_curve.png
+[2025-09-04 12:19:31] [Rank 0] step:10001/10000 train_time:420622ms step_avg:42.06ms
+[2025-09-04 12:19:31] [Rank 0] step:10001/10000 train_time:420622ms step_avg:42.06ms
+[2025-09-04 12:19:31] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 12:19:31 2025 ---
+[2025-09-04 12:19:31] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 12:19:31 2025 ---
+[2025-09-04 12:19:31] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB
+[2025-09-04 12:19:31] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/config.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..1f6eb333994305398667d7ed85d350c39cc4b3e2
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/config.json
@@ -0,0 +1,29 @@
+{
+    "cli_args": {
+        "unet": false,
+        "seed": 45,
+        "optimizer_mode": 10,
+        "model_parameterization": "qkvo",
+        "per_group_k": 100,
+        "muon_lr": 0.002,
+        "adam_lr": 0.002,
+        "base_dir": "logs_qa_muon/diff_modes",
+        "sgd_lr": 0.01,
+        "m_val": 15,
+        "qa_jsonl_path": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl"
+    },
+    "hyperparameters": {
+        "train_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin",
+        "val_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin",
+        "val_tokens": 491520,
+        "train_seq_len": 3072,
+        "val_seq_len": 16384,
+        "num_iterations": 10000,
+        "cooldown_frac": 0.8,
+        "vocab_size": 50257,
+        "val_loss_every": 500,
+        "save_checkpoint": false
+    },
+    "run_uuid_for_log": "6c733232-2eb7-41a9-b008-521a86795de0",
+    "script_code_logged_at_start": true
+}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/fixed_eval_indices.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/fixed_eval_indices.json
new file mode 100644
index 0000000000000000000000000000000000000000..a823775225c5e592eb10700e5e0319b0491b1eb6
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/fixed_eval_indices.json
@@ -0,0 +1 @@
+{"1": [1238956, 182074, 1437575, 1061037, 383150, 1176376, 926, 823011, 832520, 1266421, 512738, 144357, 848076, 890204, 213997, 95146, 261767, 467731, 832231, 217985, 913168, 107253, 1361828, 61314, 1230420, 1133619, 146690, 429587, 419151, 58695, 1579770, 503799, 1421284, 882534, 1022637, 785343, 1154604, 67783, 1325109, 243941, 1213240, 438111, 460295, 269373, 538055, 1347006, 71775, 255496, 299906, 1227973, 815402, 190082, 1304077, 1023347, 613801, 983830, 1284420, 389321, 1625224, 717538, 1172273, 992184, 1181312, 1014039, 885952, 1538489, 158933, 1667270, 1250445, 958097, 1458224, 1306495, 62945, 733843, 1360200, 540493, 762461, 501460, 1208142, 1180559, 1333588, 690481, 355756, 618511, 733586, 650301, 799437, 165533, 1238977, 323078, 1485080, 609610, 1212241, 606952, 1253407, 1420922, 327112, 701, 777907, 1626516], "0": [1390189, 1220977, 1312259, 1201125, 1235379, 1272843, 344142, 1119560, 856330, 766456, 1145928, 550624, 684681, 1309879, 54620, 1126124, 1278077, 910610, 1532044, 73645, 460165, 462723, 457959, 476877, 320873, 1408316, 989156, 1276325, 338260, 577743, 1136972, 25115, 211215, 1296818, 968705, 436781, 318224, 385325, 1516441, 533464, 1628693, 703399, 579670, 1518978, 305154, 1625960, 1400284, 713841, 1652150, 811009, 336920, 881015, 512030, 1347531, 1005706, 412167, 173312, 1136816, 343017, 537797, 1489267, 935475, 703854, 1570584, 1504269, 1458356, 1310700, 1242505, 509976, 22096, 788120, 1616850, 1591198, 1151224, 797820, 823995, 1153364, 1060220, 727674, 580729, 148912, 334290, 686098, 1633151, 1130523, 1012440, 1374480, 953410, 1381730, 1376118, 932096, 739115, 739014, 263875, 1400308, 556177, 1341771, 348626, 575350, 254846], "10": [748592, 515936, 1645691, 329828, 1104748, 1453439, 406461, 1557507, 1146644, 1398421, 163667, 228310, 407739, 585218, 785725, 1371077, 134834, 15431, 1303784, 273099, 1530332, 1320219, 1329247, 753603, 664548, 501873, 1512467, 583734, 1355643, 600440, 428998, 1426876, 121341, 1456507, 892281, 1242339, 1304508, 968336, 1593981, 634471, 1253799, 593156, 1486078, 712425, 441285, 618217, 1024661, 1395408, 1423117, 1463918, 121752, 800727, 52402, 1185143, 1651644, 1583572, 1302889, 1040220, 1234122, 333457, 1552652, 1567291, 1405501, 222708, 1469496, 875611, 894032, 783951, 934281, 109533, 198744, 643064, 197926, 224405, 1169998, 956511, 1096712, 1213224, 828780, 204659, 564225, 257079, 754287, 1598723, 821179, 1156648, 60674, 290993, 1215838, 1556010, 1025931, 687671, 252775, 771448, 574110, 578615, 279536, 1503737, 1089787, 2338], "14": [1602125, 1388977, 709154, 666138, 1446614, 572103, 99632, 25794, 1487658, 15077, 1600205, 682155, 679850, 1462049, 970283, 81336, 378196, 1627374, 832926, 1217832, 740140, 1336941, 793290, 1441127, 297494, 303372, 339524, 1363322, 525184, 593339, 61055, 1468219, 1110310, 1061855, 1046762, 410655, 449257, 530144, 1362373, 241552, 926068, 1618007, 629406, 327589, 263223, 1525290, 621372, 606355, 534601, 126046, 934024, 1181343, 1234759, 295773, 838035, 1447427, 1395837, 1568116, 486923, 465204, 98192, 1145989, 271529, 720487, 1091877, 653313, 872526, 13890, 111971, 671086, 301528, 50698, 455373, 786098, 329199, 599178, 644737, 1404606, 33829, 279256, 161938, 309352, 1537653, 21723, 309413, 478184, 340938, 1080124, 1375458, 1301528, 331423, 212589, 1220931, 112707, 454022, 1466562, 1238127, 341561, 1497264, 247245], "3": [143966, 715765, 455823, 1414629, 814361, 278667, 700211, 1286614, 1601013, 923359, 1105960, 549680, 371512, 732263, 1236795, 432942, 1379421, 488467, 679369, 1200930, 479010, 683138, 700149, 1422714, 348019, 1263383, 1044275, 820031, 969830, 75153, 1410929, 394116, 1486865, 299414, 264214, 1416493, 428688, 107076, 753671, 674417, 1456072, 307080, 1259105, 1563344, 1198484, 456551, 1061513, 1128111, 922541, 1040277, 1360208, 1411429, 259514, 883801, 716781, 1004987, 285196, 266162, 781115, 385241, 252691, 1398963, 1440582, 1300653, 756663, 231617, 1319447, 42576, 1434337, 20658, 581284, 178420, 1626533, 694079, 1531414, 1631177, 1478512, 647215, 706568, 303323, 10540, 1010611, 1150313, 159016, 1614581, 1257722, 470491, 1333309, 94699, 756222, 1456446, 296132, 1186414, 1591257, 1155050, 789439, 1472373, 1403759, 1451276, 745193], "4": [1227524, 563788, 436964, 1380176, 1648232, 267090, 273744, 1031517, 1580586, 416846, 1309866, 114170, 1649739, 1161604, 690741, 1432068, 224570, 322835, 1559902, 479718, 709757, 410142, 1668701, 1143278, 968258, 406289, 884233, 1080971, 269480, 631196, 421463, 374502, 430292, 1309826, 1422997, 1439354, 131354, 975504, 84660, 1590390, 288064, 1419944, 695652, 320713, 260376, 222416, 937739, 1104508, 1323311, 962301, 1060820, 1117857, 280792, 692359, 656651, 358878, 828406, 598701, 127434, 1511259, 1332605, 1178325, 1281187, 203589, 792701, 684176, 543034, 980077, 982509, 640126, 1252418, 1529627, 519645, 359068, 992779, 550843, 1493488, 447632, 1512599, 361760, 626672, 89042, 604248, 574827, 87735, 1432602, 473061, 668609, 925422, 311417, 71495, 682067, 221270, 600053, 673169, 1408933, 719302, 788676, 998101, 243902], "15": [1444113, 1102795, 1149896, 1181364, 349047, 1076908, 816006, 945310, 197585, 276574, 946546, 1505399, 740223, 1154813, 1312891, 1261018, 249986, 1504592, 647427, 915615, 1245136, 1288301, 802239, 410533, 805664, 1013403, 498363, 777073, 1508176, 18111, 1147340, 880289, 215773, 924760, 1301970, 318135, 1524489, 103475, 872748, 1568512, 722849, 1364157, 838956, 1451408, 970208, 1350756, 1498244, 980461, 1098677, 1141812, 463859, 95672, 1220365, 554451, 524014, 1161048, 17478, 113548, 1027741, 1657643, 1462809, 1279104, 85588, 241936, 510707, 151817, 1191227, 574942, 735662, 1046248, 1056729, 96279, 781378, 1042711, 1602339, 1503514, 299087, 898968, 553537, 293167, 897565, 572125, 425215, 916907, 1210531, 1013275, 201532, 883447, 1641610, 1185682, 425791, 831133, 1291227, 1099064, 934983, 588125, 1284591, 1163873, 524656, 1230762], "13": [1122510, 535854, 845319, 585144, 884184, 441130, 934654, 1046629, 1512962, 1190830, 1462685, 90530, 894615, 1527783, 415323, 90852, 1556905, 1002406, 366356, 1386022, 846952, 928360, 3138, 406205, 78588, 1208032, 1210206, 257177, 137047, 711380, 693628, 105893, 392053, 1301525, 677934, 1549630, 1134520, 447564, 1194061, 672969, 740755, 1045536, 633455, 1336343, 613147, 1161931, 234095, 1246643, 643373, 834221, 1351442, 1502069, 1401555, 759927, 1320337, 666063, 291222, 956633, 400583, 963313, 683474, 534882, 1101312, 1519551, 1513194, 752124, 1028953, 1383362, 169732, 785294, 160568, 1053580, 774534, 372048, 292049, 791470, 1318537, 141966, 510573, 1440830, 930544, 1352026, 463299, 1245306, 159755, 1566413, 316835, 1135200, 266962, 112892, 253603, 1328380, 703630, 21798, 268463, 7112, 1232576, 374978, 240715, 1299198], "2": [463910, 845243, 985993, 402000, 962033, 473711, 108588, 1074932, 539045, 1162793, 1086008, 778638, 751876, 53650, 465364, 803834, 820717, 9490, 473708, 26912, 970677, 1441139, 231484, 749139, 286603, 1346834, 1641988, 699767, 816768, 1220504, 978431, 1331006, 611334, 829473, 271532, 566760, 332292, 1326050, 887731, 597530, 940973, 491159, 1150471, 122010, 225465, 382413, 1207486, 1568920, 992793, 896409, 1479386, 1539347, 759024, 293232, 1345256, 1282360, 956014, 371580, 1202087, 1638038, 892501, 38073, 1651323, 1103388, 1315239, 79005, 1045045, 1230831, 1206253, 1376976, 140037, 320204, 274836, 663525, 397448, 666367, 1501713, 570453, 836424, 694423, 157979, 279074, 1235605, 536170, 301830, 778556, 1069897, 181207, 1445408, 1082031, 969552, 1083538, 1395871, 168187, 5476, 514815, 194418, 1038889, 1288748, 1535767], "8": [258910, 518286, 80108, 811775, 937854, 1623550, 1592007, 661848, 1484674, 300, 104226, 1075215, 1190104, 8885, 1062956, 88002, 1020623, 189243, 1265153, 1472857, 301412, 21371, 441401, 1078954, 1648642, 894192, 1083738, 1182805, 1347655, 316568, 1063971, 255177, 186613, 550617, 102173, 563694, 465966, 965224, 372481, 929520, 1423441, 6575, 862309, 144894, 891435, 123628, 653301, 1378790, 1247156, 1129087, 1381991, 410876, 391895, 202136, 868149, 233421, 564464, 1174972, 1385073, 238326, 483488, 507729, 1329666, 955266, 479623, 902401, 1343890, 827390, 371073, 1480060, 1241650, 1670957, 644144, 1371709, 1324243, 670126, 23715, 1566067, 70858, 522312, 558100, 301714, 359605, 1401913, 1060057, 1027855, 987700, 720084, 324937, 562646, 390910, 1398541, 1097298, 1626769, 1461630, 1430950, 650075, 1025475, 274735, 650932], "7": [181761, 629797, 231148, 365747, 1229714, 500408, 198146, 500703, 1250728, 1578820, 753672, 84476, 1379367, 355424, 1591855, 1400949, 1194404, 912060, 884197, 320928, 1096692, 714201, 331003, 1345763, 1620998, 457231, 1542074, 1108477, 1602113, 641593, 542514, 159345, 360278, 35579, 506, 470465, 1606024, 1349783, 731258, 382125, 883800, 1652746, 382071, 99241, 20216, 1398849, 245155, 1601928, 1544592, 1632383, 1597152, 1266758, 1032571, 88287, 521379, 1423306, 579207, 1152743, 271926, 418168, 8827, 1358019, 429579, 380995, 1040420, 1067138, 283335, 1150469, 280745, 1184794, 468245, 948227, 357499, 527525, 1067827, 816862, 360563, 433490, 258517, 581805, 1089641, 1537081, 473494, 1480784, 60701, 686908, 614323, 143069, 738219, 1234467, 927103, 64088, 658420, 1353941, 1517049, 753334, 198407, 434613, 850731, 1211609], "5": [1250448, 711876, 285881, 342500, 987883, 195047, 810508, 356140, 999602, 284542, 730936, 736191, 661291, 11822, 551928, 617679, 384856, 414906, 238071, 473605, 868498, 1269087, 558407, 245083, 810169, 1456559, 1657433, 70018, 865238, 416303, 1389734, 1561961, 534163, 1098802, 1494214, 1285484, 599936, 296819, 696611, 307176, 29697, 731071, 257656, 1465441, 1668863, 166995, 1668377, 1457645, 1400335, 1525120, 33323, 437285, 459069, 620176, 231704, 846424, 1391144, 364636, 63552, 1093702, 67741, 285428, 454922, 719128, 1393391, 87651, 1418561, 1312176, 652985, 1446762, 71280, 1594253, 89848, 628629, 1117344, 882598, 1014283, 1601256, 739004, 508965, 1273914, 44245, 1267971, 1430470, 611385, 930900, 1439585, 1545682, 1300966, 1331569, 364367, 27900, 107616, 1579070, 94723, 818780, 1183546, 384575, 142627, 431847], "11": [12772, 1091320, 1136723, 560663, 715094, 1097800, 1016547, 1299126, 961155, 725788, 999768, 1277062, 650665, 1323964, 1024796, 1255124, 215111, 268904, 442949, 358645, 621958, 1666210, 1154349, 375517, 614870, 1166618, 1418680, 795635, 184985, 129300, 129366, 1343289, 1151524, 193531, 634856, 1618906, 43721, 1601844, 591772, 1322095, 1522284, 431856, 574972, 1544663, 1389257, 1113893, 1515857, 614345, 809163, 132406, 95327, 997226, 423104, 910705, 261584, 765030, 1216397, 439297, 1565563, 1595785, 1622521, 975588, 1314965, 1205327, 1501612, 1642358, 656493, 1369509, 195325, 1554913, 1661675, 790468, 511387, 1227059, 1219143, 1118001, 788451, 697651, 1575073, 1345104, 167862, 1409843, 138942, 984075, 1382364, 1299538, 214601, 1541475, 459076, 353219, 1310728, 757670, 904351, 334019, 1003259, 670104, 10386, 13556, 1097931, 1526407], "6": [295849, 644897, 1207883, 1648549, 1478750, 1486750, 1156499, 621047, 1387936, 662848, 1362507, 891859, 1188350, 1298200, 1044010, 767288, 1191230, 683760, 533545, 498960, 1421819, 366846, 54389, 1382045, 1002942, 1285004, 457209, 1074763, 823805, 1508295, 719970, 837720, 19881, 1189285, 1226048, 1078656, 1433699, 14805, 254682, 614235, 805087, 900964, 343832, 561709, 243288, 465867, 497085, 1153456, 730115, 42050, 878511, 1382508, 667820, 712856, 715963, 1499782, 1516002, 444399, 726168, 1204910, 591639, 986194, 107160, 1596752, 1239812, 591854, 536489, 1347905, 1597670, 122237, 620494, 354940, 1561646, 1352115, 74785, 163277, 590349, 1429584, 1440950, 848184, 1580723, 1631128, 979479, 758861, 84655, 1079672, 673692, 1315324, 167985, 1353485, 21815, 1175739, 1360408, 1062086, 1369999, 733863, 1371964, 61279, 1152056, 220910], "9": [780278, 1156801, 827882, 1095960, 874763, 1330006, 979412, 1154745, 1399441, 515095, 1505401, 1454104, 1563117, 1316054, 57098, 43367, 1175749, 1257252, 865038, 624410, 186727, 907758, 126168, 1353431, 167806, 1021404, 1585615, 846939, 552247, 1400471, 68525, 190495, 1556857, 181528, 47677, 136468, 189446, 269049, 262467, 503841, 1367168, 382007, 486946, 621235, 622683, 469399, 1537192, 707594, 1003091, 56021, 1354302, 799290, 1618221, 696794, 836658, 1591562, 251096, 1321886, 412122, 1352217, 946767, 1267324, 213504, 747174, 1403653, 1589967, 849743, 549269, 663800, 1315189, 451416, 1330881, 1352197, 1517840, 827169, 1228725, 1097725, 955944, 1375509, 1590213, 266039, 1558465, 1498493, 604161, 608723, 1547013, 1484549, 1530909, 1325336, 1584866, 630737, 1023780, 227549, 961233, 987591, 159267, 608424, 636687, 745397, 460687], "12": [1586709, 357745, 1274743, 202991, 1250469, 837216, 757864, 185235, 493474, 621351, 22706, 1096491, 186138, 1482718, 534525, 1157131, 1350934, 1167912, 1366603, 756644, 311945, 72417, 706444, 1464146, 1236814, 74555, 983713, 1182807, 1082263, 804850, 1028788, 1405036, 932746, 742978, 1589076, 723766, 898380, 906214, 913629, 33386, 256107, 1245317, 360574, 836359, 565572, 400905, 1269191, 1042959, 886332, 1415571, 68870, 470376, 1531, 1333957, 834102, 674685, 839119, 255119, 552181, 1239034, 490552, 468715, 963242, 1612174, 21991, 1157603, 975554, 1342832, 883818, 1475260, 1666539, 1252155, 986632, 906429, 1495805, 561464, 1219097, 567704, 1463586, 1571831, 689821, 481310, 1407520, 97442, 512704, 962995, 876670, 956570, 1292534, 914981, 1655195, 870222, 196978, 617479, 1610748, 128227, 213356, 1590232, 230140, 942835]}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..8fee3b3aa7c82fcbfe4da6844135b298f508ade5
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cac34c00873d9fce57646821a9a2e0ebf7ebcbebb5a84ba617550c0952c75f62
+size 346660
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..f6ff04d63043c3757a301ce4f8c5abae4805a102
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:769a43a054e7ed360a913ac958e8e624a5a36068b4f547d60e09b2e92818e503
+size 441844
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..9c9256e313a16c84767a6b63aeecf059d79a804d
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59feff9aa3650a50486a772116f2769f59b6ed58214dc1f8fb61c95fc0982f00
+size 96613
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..1486755b9f27f66d538c3cb153ea3fc6eb73de5b
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aec22952182df7f80673015e59dc198d8dac5501b47e991b7b11fd764187da55
+size 113885
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/training_log_6c733232-2eb7-41a9-b008-521a86795de0.txt b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/training_log_6c733232-2eb7-41a9-b008-521a86795de0.txt
new file mode 100644
index 0000000000000000000000000000000000000000..442fa4243ede593903b2d286e66aad14f3980258
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/training_log_6c733232-2eb7-41a9-b008-521a86795de0.txt
@@ -0,0 +1,5236 @@
+[2025-09-04 12:19:52] [Rank 0] PRINT: --- Script Start: Thu Sep  4 12:19:52 2025 ---
+[2025-09-04 12:19:52] [Rank 0] PRINT: --- Script Start: Thu Sep  4 12:19:52 2025 ---
+[2025-09-04 12:19:52] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=45, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 12:19:52] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=45, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 12:19:52] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 12:19:52] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 12:19:52] [Rank 0] PRINT: Using fixed seed: 45
+[2025-09-04 12:19:52] [Rank 0] PRINT: Using fixed seed: 45
+[2025-09-04 12:19:52] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45
+[2025-09-04 12:19:52] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45
+[2025-09-04 12:19:52] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 12:19:52] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 12:19:52] [Rank 0] PRINT: Constructing model...
+[2025-09-04 12:19:52] [Rank 0] PRINT: Constructing model...
+[2025-09-04 12:19:54] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 12:19:54] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 12:19:54] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 12:19:54] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 12:19:54] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 12:19:54] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 12:19:57] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 12:19:57] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 12:19:57] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 12:19:57] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 12:19:57] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 12:19:57] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 12:19:57] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 12:19:57] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 12:19:58] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 12:19:58] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 12:19:58] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 12:19:58] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 12:19:58] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 12:19:58] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 12:19:58] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 12:19:58] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 12:19:58] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 12:19:58] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 12:19:58] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 12:19:58] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 12:19:58] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 12:19:58] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 12:20:02] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 12:20:02] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 12:20:02] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 12:20:02] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 12:20:45] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 12:20:45] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 12:20:45] [Rank 0] PRINT: Starting training...
+[2025-09-04 12:20:45] [Rank 0] PRINT: Starting training...
+[2025-09-04 12:20:51] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/fixed_eval_indices.json
+[2025-09-04 12:20:51] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/fixed_eval_indices.json
+[2025-09-04 12:20:51] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:20:51] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:20:56] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 12:20:56] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 12:21:33] [Rank 0] step:21/10000 train_time:37656ms step_avg:1793.13ms
+[2025-09-04 12:21:33] [Rank 0] step:21/10000 train_time:37656ms step_avg:1793.13ms
+[2025-09-04 12:21:34] [Rank 0] step:41/10000 train_time:38402ms step_avg:936.63ms
+[2025-09-04 12:21:34] [Rank 0] step:41/10000 train_time:38402ms step_avg:936.63ms
+[2025-09-04 12:21:35] [Rank 0] step:61/10000 train_time:39147ms step_avg:641.75ms
+[2025-09-04 12:21:35] [Rank 0] step:61/10000 train_time:39147ms step_avg:641.75ms
+[2025-09-04 12:21:36] [Rank 0] step:81/10000 train_time:39891ms step_avg:492.48ms
+[2025-09-04 12:21:36] [Rank 0] step:81/10000 train_time:39891ms step_avg:492.48ms
+[2025-09-04 12:21:36] [Rank 0] step:101/10000 train_time:40636ms step_avg:402.33ms
+[2025-09-04 12:21:36] [Rank 0] step:101/10000 train_time:40636ms step_avg:402.33ms
+[2025-09-04 12:21:37] [Rank 0] step:121/10000 train_time:41380ms step_avg:341.98ms
+[2025-09-04 12:21:37] [Rank 0] step:121/10000 train_time:41380ms step_avg:341.98ms
+[2025-09-04 12:21:38] [Rank 0] step:141/10000 train_time:42124ms step_avg:298.75ms
+[2025-09-04 12:21:38] [Rank 0] step:141/10000 train_time:42124ms step_avg:298.75ms
+[2025-09-04 12:21:39] [Rank 0] step:161/10000 train_time:42868ms step_avg:266.26ms
+[2025-09-04 12:21:39] [Rank 0] step:161/10000 train_time:42868ms step_avg:266.26ms
+[2025-09-04 12:21:39] [Rank 0] step:181/10000 train_time:43612ms step_avg:240.95ms
+[2025-09-04 12:21:39] [Rank 0] step:181/10000 train_time:43612ms step_avg:240.95ms
+[2025-09-04 12:21:40] [Rank 0] step:201/10000 train_time:44356ms step_avg:220.67ms
+[2025-09-04 12:21:40] [Rank 0] step:201/10000 train_time:44356ms step_avg:220.67ms
+[2025-09-04 12:21:41] [Rank 0] step:221/10000 train_time:45100ms step_avg:204.07ms
+[2025-09-04 12:21:41] [Rank 0] step:221/10000 train_time:45100ms step_avg:204.07ms
+[2025-09-04 12:21:41] [Rank 0] step:241/10000 train_time:45844ms step_avg:190.22ms
+[2025-09-04 12:21:41] [Rank 0] step:241/10000 train_time:45844ms step_avg:190.22ms
+[2025-09-04 12:21:42] [Rank 0] step:261/10000 train_time:46589ms step_avg:178.50ms
+[2025-09-04 12:21:42] [Rank 0] step:261/10000 train_time:46589ms step_avg:178.50ms
+[2025-09-04 12:21:43] [Rank 0] step:281/10000 train_time:47333ms step_avg:168.45ms
+[2025-09-04 12:21:43] [Rank 0] step:281/10000 train_time:47333ms step_avg:168.45ms
+[2025-09-04 12:21:44] [Rank 0] step:301/10000 train_time:48077ms step_avg:159.72ms
+[2025-09-04 12:21:44] [Rank 0] step:301/10000 train_time:48077ms step_avg:159.72ms
+[2025-09-04 12:21:44] [Rank 0] step:321/10000 train_time:48822ms step_avg:152.09ms
+[2025-09-04 12:21:44] [Rank 0] step:321/10000 train_time:48822ms step_avg:152.09ms
+[2025-09-04 12:21:45] [Rank 0] step:341/10000 train_time:49567ms step_avg:145.36ms
+[2025-09-04 12:21:45] [Rank 0] step:341/10000 train_time:49567ms step_avg:145.36ms
+[2025-09-04 12:21:46] [Rank 0] step:361/10000 train_time:50312ms step_avg:139.37ms
+[2025-09-04 12:21:46] [Rank 0] step:361/10000 train_time:50312ms step_avg:139.37ms
+[2025-09-04 12:21:47] [Rank 0] step:381/10000 train_time:51056ms step_avg:134.00ms
+[2025-09-04 12:21:47] [Rank 0] step:381/10000 train_time:51056ms step_avg:134.00ms
+[2025-09-04 12:21:47] [Rank 0] step:401/10000 train_time:51800ms step_avg:129.18ms
+[2025-09-04 12:21:47] [Rank 0] step:401/10000 train_time:51800ms step_avg:129.18ms
+[2025-09-04 12:21:48] [Rank 0] step:421/10000 train_time:52544ms step_avg:124.81ms
+[2025-09-04 12:21:48] [Rank 0] step:421/10000 train_time:52544ms step_avg:124.81ms
+[2025-09-04 12:21:49] [Rank 0] step:441/10000 train_time:53289ms step_avg:120.84ms
+[2025-09-04 12:21:49] [Rank 0] step:441/10000 train_time:53289ms step_avg:120.84ms
+[2025-09-04 12:21:50] [Rank 0] step:461/10000 train_time:54033ms step_avg:117.21ms
+[2025-09-04 12:21:50] [Rank 0] step:461/10000 train_time:54033ms step_avg:117.21ms
+[2025-09-04 12:21:50] [Rank 0] step:481/10000 train_time:54777ms step_avg:113.88ms
+[2025-09-04 12:21:50] [Rank 0] step:481/10000 train_time:54777ms step_avg:113.88ms
+[2025-09-04 12:21:51] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:21:51] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:21:52] [Rank 0] PRINT: step:500/10000 train_loss:3.1297 val_loss:1.1157 train_time:55527ms step_avg:111.05ms
+[2025-09-04 12:21:52] [Rank 0] PRINT: step:500/10000 train_loss:3.1297 val_loss:1.1157 train_time:55527ms step_avg:111.05ms
+[2025-09-04 12:21:52] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:21:52] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:21:52] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:21:52] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:23:29] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:23:29] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:23:29] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:23:29] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:23:29] [Rank 0]   Total Loss: 3.9011
+[2025-09-04 12:23:29] [Rank 0]   Total Loss: 3.9011
+[2025-09-04 12:23:29] [Rank 0]   Total FTA (Unweighted): 0.4788
+[2025-09-04 12:23:29] [Rank 0]   Total FTA (Unweighted): 0.4788
+[2025-09-04 12:23:29] [Rank 0]   Total FTA (Weighted):   0.4788
+[2025-09-04 12:23:29] [Rank 0]   Total FTA (Weighted):   0.4788
+[2025-09-04 12:23:29] [Rank 0]   Group 0 Loss: 3.4765
+[2025-09-04 12:23:29] [Rank 0]   Group 0 Loss: 3.4765
+[2025-09-04 12:23:29] [Rank 0]   Group 1 Loss: 3.2393
+[2025-09-04 12:23:29] [Rank 0]   Group 1 Loss: 3.2393
+[2025-09-04 12:23:29] [Rank 0]   Group 2 Loss: 3.1438
+[2025-09-04 12:23:29] [Rank 0]   Group 2 Loss: 3.1438
+[2025-09-04 12:23:29] [Rank 0]   Group 3 Loss: 3.4699
+[2025-09-04 12:23:29] [Rank 0]   Group 3 Loss: 3.4699
+[2025-09-04 12:23:29] [Rank 0]   Group 4 Loss: 3.5028
+[2025-09-04 12:23:29] [Rank 0]   Group 4 Loss: 3.5028
+[2025-09-04 12:23:29] [Rank 0]   Group 5 Loss: 3.5695
+[2025-09-04 12:23:29] [Rank 0]   Group 5 Loss: 3.5695
+[2025-09-04 12:23:29] [Rank 0]   Group 6 Loss: 3.6227
+[2025-09-04 12:23:29] [Rank 0]   Group 6 Loss: 3.6227
+[2025-09-04 12:23:29] [Rank 0]   Group 7 Loss: 3.7242
+[2025-09-04 12:23:29] [Rank 0]   Group 7 Loss: 3.7242
+[2025-09-04 12:23:29] [Rank 0]   Group 8 Loss: 4.0064
+[2025-09-04 12:23:29] [Rank 0]   Group 8 Loss: 4.0064
+[2025-09-04 12:23:29] [Rank 0]   Group 9 Loss: 4.0920
+[2025-09-04 12:23:29] [Rank 0]   Group 9 Loss: 4.0920
+[2025-09-04 12:23:29] [Rank 0]   Group 10 Loss: 4.2635
+[2025-09-04 12:23:29] [Rank 0]   Group 10 Loss: 4.2635
+[2025-09-04 12:23:29] [Rank 0]   Group 11 Loss: 4.3493
+[2025-09-04 12:23:29] [Rank 0]   Group 11 Loss: 4.3493
+[2025-09-04 12:23:29] [Rank 0]   Group 12 Loss: 4.4013
+[2025-09-04 12:23:29] [Rank 0]   Group 12 Loss: 4.4013
+[2025-09-04 12:23:29] [Rank 0]   Group 13 Loss: 4.5354
+[2025-09-04 12:23:29] [Rank 0]   Group 13 Loss: 4.5354
+[2025-09-04 12:23:29] [Rank 0]   Group 14 Loss: 4.4987
+[2025-09-04 12:23:29] [Rank 0]   Group 14 Loss: 4.4987
+[2025-09-04 12:23:29] [Rank 0]   Group 15 Loss: 4.5225
+[2025-09-04 12:23:29] [Rank 0]   Group 15 Loss: 4.5225
+[2025-09-04 12:23:29] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:23:29] [Rank 0]   Group 4 FTA: 0.9500
+[2025-09-04 12:23:29] [Rank 0]   Group 4 FTA: 0.9500
+[2025-09-04 12:23:29] [Rank 0]   Group 5 FTA: 0.6400
+[2025-09-04 12:23:29] [Rank 0]   Group 5 FTA: 0.6400
+[2025-09-04 12:23:29] [Rank 0]   Group 6 FTA: 0.4800
+[2025-09-04 12:23:29] [Rank 0]   Group 6 FTA: 0.4800
+[2025-09-04 12:23:29] [Rank 0]   Group 7 FTA: 0.4200
+[2025-09-04 12:23:29] [Rank 0]   Group 7 FTA: 0.4200
+[2025-09-04 12:23:29] [Rank 0]   Group 8 FTA: 0.3600
+[2025-09-04 12:23:29] [Rank 0]   Group 8 FTA: 0.3600
+[2025-09-04 12:23:29] [Rank 0]   Group 9 FTA: 0.2000
+[2025-09-04 12:23:29] [Rank 0]   Group 9 FTA: 0.2000
+[2025-09-04 12:23:29] [Rank 0]   Group 10 FTA: 0.1400
+[2025-09-04 12:23:29] [Rank 0]   Group 10 FTA: 0.1400
+[2025-09-04 12:23:29] [Rank 0]   Group 11 FTA: 0.0700
+[2025-09-04 12:23:29] [Rank 0]   Group 11 FTA: 0.0700
+[2025-09-04 12:23:29] [Rank 0]   Group 12 FTA: 0.1000
+[2025-09-04 12:23:29] [Rank 0]   Group 12 FTA: 0.1000
+[2025-09-04 12:23:29] [Rank 0]   Group 13 FTA: 0.1200
+[2025-09-04 12:23:29] [Rank 0]   Group 13 FTA: 0.1200
+[2025-09-04 12:23:29] [Rank 0]   Group 14 FTA: 0.1100
+[2025-09-04 12:23:29] [Rank 0]   Group 14 FTA: 0.1100
+[2025-09-04 12:23:29] [Rank 0]   Group 15 FTA: 0.0700
+[2025-09-04 12:23:29] [Rank 0]   Group 15 FTA: 0.0700
+[2025-09-04 12:23:29] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:23:29] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:23:30] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:23:30] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:23:30] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:23:30] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:23:30] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:23:30] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:23:30] [Rank 0] step:501/10000 train_time:55545ms step_avg:110.87ms
+[2025-09-04 12:23:30] [Rank 0] step:501/10000 train_time:55545ms step_avg:110.87ms
+[2025-09-04 12:23:31] [Rank 0] step:521/10000 train_time:56399ms step_avg:108.25ms
+[2025-09-04 12:23:31] [Rank 0] step:521/10000 train_time:56399ms step_avg:108.25ms
+[2025-09-04 12:23:32] [Rank 0] step:541/10000 train_time:57143ms step_avg:105.63ms
+[2025-09-04 12:23:32] [Rank 0] step:541/10000 train_time:57143ms step_avg:105.63ms
+[2025-09-04 12:23:33] [Rank 0] step:561/10000 train_time:57888ms step_avg:103.19ms
+[2025-09-04 12:23:33] [Rank 0] step:561/10000 train_time:57888ms step_avg:103.19ms
+[2025-09-04 12:23:33] [Rank 0] step:581/10000 train_time:58851ms step_avg:101.29ms
+[2025-09-04 12:23:33] [Rank 0] step:581/10000 train_time:58851ms step_avg:101.29ms
+[2025-09-04 12:23:34] [Rank 0] step:601/10000 train_time:59596ms step_avg:99.16ms
+[2025-09-04 12:23:34] [Rank 0] step:601/10000 train_time:59596ms step_avg:99.16ms
+[2025-09-04 12:23:35] [Rank 0] step:621/10000 train_time:60340ms step_avg:97.17ms
+[2025-09-04 12:23:35] [Rank 0] step:621/10000 train_time:60340ms step_avg:97.17ms
+[2025-09-04 12:23:36] [Rank 0] step:641/10000 train_time:61084ms step_avg:95.30ms
+[2025-09-04 12:23:36] [Rank 0] step:641/10000 train_time:61084ms step_avg:95.30ms
+[2025-09-04 12:23:36] [Rank 0] step:661/10000 train_time:61829ms step_avg:93.54ms
+[2025-09-04 12:23:36] [Rank 0] step:661/10000 train_time:61829ms step_avg:93.54ms
+[2025-09-04 12:23:37] [Rank 0] step:681/10000 train_time:62574ms step_avg:91.89ms
+[2025-09-04 12:23:37] [Rank 0] step:681/10000 train_time:62574ms step_avg:91.89ms
+[2025-09-04 12:23:38] [Rank 0] step:701/10000 train_time:63319ms step_avg:90.33ms
+[2025-09-04 12:23:38] [Rank 0] step:701/10000 train_time:63319ms step_avg:90.33ms
+[2025-09-04 12:23:39] [Rank 0] step:721/10000 train_time:64064ms step_avg:88.85ms
+[2025-09-04 12:23:39] [Rank 0] step:721/10000 train_time:64064ms step_avg:88.85ms
+[2025-09-04 12:23:39] [Rank 0] step:741/10000 train_time:64809ms step_avg:87.46ms
+[2025-09-04 12:23:39] [Rank 0] step:741/10000 train_time:64809ms step_avg:87.46ms
+[2025-09-04 12:23:40] [Rank 0] step:761/10000 train_time:65558ms step_avg:86.15ms
+[2025-09-04 12:23:40] [Rank 0] step:761/10000 train_time:65558ms step_avg:86.15ms
+[2025-09-04 12:23:41] [Rank 0] step:781/10000 train_time:66307ms step_avg:84.90ms
+[2025-09-04 12:23:41] [Rank 0] step:781/10000 train_time:66307ms step_avg:84.90ms
+[2025-09-04 12:23:42] [Rank 0] step:801/10000 train_time:67056ms step_avg:83.72ms
+[2025-09-04 12:23:42] [Rank 0] step:801/10000 train_time:67056ms step_avg:83.72ms
+[2025-09-04 12:23:43] [Rank 0] step:821/10000 train_time:68486ms step_avg:83.42ms
+[2025-09-04 12:23:43] [Rank 0] step:821/10000 train_time:68486ms step_avg:83.42ms
+[2025-09-04 12:23:44] [Rank 0] step:841/10000 train_time:69236ms step_avg:82.33ms
+[2025-09-04 12:23:44] [Rank 0] step:841/10000 train_time:69236ms step_avg:82.33ms
+[2025-09-04 12:23:45] [Rank 0] step:861/10000 train_time:69986ms step_avg:81.28ms
+[2025-09-04 12:23:45] [Rank 0] step:861/10000 train_time:69986ms step_avg:81.28ms
+[2025-09-04 12:23:45] [Rank 0] step:881/10000 train_time:70736ms step_avg:80.29ms
+[2025-09-04 12:23:45] [Rank 0] step:881/10000 train_time:70736ms step_avg:80.29ms
+[2025-09-04 12:23:46] [Rank 0] step:901/10000 train_time:71485ms step_avg:79.34ms
+[2025-09-04 12:23:46] [Rank 0] step:901/10000 train_time:71485ms step_avg:79.34ms
+[2025-09-04 12:23:47] [Rank 0] step:921/10000 train_time:72235ms step_avg:78.43ms
+[2025-09-04 12:23:47] [Rank 0] step:921/10000 train_time:72235ms step_avg:78.43ms
+[2025-09-04 12:23:48] [Rank 0] step:941/10000 train_time:72985ms step_avg:77.56ms
+[2025-09-04 12:23:48] [Rank 0] step:941/10000 train_time:72985ms step_avg:77.56ms
+[2025-09-04 12:23:48] [Rank 0] step:961/10000 train_time:73735ms step_avg:76.73ms
+[2025-09-04 12:23:48] [Rank 0] step:961/10000 train_time:73735ms step_avg:76.73ms
+[2025-09-04 12:23:49] [Rank 0] step:981/10000 train_time:74485ms step_avg:75.93ms
+[2025-09-04 12:23:49] [Rank 0] step:981/10000 train_time:74485ms step_avg:75.93ms
+[2025-09-04 12:23:50] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:23:50] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:23:50] [Rank 0] PRINT: step:1000/10000 train_loss:0.9695 val_loss:0.8652 train_time:75240ms step_avg:75.24ms
+[2025-09-04 12:23:50] [Rank 0] PRINT: step:1000/10000 train_loss:0.9695 val_loss:0.8652 train_time:75240ms step_avg:75.24ms
+[2025-09-04 12:23:50] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:23:50] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:23:50] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:23:50] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:25:28] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:25:28] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:25:28] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:25:28] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:25:28] [Rank 0]   Total Loss: 4.0618
+[2025-09-04 12:25:28] [Rank 0]   Total Loss: 4.0618
+[2025-09-04 12:25:28] [Rank 0]   Total FTA (Unweighted): 0.6781
+[2025-09-04 12:25:28] [Rank 0]   Total FTA (Unweighted): 0.6781
+[2025-09-04 12:25:28] [Rank 0]   Total FTA (Weighted):   0.6781
+[2025-09-04 12:25:28] [Rank 0]   Total FTA (Weighted):   0.6781
+[2025-09-04 12:25:28] [Rank 0]   Group 0 Loss: 3.9511
+[2025-09-04 12:25:28] [Rank 0]   Group 0 Loss: 3.9511
+[2025-09-04 12:25:28] [Rank 0]   Group 1 Loss: 3.5616
+[2025-09-04 12:25:28] [Rank 0]   Group 1 Loss: 3.5616
+[2025-09-04 12:25:28] [Rank 0]   Group 2 Loss: 3.4543
+[2025-09-04 12:25:28] [Rank 0]   Group 2 Loss: 3.4543
+[2025-09-04 12:25:28] [Rank 0]   Group 3 Loss: 3.8786
+[2025-09-04 12:25:28] [Rank 0]   Group 3 Loss: 3.8786
+[2025-09-04 12:25:28] [Rank 0]   Group 4 Loss: 3.8111
+[2025-09-04 12:25:28] [Rank 0]   Group 4 Loss: 3.8111
+[2025-09-04 12:25:28] [Rank 0]   Group 5 Loss: 3.8127
+[2025-09-04 12:25:28] [Rank 0]   Group 5 Loss: 3.8127
+[2025-09-04 12:25:28] [Rank 0]   Group 6 Loss: 3.8041
+[2025-09-04 12:25:28] [Rank 0]   Group 6 Loss: 3.8041
+[2025-09-04 12:25:28] [Rank 0]   Group 7 Loss: 3.8256
+[2025-09-04 12:25:28] [Rank 0]   Group 7 Loss: 3.8256
+[2025-09-04 12:25:28] [Rank 0]   Group 8 Loss: 4.0190
+[2025-09-04 12:25:28] [Rank 0]   Group 8 Loss: 4.0190
+[2025-09-04 12:25:28] [Rank 0]   Group 9 Loss: 4.0426
+[2025-09-04 12:25:28] [Rank 0]   Group 9 Loss: 4.0426
+[2025-09-04 12:25:28] [Rank 0]   Group 10 Loss: 4.2365
+[2025-09-04 12:25:28] [Rank 0]   Group 10 Loss: 4.2365
+[2025-09-04 12:25:28] [Rank 0]   Group 11 Loss: 4.3679
+[2025-09-04 12:25:28] [Rank 0]   Group 11 Loss: 4.3679
+[2025-09-04 12:25:28] [Rank 0]   Group 12 Loss: 4.3770
+[2025-09-04 12:25:28] [Rank 0]   Group 12 Loss: 4.3770
+[2025-09-04 12:25:28] [Rank 0]   Group 13 Loss: 4.5521
+[2025-09-04 12:25:28] [Rank 0]   Group 13 Loss: 4.5521
+[2025-09-04 12:25:28] [Rank 0]   Group 14 Loss: 4.6184
+[2025-09-04 12:25:28] [Rank 0]   Group 14 Loss: 4.6184
+[2025-09-04 12:25:28] [Rank 0]   Group 15 Loss: 4.6759
+[2025-09-04 12:25:28] [Rank 0]   Group 15 Loss: 4.6759
+[2025-09-04 12:25:28] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:25:28] [Rank 0]   Group 7 FTA: 0.8700
+[2025-09-04 12:25:28] [Rank 0]   Group 7 FTA: 0.8700
+[2025-09-04 12:25:28] [Rank 0]   Group 8 FTA: 0.7800
+[2025-09-04 12:25:28] [Rank 0]   Group 8 FTA: 0.7800
+[2025-09-04 12:25:28] [Rank 0]   Group 9 FTA: 0.6500
+[2025-09-04 12:25:28] [Rank 0]   Group 9 FTA: 0.6500
+[2025-09-04 12:25:28] [Rank 0]   Group 10 FTA: 0.6200
+[2025-09-04 12:25:28] [Rank 0]   Group 10 FTA: 0.6200
+[2025-09-04 12:25:28] [Rank 0]   Group 11 FTA: 0.3700
+[2025-09-04 12:25:28] [Rank 0]   Group 11 FTA: 0.3700
+[2025-09-04 12:25:28] [Rank 0]   Group 12 FTA: 0.2000
+[2025-09-04 12:25:28] [Rank 0]   Group 12 FTA: 0.2000
+[2025-09-04 12:25:28] [Rank 0]   Group 13 FTA: 0.1200
+[2025-09-04 12:25:28] [Rank 0]   Group 13 FTA: 0.1200
+[2025-09-04 12:25:28] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 12:25:28] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 12:25:28] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 12:25:28] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 12:25:28] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:25:28] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:25:29] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:25:29] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:25:29] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:25:29] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:25:29] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:25:29] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:25:29] [Rank 0] step:1001/10000 train_time:75256ms step_avg:75.18ms
+[2025-09-04 12:25:29] [Rank 0] step:1001/10000 train_time:75256ms step_avg:75.18ms
+[2025-09-04 12:25:30] [Rank 0] step:1021/10000 train_time:76012ms step_avg:74.45ms
+[2025-09-04 12:25:30] [Rank 0] step:1021/10000 train_time:76012ms step_avg:74.45ms
+[2025-09-04 12:25:31] [Rank 0] step:1041/10000 train_time:76762ms step_avg:73.74ms
+[2025-09-04 12:25:31] [Rank 0] step:1041/10000 train_time:76762ms step_avg:73.74ms
+[2025-09-04 12:25:32] [Rank 0] step:1061/10000 train_time:77510ms step_avg:73.05ms
+[2025-09-04 12:25:32] [Rank 0] step:1061/10000 train_time:77510ms step_avg:73.05ms
+[2025-09-04 12:25:32] [Rank 0] step:1081/10000 train_time:78260ms step_avg:72.40ms
+[2025-09-04 12:25:32] [Rank 0] step:1081/10000 train_time:78260ms step_avg:72.40ms
+[2025-09-04 12:25:33] [Rank 0] step:1101/10000 train_time:79011ms step_avg:71.76ms
+[2025-09-04 12:25:33] [Rank 0] step:1101/10000 train_time:79011ms step_avg:71.76ms
+[2025-09-04 12:25:34] [Rank 0] step:1121/10000 train_time:79759ms step_avg:71.15ms
+[2025-09-04 12:25:34] [Rank 0] step:1121/10000 train_time:79759ms step_avg:71.15ms
+[2025-09-04 12:25:35] [Rank 0] step:1141/10000 train_time:80508ms step_avg:70.56ms
+[2025-09-04 12:25:35] [Rank 0] step:1141/10000 train_time:80508ms step_avg:70.56ms
+[2025-09-04 12:25:35] [Rank 0] step:1161/10000 train_time:81256ms step_avg:69.99ms
+[2025-09-04 12:25:35] [Rank 0] step:1161/10000 train_time:81256ms step_avg:69.99ms
+[2025-09-04 12:25:36] [Rank 0] step:1181/10000 train_time:82005ms step_avg:69.44ms
+[2025-09-04 12:25:36] [Rank 0] step:1181/10000 train_time:82005ms step_avg:69.44ms
+[2025-09-04 12:25:37] [Rank 0] step:1201/10000 train_time:83000ms step_avg:69.11ms
+[2025-09-04 12:25:37] [Rank 0] step:1201/10000 train_time:83000ms step_avg:69.11ms
+[2025-09-04 12:25:38] [Rank 0] step:1221/10000 train_time:83748ms step_avg:68.59ms
+[2025-09-04 12:25:38] [Rank 0] step:1221/10000 train_time:83748ms step_avg:68.59ms
+[2025-09-04 12:25:39] [Rank 0] step:1241/10000 train_time:84497ms step_avg:68.09ms
+[2025-09-04 12:25:39] [Rank 0] step:1241/10000 train_time:84497ms step_avg:68.09ms
+[2025-09-04 12:25:40] [Rank 0] step:1261/10000 train_time:85494ms step_avg:67.80ms
+[2025-09-04 12:25:40] [Rank 0] step:1261/10000 train_time:85494ms step_avg:67.80ms
+[2025-09-04 12:25:40] [Rank 0] step:1281/10000 train_time:86243ms step_avg:67.32ms
+[2025-09-04 12:25:40] [Rank 0] step:1281/10000 train_time:86243ms step_avg:67.32ms
+[2025-09-04 12:25:41] [Rank 0] step:1301/10000 train_time:86992ms step_avg:66.87ms
+[2025-09-04 12:25:41] [Rank 0] step:1301/10000 train_time:86992ms step_avg:66.87ms
+[2025-09-04 12:25:42] [Rank 0] step:1321/10000 train_time:87740ms step_avg:66.42ms
+[2025-09-04 12:25:42] [Rank 0] step:1321/10000 train_time:87740ms step_avg:66.42ms
+[2025-09-04 12:25:43] [Rank 0] step:1341/10000 train_time:88489ms step_avg:65.99ms
+[2025-09-04 12:25:43] [Rank 0] step:1341/10000 train_time:88489ms step_avg:65.99ms
+[2025-09-04 12:25:43] [Rank 0] step:1361/10000 train_time:89238ms step_avg:65.57ms
+[2025-09-04 12:25:43] [Rank 0] step:1361/10000 train_time:89238ms step_avg:65.57ms
+[2025-09-04 12:25:44] [Rank 0] step:1381/10000 train_time:89987ms step_avg:65.16ms
+[2025-09-04 12:25:44] [Rank 0] step:1381/10000 train_time:89987ms step_avg:65.16ms
+[2025-09-04 12:25:45] [Rank 0] step:1401/10000 train_time:90736ms step_avg:64.77ms
+[2025-09-04 12:25:45] [Rank 0] step:1401/10000 train_time:90736ms step_avg:64.77ms
+[2025-09-04 12:25:46] [Rank 0] step:1421/10000 train_time:91485ms step_avg:64.38ms
+[2025-09-04 12:25:46] [Rank 0] step:1421/10000 train_time:91485ms step_avg:64.38ms
+[2025-09-04 12:25:46] [Rank 0] step:1441/10000 train_time:92235ms step_avg:64.01ms
+[2025-09-04 12:25:46] [Rank 0] step:1441/10000 train_time:92235ms step_avg:64.01ms
+[2025-09-04 12:25:47] [Rank 0] step:1461/10000 train_time:92985ms step_avg:63.64ms
+[2025-09-04 12:25:47] [Rank 0] step:1461/10000 train_time:92985ms step_avg:63.64ms
+[2025-09-04 12:25:48] [Rank 0] step:1481/10000 train_time:93734ms step_avg:63.29ms
+[2025-09-04 12:25:48] [Rank 0] step:1481/10000 train_time:93734ms step_avg:63.29ms
+[2025-09-04 12:25:48] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:25:48] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:25:49] [Rank 0] PRINT: step:1500/10000 train_loss:0.8290 val_loss:0.7825 train_time:94489ms step_avg:62.99ms
+[2025-09-04 12:25:49] [Rank 0] PRINT: step:1500/10000 train_loss:0.8290 val_loss:0.7825 train_time:94489ms step_avg:62.99ms
+[2025-09-04 12:25:49] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:25:49] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:25:49] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:25:49] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:27:27] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:27:27] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:27:27] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:27:27] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:27:27] [Rank 0]   Total Loss: 4.4906
+[2025-09-04 12:27:27] [Rank 0]   Total Loss: 4.4906
+[2025-09-04 12:27:27] [Rank 0]   Total FTA (Unweighted): 0.7444
+[2025-09-04 12:27:27] [Rank 0]   Total FTA (Unweighted): 0.7444
+[2025-09-04 12:27:27] [Rank 0]   Total FTA (Weighted):   0.7444
+[2025-09-04 12:27:27] [Rank 0]   Total FTA (Weighted):   0.7444
+[2025-09-04 12:27:27] [Rank 0]   Group 0 Loss: 4.4573
+[2025-09-04 12:27:27] [Rank 0]   Group 0 Loss: 4.4573
+[2025-09-04 12:27:27] [Rank 0]   Group 1 Loss: 4.0623
+[2025-09-04 12:27:27] [Rank 0]   Group 1 Loss: 4.0623
+[2025-09-04 12:27:27] [Rank 0]   Group 2 Loss: 3.8363
+[2025-09-04 12:27:27] [Rank 0]   Group 2 Loss: 3.8363
+[2025-09-04 12:27:27] [Rank 0]   Group 3 Loss: 4.3381
+[2025-09-04 12:27:27] [Rank 0]   Group 3 Loss: 4.3381
+[2025-09-04 12:27:27] [Rank 0]   Group 4 Loss: 4.3225
+[2025-09-04 12:27:27] [Rank 0]   Group 4 Loss: 4.3225
+[2025-09-04 12:27:27] [Rank 0]   Group 5 Loss: 4.3203
+[2025-09-04 12:27:27] [Rank 0]   Group 5 Loss: 4.3203
+[2025-09-04 12:27:27] [Rank 0]   Group 6 Loss: 4.2585
+[2025-09-04 12:27:27] [Rank 0]   Group 6 Loss: 4.2585
+[2025-09-04 12:27:27] [Rank 0]   Group 7 Loss: 4.3033
+[2025-09-04 12:27:27] [Rank 0]   Group 7 Loss: 4.3033
+[2025-09-04 12:27:27] [Rank 0]   Group 8 Loss: 4.4381
+[2025-09-04 12:27:27] [Rank 0]   Group 8 Loss: 4.4381
+[2025-09-04 12:27:27] [Rank 0]   Group 9 Loss: 4.4506
+[2025-09-04 12:27:27] [Rank 0]   Group 9 Loss: 4.4506
+[2025-09-04 12:27:27] [Rank 0]   Group 10 Loss: 4.6163
+[2025-09-04 12:27:27] [Rank 0]   Group 10 Loss: 4.6163
+[2025-09-04 12:27:27] [Rank 0]   Group 11 Loss: 4.7339
+[2025-09-04 12:27:27] [Rank 0]   Group 11 Loss: 4.7339
+[2025-09-04 12:27:27] [Rank 0]   Group 12 Loss: 4.7223
+[2025-09-04 12:27:27] [Rank 0]   Group 12 Loss: 4.7223
+[2025-09-04 12:27:27] [Rank 0]   Group 13 Loss: 4.8817
+[2025-09-04 12:27:27] [Rank 0]   Group 13 Loss: 4.8817
+[2025-09-04 12:27:27] [Rank 0]   Group 14 Loss: 4.9869
+[2025-09-04 12:27:27] [Rank 0]   Group 14 Loss: 4.9869
+[2025-09-04 12:27:27] [Rank 0]   Group 15 Loss: 5.1215
+[2025-09-04 12:27:27] [Rank 0]   Group 15 Loss: 5.1215
+[2025-09-04 12:27:27] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:27:27] [Rank 0]   Group 8 FTA: 0.9500
+[2025-09-04 12:27:27] [Rank 0]   Group 8 FTA: 0.9500
+[2025-09-04 12:27:27] [Rank 0]   Group 9 FTA: 0.8100
+[2025-09-04 12:27:27] [Rank 0]   Group 9 FTA: 0.8100
+[2025-09-04 12:27:27] [Rank 0]   Group 10 FTA: 0.7800
+[2025-09-04 12:27:27] [Rank 0]   Group 10 FTA: 0.7800
+[2025-09-04 12:27:27] [Rank 0]   Group 11 FTA: 0.6300
+[2025-09-04 12:27:27] [Rank 0]   Group 11 FTA: 0.6300
+[2025-09-04 12:27:27] [Rank 0]   Group 12 FTA: 0.3100
+[2025-09-04 12:27:27] [Rank 0]   Group 12 FTA: 0.3100
+[2025-09-04 12:27:27] [Rank 0]   Group 13 FTA: 0.2200
+[2025-09-04 12:27:27] [Rank 0]   Group 13 FTA: 0.2200
+[2025-09-04 12:27:27] [Rank 0]   Group 14 FTA: 0.1200
+[2025-09-04 12:27:27] [Rank 0]   Group 14 FTA: 0.1200
+[2025-09-04 12:27:27] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 12:27:27] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 12:27:27] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:27:27] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:27:28] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:27:28] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:27:28] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:27:28] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:27:28] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:27:28] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:27:28] [Rank 0] step:1501/10000 train_time:94504ms step_avg:62.96ms
+[2025-09-04 12:27:28] [Rank 0] step:1501/10000 train_time:94504ms step_avg:62.96ms
+[2025-09-04 12:27:29] [Rank 0] step:1521/10000 train_time:95271ms step_avg:62.64ms
+[2025-09-04 12:27:29] [Rank 0] step:1521/10000 train_time:95271ms step_avg:62.64ms
+[2025-09-04 12:27:30] [Rank 0] step:1541/10000 train_time:96019ms step_avg:62.31ms
+[2025-09-04 12:27:30] [Rank 0] step:1541/10000 train_time:96019ms step_avg:62.31ms
+[2025-09-04 12:27:30] [Rank 0] step:1561/10000 train_time:96769ms step_avg:61.99ms
+[2025-09-04 12:27:30] [Rank 0] step:1561/10000 train_time:96769ms step_avg:61.99ms
+[2025-09-04 12:27:31] [Rank 0] step:1581/10000 train_time:97518ms step_avg:61.68ms
+[2025-09-04 12:27:31] [Rank 0] step:1581/10000 train_time:97518ms step_avg:61.68ms
+[2025-09-04 12:27:32] [Rank 0] step:1601/10000 train_time:98267ms step_avg:61.38ms
+[2025-09-04 12:27:32] [Rank 0] step:1601/10000 train_time:98267ms step_avg:61.38ms
+[2025-09-04 12:27:33] [Rank 0] step:1621/10000 train_time:99016ms step_avg:61.08ms
+[2025-09-04 12:27:33] [Rank 0] step:1621/10000 train_time:99016ms step_avg:61.08ms
+[2025-09-04 12:27:34] [Rank 0] step:1641/10000 train_time:100038ms step_avg:60.96ms
+[2025-09-04 12:27:34] [Rank 0] step:1641/10000 train_time:100038ms step_avg:60.96ms
+[2025-09-04 12:27:34] [Rank 0] step:1661/10000 train_time:100788ms step_avg:60.68ms
+[2025-09-04 12:27:34] [Rank 0] step:1661/10000 train_time:100788ms step_avg:60.68ms
+[2025-09-04 12:27:35] [Rank 0] step:1681/10000 train_time:101536ms step_avg:60.40ms
+[2025-09-04 12:27:35] [Rank 0] step:1681/10000 train_time:101536ms step_avg:60.40ms
+[2025-09-04 12:27:36] [Rank 0] step:1701/10000 train_time:102285ms step_avg:60.13ms
+[2025-09-04 12:27:36] [Rank 0] step:1701/10000 train_time:102285ms step_avg:60.13ms
+[2025-09-04 12:27:37] [Rank 0] step:1721/10000 train_time:103035ms step_avg:59.87ms
+[2025-09-04 12:27:37] [Rank 0] step:1721/10000 train_time:103035ms step_avg:59.87ms
+[2025-09-04 12:27:37] [Rank 0] step:1741/10000 train_time:103783ms step_avg:59.61ms
+[2025-09-04 12:27:37] [Rank 0] step:1741/10000 train_time:103783ms step_avg:59.61ms
+[2025-09-04 12:27:38] [Rank 0] step:1761/10000 train_time:104532ms step_avg:59.36ms
+[2025-09-04 12:27:38] [Rank 0] step:1761/10000 train_time:104532ms step_avg:59.36ms
+[2025-09-04 12:27:39] [Rank 0] step:1781/10000 train_time:105282ms step_avg:59.11ms
+[2025-09-04 12:27:39] [Rank 0] step:1781/10000 train_time:105282ms step_avg:59.11ms
+[2025-09-04 12:27:40] [Rank 0] step:1801/10000 train_time:106031ms step_avg:58.87ms
+[2025-09-04 12:27:40] [Rank 0] step:1801/10000 train_time:106031ms step_avg:58.87ms
+[2025-09-04 12:27:40] [Rank 0] step:1821/10000 train_time:106780ms step_avg:58.64ms
+[2025-09-04 12:27:40] [Rank 0] step:1821/10000 train_time:106780ms step_avg:58.64ms
+[2025-09-04 12:27:41] [Rank 0] step:1841/10000 train_time:107529ms step_avg:58.41ms
+[2025-09-04 12:27:41] [Rank 0] step:1841/10000 train_time:107529ms step_avg:58.41ms
+[2025-09-04 12:27:42] [Rank 0] step:1861/10000 train_time:108278ms step_avg:58.18ms
+[2025-09-04 12:27:42] [Rank 0] step:1861/10000 train_time:108278ms step_avg:58.18ms
+[2025-09-04 12:27:43] [Rank 0] step:1881/10000 train_time:109028ms step_avg:57.96ms
+[2025-09-04 12:27:43] [Rank 0] step:1881/10000 train_time:109028ms step_avg:57.96ms
+[2025-09-04 12:27:44] [Rank 0] step:1901/10000 train_time:110028ms step_avg:57.88ms
+[2025-09-04 12:27:44] [Rank 0] step:1901/10000 train_time:110028ms step_avg:57.88ms
+[2025-09-04 12:27:44] [Rank 0] step:1921/10000 train_time:110777ms step_avg:57.67ms
+[2025-09-04 12:27:44] [Rank 0] step:1921/10000 train_time:110777ms step_avg:57.67ms
+[2025-09-04 12:27:45] [Rank 0] step:1941/10000 train_time:111527ms step_avg:57.46ms
+[2025-09-04 12:27:45] [Rank 0] step:1941/10000 train_time:111527ms step_avg:57.46ms
+[2025-09-04 12:27:46] [Rank 0] step:1961/10000 train_time:112562ms step_avg:57.40ms
+[2025-09-04 12:27:46] [Rank 0] step:1961/10000 train_time:112562ms step_avg:57.40ms
+[2025-09-04 12:27:47] [Rank 0] step:1981/10000 train_time:113311ms step_avg:57.20ms
+[2025-09-04 12:27:47] [Rank 0] step:1981/10000 train_time:113311ms step_avg:57.20ms
+[2025-09-04 12:27:48] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:27:48] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:27:48] [Rank 0] PRINT: step:2000/10000 train_loss:0.7702 val_loss:0.7376 train_time:114065ms step_avg:57.03ms
+[2025-09-04 12:27:48] [Rank 0] PRINT: step:2000/10000 train_loss:0.7702 val_loss:0.7376 train_time:114065ms step_avg:57.03ms
+[2025-09-04 12:27:48] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:27:48] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:27:48] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:27:48] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:29:25] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:29:25] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:29:25] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:29:25] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:29:25] [Rank 0]   Total Loss: 4.5338
+[2025-09-04 12:29:25] [Rank 0]   Total Loss: 4.5338
+[2025-09-04 12:29:25] [Rank 0]   Total FTA (Unweighted): 0.8019
+[2025-09-04 12:29:25] [Rank 0]   Total FTA (Unweighted): 0.8019
+[2025-09-04 12:29:25] [Rank 0]   Total FTA (Weighted):   0.8019
+[2025-09-04 12:29:25] [Rank 0]   Total FTA (Weighted):   0.8019
+[2025-09-04 12:29:25] [Rank 0]   Group 0 Loss: 4.5056
+[2025-09-04 12:29:25] [Rank 0]   Group 0 Loss: 4.5056
+[2025-09-04 12:29:25] [Rank 0]   Group 1 Loss: 4.0318
+[2025-09-04 12:29:25] [Rank 0]   Group 1 Loss: 4.0318
+[2025-09-04 12:29:25] [Rank 0]   Group 2 Loss: 3.8883
+[2025-09-04 12:29:25] [Rank 0]   Group 2 Loss: 3.8883
+[2025-09-04 12:29:25] [Rank 0]   Group 3 Loss: 4.4717
+[2025-09-04 12:29:25] [Rank 0]   Group 3 Loss: 4.4717
+[2025-09-04 12:29:25] [Rank 0]   Group 4 Loss: 4.3642
+[2025-09-04 12:29:25] [Rank 0]   Group 4 Loss: 4.3642
+[2025-09-04 12:29:25] [Rank 0]   Group 5 Loss: 4.4204
+[2025-09-04 12:29:25] [Rank 0]   Group 5 Loss: 4.4204
+[2025-09-04 12:29:25] [Rank 0]   Group 6 Loss: 4.4170
+[2025-09-04 12:29:25] [Rank 0]   Group 6 Loss: 4.4170
+[2025-09-04 12:29:25] [Rank 0]   Group 7 Loss: 4.3651
+[2025-09-04 12:29:25] [Rank 0]   Group 7 Loss: 4.3651
+[2025-09-04 12:29:25] [Rank 0]   Group 8 Loss: 4.5472
+[2025-09-04 12:29:25] [Rank 0]   Group 8 Loss: 4.5472
+[2025-09-04 12:29:25] [Rank 0]   Group 9 Loss: 4.5301
+[2025-09-04 12:29:25] [Rank 0]   Group 9 Loss: 4.5301
+[2025-09-04 12:29:25] [Rank 0]   Group 10 Loss: 4.6620
+[2025-09-04 12:29:25] [Rank 0]   Group 10 Loss: 4.6620
+[2025-09-04 12:29:25] [Rank 0]   Group 11 Loss: 4.7741
+[2025-09-04 12:29:25] [Rank 0]   Group 11 Loss: 4.7741
+[2025-09-04 12:29:25] [Rank 0]   Group 12 Loss: 4.7267
+[2025-09-04 12:29:25] [Rank 0]   Group 12 Loss: 4.7267
+[2025-09-04 12:29:25] [Rank 0]   Group 13 Loss: 4.8758
+[2025-09-04 12:29:25] [Rank 0]   Group 13 Loss: 4.8758
+[2025-09-04 12:29:25] [Rank 0]   Group 14 Loss: 4.9003
+[2025-09-04 12:29:25] [Rank 0]   Group 14 Loss: 4.9003
+[2025-09-04 12:29:25] [Rank 0]   Group 15 Loss: 5.0609
+[2025-09-04 12:29:25] [Rank 0]   Group 15 Loss: 5.0609
+[2025-09-04 12:29:25] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:29:25] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 12:29:25] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 12:29:25] [Rank 0]   Group 9 FTA: 0.9100
+[2025-09-04 12:29:25] [Rank 0]   Group 9 FTA: 0.9100
+[2025-09-04 12:29:25] [Rank 0]   Group 10 FTA: 0.9400
+[2025-09-04 12:29:25] [Rank 0]   Group 10 FTA: 0.9400
+[2025-09-04 12:29:25] [Rank 0]   Group 11 FTA: 0.8400
+[2025-09-04 12:29:25] [Rank 0]   Group 11 FTA: 0.8400
+[2025-09-04 12:29:25] [Rank 0]   Group 12 FTA: 0.6000
+[2025-09-04 12:29:25] [Rank 0]   Group 12 FTA: 0.6000
+[2025-09-04 12:29:25] [Rank 0]   Group 13 FTA: 0.2700
+[2025-09-04 12:29:25] [Rank 0]   Group 13 FTA: 0.2700
+[2025-09-04 12:29:25] [Rank 0]   Group 14 FTA: 0.1900
+[2025-09-04 12:29:25] [Rank 0]   Group 14 FTA: 0.1900
+[2025-09-04 12:29:25] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 12:29:25] [Rank 0]   Group 15 FTA: 0.0900
+[2025-09-04 12:29:26] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:29:26] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:29:26] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:29:26] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:29:26] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:29:26] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:29:27] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:29:27] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:29:27] [Rank 0] step:2001/10000 train_time:114080ms step_avg:57.01ms
+[2025-09-04 12:29:27] [Rank 0] step:2001/10000 train_time:114080ms step_avg:57.01ms
+[2025-09-04 12:29:28] [Rank 0] step:2021/10000 train_time:115095ms step_avg:56.95ms
+[2025-09-04 12:29:28] [Rank 0] step:2021/10000 train_time:115095ms step_avg:56.95ms
+[2025-09-04 12:29:28] [Rank 0] step:2041/10000 train_time:115844ms step_avg:56.76ms
+[2025-09-04 12:29:28] [Rank 0] step:2041/10000 train_time:115844ms step_avg:56.76ms
+[2025-09-04 12:29:29] [Rank 0] step:2061/10000 train_time:116594ms step_avg:56.57ms
+[2025-09-04 12:29:29] [Rank 0] step:2061/10000 train_time:116594ms step_avg:56.57ms
+[2025-09-04 12:29:30] [Rank 0] step:2081/10000 train_time:117342ms step_avg:56.39ms
+[2025-09-04 12:29:30] [Rank 0] step:2081/10000 train_time:117342ms step_avg:56.39ms
+[2025-09-04 12:29:31] [Rank 0] step:2101/10000 train_time:118091ms step_avg:56.21ms
+[2025-09-04 12:29:31] [Rank 0] step:2101/10000 train_time:118091ms step_avg:56.21ms
+[2025-09-04 12:29:31] [Rank 0] step:2121/10000 train_time:118839ms step_avg:56.03ms
+[2025-09-04 12:29:31] [Rank 0] step:2121/10000 train_time:118839ms step_avg:56.03ms
+[2025-09-04 12:29:32] [Rank 0] step:2141/10000 train_time:119588ms step_avg:55.86ms
+[2025-09-04 12:29:32] [Rank 0] step:2141/10000 train_time:119588ms step_avg:55.86ms
+[2025-09-04 12:29:33] [Rank 0] step:2161/10000 train_time:120336ms step_avg:55.69ms
+[2025-09-04 12:29:33] [Rank 0] step:2161/10000 train_time:120336ms step_avg:55.69ms
+[2025-09-04 12:29:34] [Rank 0] step:2181/10000 train_time:121085ms step_avg:55.52ms
+[2025-09-04 12:29:34] [Rank 0] step:2181/10000 train_time:121085ms step_avg:55.52ms
+[2025-09-04 12:29:34] [Rank 0] step:2201/10000 train_time:121833ms step_avg:55.35ms
+[2025-09-04 12:29:34] [Rank 0] step:2201/10000 train_time:121833ms step_avg:55.35ms
+[2025-09-04 12:29:35] [Rank 0] step:2221/10000 train_time:122581ms step_avg:55.19ms
+[2025-09-04 12:29:35] [Rank 0] step:2221/10000 train_time:122581ms step_avg:55.19ms
+[2025-09-04 12:29:36] [Rank 0] step:2241/10000 train_time:123339ms step_avg:55.04ms
+[2025-09-04 12:29:36] [Rank 0] step:2241/10000 train_time:123339ms step_avg:55.04ms
+[2025-09-04 12:29:37] [Rank 0] step:2261/10000 train_time:124098ms step_avg:54.89ms
+[2025-09-04 12:29:37] [Rank 0] step:2261/10000 train_time:124098ms step_avg:54.89ms
+[2025-09-04 12:29:37] [Rank 0] step:2281/10000 train_time:124857ms step_avg:54.74ms
+[2025-09-04 12:29:37] [Rank 0] step:2281/10000 train_time:124857ms step_avg:54.74ms
+[2025-09-04 12:29:38] [Rank 0] step:2301/10000 train_time:125616ms step_avg:54.59ms
+[2025-09-04 12:29:38] [Rank 0] step:2301/10000 train_time:125616ms step_avg:54.59ms
+[2025-09-04 12:29:39] [Rank 0] step:2321/10000 train_time:126375ms step_avg:54.45ms
+[2025-09-04 12:29:39] [Rank 0] step:2321/10000 train_time:126375ms step_avg:54.45ms
+[2025-09-04 12:29:40] [Rank 0] step:2341/10000 train_time:127135ms step_avg:54.31ms
+[2025-09-04 12:29:40] [Rank 0] step:2341/10000 train_time:127135ms step_avg:54.31ms
+[2025-09-04 12:29:40] [Rank 0] step:2361/10000 train_time:127895ms step_avg:54.17ms
+[2025-09-04 12:29:40] [Rank 0] step:2361/10000 train_time:127895ms step_avg:54.17ms
+[2025-09-04 12:29:41] [Rank 0] step:2381/10000 train_time:128655ms step_avg:54.03ms
+[2025-09-04 12:29:41] [Rank 0] step:2381/10000 train_time:128655ms step_avg:54.03ms
+[2025-09-04 12:29:42] [Rank 0] step:2401/10000 train_time:129415ms step_avg:53.90ms
+[2025-09-04 12:29:42] [Rank 0] step:2401/10000 train_time:129415ms step_avg:53.90ms
+[2025-09-04 12:29:43] [Rank 0] step:2421/10000 train_time:130177ms step_avg:53.77ms
+[2025-09-04 12:29:43] [Rank 0] step:2421/10000 train_time:130177ms step_avg:53.77ms
+[2025-09-04 12:29:43] [Rank 0] step:2441/10000 train_time:130936ms step_avg:53.64ms
+[2025-09-04 12:29:43] [Rank 0] step:2441/10000 train_time:130936ms step_avg:53.64ms
+[2025-09-04 12:29:44] [Rank 0] step:2461/10000 train_time:131695ms step_avg:53.51ms
+[2025-09-04 12:29:44] [Rank 0] step:2461/10000 train_time:131695ms step_avg:53.51ms
+[2025-09-04 12:29:45] [Rank 0] step:2481/10000 train_time:132455ms step_avg:53.39ms
+[2025-09-04 12:29:45] [Rank 0] step:2481/10000 train_time:132455ms step_avg:53.39ms
+[2025-09-04 12:29:46] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:29:46] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:29:46] [Rank 0] PRINT: step:2500/10000 train_loss:0.7331 val_loss:0.7052 train_time:133219ms step_avg:53.29ms
+[2025-09-04 12:29:46] [Rank 0] PRINT: step:2500/10000 train_loss:0.7331 val_loss:0.7052 train_time:133219ms step_avg:53.29ms
+[2025-09-04 12:29:46] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:29:46] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:29:46] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:29:46] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:31:23] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:31:23] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:31:23] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:31:23] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:31:23] [Rank 0]   Total Loss: 4.6467
+[2025-09-04 12:31:23] [Rank 0]   Total Loss: 4.6467
+[2025-09-04 12:31:23] [Rank 0]   Total FTA (Unweighted): 0.8337
+[2025-09-04 12:31:23] [Rank 0]   Total FTA (Unweighted): 0.8337
+[2025-09-04 12:31:23] [Rank 0]   Total FTA (Weighted):   0.8337
+[2025-09-04 12:31:23] [Rank 0]   Total FTA (Weighted):   0.8337
+[2025-09-04 12:31:23] [Rank 0]   Group 0 Loss: 4.6394
+[2025-09-04 12:31:23] [Rank 0]   Group 0 Loss: 4.6394
+[2025-09-04 12:31:23] [Rank 0]   Group 1 Loss: 4.2380
+[2025-09-04 12:31:23] [Rank 0]   Group 1 Loss: 4.2380
+[2025-09-04 12:31:23] [Rank 0]   Group 2 Loss: 3.9782
+[2025-09-04 12:31:23] [Rank 0]   Group 2 Loss: 3.9782
+[2025-09-04 12:31:23] [Rank 0]   Group 3 Loss: 4.6408
+[2025-09-04 12:31:23] [Rank 0]   Group 3 Loss: 4.6408
+[2025-09-04 12:31:23] [Rank 0]   Group 4 Loss: 4.5188
+[2025-09-04 12:31:23] [Rank 0]   Group 4 Loss: 4.5188
+[2025-09-04 12:31:23] [Rank 0]   Group 5 Loss: 4.4764
+[2025-09-04 12:31:23] [Rank 0]   Group 5 Loss: 4.4764
+[2025-09-04 12:31:23] [Rank 0]   Group 6 Loss: 4.5490
+[2025-09-04 12:31:23] [Rank 0]   Group 6 Loss: 4.5490
+[2025-09-04 12:31:23] [Rank 0]   Group 7 Loss: 4.5136
+[2025-09-04 12:31:23] [Rank 0]   Group 7 Loss: 4.5136
+[2025-09-04 12:31:23] [Rank 0]   Group 8 Loss: 4.7037
+[2025-09-04 12:31:23] [Rank 0]   Group 8 Loss: 4.7037
+[2025-09-04 12:31:23] [Rank 0]   Group 9 Loss: 4.6511
+[2025-09-04 12:31:23] [Rank 0]   Group 9 Loss: 4.6511
+[2025-09-04 12:31:23] [Rank 0]   Group 10 Loss: 4.7657
+[2025-09-04 12:31:23] [Rank 0]   Group 10 Loss: 4.7657
+[2025-09-04 12:31:23] [Rank 0]   Group 11 Loss: 4.8162
+[2025-09-04 12:31:23] [Rank 0]   Group 11 Loss: 4.8162
+[2025-09-04 12:31:23] [Rank 0]   Group 12 Loss: 4.8337
+[2025-09-04 12:31:23] [Rank 0]   Group 12 Loss: 4.8337
+[2025-09-04 12:31:23] [Rank 0]   Group 13 Loss: 4.9715
+[2025-09-04 12:31:23] [Rank 0]   Group 13 Loss: 4.9715
+[2025-09-04 12:31:23] [Rank 0]   Group 14 Loss: 4.9559
+[2025-09-04 12:31:23] [Rank 0]   Group 14 Loss: 4.9559
+[2025-09-04 12:31:23] [Rank 0]   Group 15 Loss: 5.0948
+[2025-09-04 12:31:23] [Rank 0]   Group 15 Loss: 5.0948
+[2025-09-04 12:31:23] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:31:23] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 12:31:23] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 12:31:23] [Rank 0]   Group 10 FTA: 0.9700
+[2025-09-04 12:31:23] [Rank 0]   Group 10 FTA: 0.9700
+[2025-09-04 12:31:23] [Rank 0]   Group 11 FTA: 0.8900
+[2025-09-04 12:31:23] [Rank 0]   Group 11 FTA: 0.8900
+[2025-09-04 12:31:23] [Rank 0]   Group 12 FTA: 0.7900
+[2025-09-04 12:31:23] [Rank 0]   Group 12 FTA: 0.7900
+[2025-09-04 12:31:23] [Rank 0]   Group 13 FTA: 0.4700
+[2025-09-04 12:31:23] [Rank 0]   Group 13 FTA: 0.4700
+[2025-09-04 12:31:23] [Rank 0]   Group 14 FTA: 0.1700
+[2025-09-04 12:31:23] [Rank 0]   Group 14 FTA: 0.1700
+[2025-09-04 12:31:23] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 12:31:23] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 12:31:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:31:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:31:24] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:31:24] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:31:24] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:31:24] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:31:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:31:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:31:25] [Rank 0] step:2501/10000 train_time:133235ms step_avg:53.27ms
+[2025-09-04 12:31:25] [Rank 0] step:2501/10000 train_time:133235ms step_avg:53.27ms
+[2025-09-04 12:31:26] [Rank 0] step:2521/10000 train_time:134007ms step_avg:53.16ms
+[2025-09-04 12:31:26] [Rank 0] step:2521/10000 train_time:134007ms step_avg:53.16ms
+[2025-09-04 12:31:26] [Rank 0] step:2541/10000 train_time:134767ms step_avg:53.04ms
+[2025-09-04 12:31:26] [Rank 0] step:2541/10000 train_time:134767ms step_avg:53.04ms
+[2025-09-04 12:31:27] [Rank 0] step:2561/10000 train_time:135527ms step_avg:52.92ms
+[2025-09-04 12:31:27] [Rank 0] step:2561/10000 train_time:135527ms step_avg:52.92ms
+[2025-09-04 12:31:28] [Rank 0] step:2581/10000 train_time:136285ms step_avg:52.80ms
+[2025-09-04 12:31:28] [Rank 0] step:2581/10000 train_time:136285ms step_avg:52.80ms
+[2025-09-04 12:31:29] [Rank 0] step:2601/10000 train_time:137044ms step_avg:52.69ms
+[2025-09-04 12:31:29] [Rank 0] step:2601/10000 train_time:137044ms step_avg:52.69ms
+[2025-09-04 12:31:29] [Rank 0] step:2621/10000 train_time:137804ms step_avg:52.58ms
+[2025-09-04 12:31:29] [Rank 0] step:2621/10000 train_time:137804ms step_avg:52.58ms
+[2025-09-04 12:31:30] [Rank 0] step:2641/10000 train_time:138563ms step_avg:52.47ms
+[2025-09-04 12:31:30] [Rank 0] step:2641/10000 train_time:138563ms step_avg:52.47ms
+[2025-09-04 12:31:31] [Rank 0] step:2661/10000 train_time:139323ms step_avg:52.36ms
+[2025-09-04 12:31:31] [Rank 0] step:2661/10000 train_time:139323ms step_avg:52.36ms
+[2025-09-04 12:31:32] [Rank 0] step:2681/10000 train_time:140082ms step_avg:52.25ms
+[2025-09-04 12:31:32] [Rank 0] step:2681/10000 train_time:140082ms step_avg:52.25ms
+[2025-09-04 12:31:32] [Rank 0] step:2701/10000 train_time:140842ms step_avg:52.14ms
+[2025-09-04 12:31:32] [Rank 0] step:2701/10000 train_time:140842ms step_avg:52.14ms
+[2025-09-04 12:31:33] [Rank 0] step:2721/10000 train_time:141601ms step_avg:52.04ms
+[2025-09-04 12:31:33] [Rank 0] step:2721/10000 train_time:141601ms step_avg:52.04ms
+[2025-09-04 12:31:34] [Rank 0] step:2741/10000 train_time:142360ms step_avg:51.94ms
+[2025-09-04 12:31:34] [Rank 0] step:2741/10000 train_time:142360ms step_avg:51.94ms
+[2025-09-04 12:31:35] [Rank 0] step:2761/10000 train_time:143119ms step_avg:51.84ms
+[2025-09-04 12:31:35] [Rank 0] step:2761/10000 train_time:143119ms step_avg:51.84ms
+[2025-09-04 12:31:35] [Rank 0] step:2781/10000 train_time:143878ms step_avg:51.74ms
+[2025-09-04 12:31:35] [Rank 0] step:2781/10000 train_time:143878ms step_avg:51.74ms
+[2025-09-04 12:31:36] [Rank 0] step:2801/10000 train_time:144638ms step_avg:51.64ms
+[2025-09-04 12:31:36] [Rank 0] step:2801/10000 train_time:144638ms step_avg:51.64ms
+[2025-09-04 12:31:37] [Rank 0] step:2821/10000 train_time:145670ms step_avg:51.64ms
+[2025-09-04 12:31:37] [Rank 0] step:2821/10000 train_time:145670ms step_avg:51.64ms
+[2025-09-04 12:31:38] [Rank 0] step:2841/10000 train_time:146429ms step_avg:51.54ms
+[2025-09-04 12:31:38] [Rank 0] step:2841/10000 train_time:146429ms step_avg:51.54ms
+[2025-09-04 12:31:39] [Rank 0] step:2861/10000 train_time:147188ms step_avg:51.45ms
+[2025-09-04 12:31:39] [Rank 0] step:2861/10000 train_time:147188ms step_avg:51.45ms
+[2025-09-04 12:31:40] [Rank 0] step:2881/10000 train_time:147949ms step_avg:51.35ms
+[2025-09-04 12:31:40] [Rank 0] step:2881/10000 train_time:147949ms step_avg:51.35ms
+[2025-09-04 12:31:40] [Rank 0] step:2901/10000 train_time:148709ms step_avg:51.26ms
+[2025-09-04 12:31:40] [Rank 0] step:2901/10000 train_time:148709ms step_avg:51.26ms
+[2025-09-04 12:31:41] [Rank 0] step:2921/10000 train_time:149469ms step_avg:51.17ms
+[2025-09-04 12:31:41] [Rank 0] step:2921/10000 train_time:149469ms step_avg:51.17ms
+[2025-09-04 12:31:42] [Rank 0] step:2941/10000 train_time:150228ms step_avg:51.08ms
+[2025-09-04 12:31:42] [Rank 0] step:2941/10000 train_time:150228ms step_avg:51.08ms
+[2025-09-04 12:31:43] [Rank 0] step:2961/10000 train_time:150988ms step_avg:50.99ms
+[2025-09-04 12:31:43] [Rank 0] step:2961/10000 train_time:150988ms step_avg:50.99ms
+[2025-09-04 12:31:43] [Rank 0] step:2981/10000 train_time:151749ms step_avg:50.91ms
+[2025-09-04 12:31:43] [Rank 0] step:2981/10000 train_time:151749ms step_avg:50.91ms
+[2025-09-04 12:31:44] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:31:44] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:31:45] [Rank 0] PRINT: step:3000/10000 train_loss:0.7066 val_loss:0.6854 train_time:152514ms step_avg:50.84ms
+[2025-09-04 12:31:45] [Rank 0] PRINT: step:3000/10000 train_loss:0.7066 val_loss:0.6854 train_time:152514ms step_avg:50.84ms
+[2025-09-04 12:31:45] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:31:45] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:31:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:31:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:33:22] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:33:22] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:33:22] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:33:22] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:33:22] [Rank 0]   Total Loss: 4.6950
+[2025-09-04 12:33:22] [Rank 0]   Total Loss: 4.6950
+[2025-09-04 12:33:22] [Rank 0]   Total FTA (Unweighted): 0.8581
+[2025-09-04 12:33:22] [Rank 0]   Total FTA (Unweighted): 0.8581
+[2025-09-04 12:33:22] [Rank 0]   Total FTA (Weighted):   0.8581
+[2025-09-04 12:33:22] [Rank 0]   Total FTA (Weighted):   0.8581
+[2025-09-04 12:33:22] [Rank 0]   Group 0 Loss: 4.6260
+[2025-09-04 12:33:22] [Rank 0]   Group 0 Loss: 4.6260
+[2025-09-04 12:33:22] [Rank 0]   Group 1 Loss: 4.2532
+[2025-09-04 12:33:22] [Rank 0]   Group 1 Loss: 4.2532
+[2025-09-04 12:33:22] [Rank 0]   Group 2 Loss: 4.2067
+[2025-09-04 12:33:22] [Rank 0]   Group 2 Loss: 4.2067
+[2025-09-04 12:33:22] [Rank 0]   Group 3 Loss: 4.6114
+[2025-09-04 12:33:22] [Rank 0]   Group 3 Loss: 4.6114
+[2025-09-04 12:33:22] [Rank 0]   Group 4 Loss: 4.6415
+[2025-09-04 12:33:22] [Rank 0]   Group 4 Loss: 4.6415
+[2025-09-04 12:33:22] [Rank 0]   Group 5 Loss: 4.5936
+[2025-09-04 12:33:22] [Rank 0]   Group 5 Loss: 4.5936
+[2025-09-04 12:33:22] [Rank 0]   Group 6 Loss: 4.5262
+[2025-09-04 12:33:22] [Rank 0]   Group 6 Loss: 4.5262
+[2025-09-04 12:33:22] [Rank 0]   Group 7 Loss: 4.6035
+[2025-09-04 12:33:22] [Rank 0]   Group 7 Loss: 4.6035
+[2025-09-04 12:33:22] [Rank 0]   Group 8 Loss: 4.7827
+[2025-09-04 12:33:22] [Rank 0]   Group 8 Loss: 4.7827
+[2025-09-04 12:33:22] [Rank 0]   Group 9 Loss: 4.7285
+[2025-09-04 12:33:22] [Rank 0]   Group 9 Loss: 4.7285
+[2025-09-04 12:33:22] [Rank 0]   Group 10 Loss: 4.8617
+[2025-09-04 12:33:22] [Rank 0]   Group 10 Loss: 4.8617
+[2025-09-04 12:33:22] [Rank 0]   Group 11 Loss: 4.8843
+[2025-09-04 12:33:22] [Rank 0]   Group 11 Loss: 4.8843
+[2025-09-04 12:33:22] [Rank 0]   Group 12 Loss: 4.8650
+[2025-09-04 12:33:22] [Rank 0]   Group 12 Loss: 4.8650
+[2025-09-04 12:33:22] [Rank 0]   Group 13 Loss: 4.9826
+[2025-09-04 12:33:22] [Rank 0]   Group 13 Loss: 4.9826
+[2025-09-04 12:33:22] [Rank 0]   Group 14 Loss: 4.9206
+[2025-09-04 12:33:22] [Rank 0]   Group 14 Loss: 4.9206
+[2025-09-04 12:33:22] [Rank 0]   Group 15 Loss: 5.0322
+[2025-09-04 12:33:22] [Rank 0]   Group 15 Loss: 5.0322
+[2025-09-04 12:33:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:33:22] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 12:33:22] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 12:33:22] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 12:33:22] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 12:33:22] [Rank 0]   Group 12 FTA: 0.9100
+[2025-09-04 12:33:22] [Rank 0]   Group 12 FTA: 0.9100
+[2025-09-04 12:33:22] [Rank 0]   Group 13 FTA: 0.5600
+[2025-09-04 12:33:22] [Rank 0]   Group 13 FTA: 0.5600
+[2025-09-04 12:33:22] [Rank 0]   Group 14 FTA: 0.1900
+[2025-09-04 12:33:22] [Rank 0]   Group 14 FTA: 0.1900
+[2025-09-04 12:33:22] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 12:33:22] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 12:33:23] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:33:23] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:33:23] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:33:23] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:33:24] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:33:24] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:33:24] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:33:24] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:33:24] [Rank 0] step:3001/10000 train_time:152529ms step_avg:50.83ms
+[2025-09-04 12:33:24] [Rank 0] step:3001/10000 train_time:152529ms step_avg:50.83ms
+[2025-09-04 12:33:25] [Rank 0] step:3021/10000 train_time:153301ms step_avg:50.75ms
+[2025-09-04 12:33:25] [Rank 0] step:3021/10000 train_time:153301ms step_avg:50.75ms
+[2025-09-04 12:33:25] [Rank 0] step:3041/10000 train_time:154060ms step_avg:50.66ms
+[2025-09-04 12:33:25] [Rank 0] step:3041/10000 train_time:154060ms step_avg:50.66ms
+[2025-09-04 12:33:26] [Rank 0] step:3061/10000 train_time:154819ms step_avg:50.58ms
+[2025-09-04 12:33:26] [Rank 0] step:3061/10000 train_time:154819ms step_avg:50.58ms
+[2025-09-04 12:33:27] [Rank 0] step:3081/10000 train_time:155577ms step_avg:50.50ms
+[2025-09-04 12:33:27] [Rank 0] step:3081/10000 train_time:155577ms step_avg:50.50ms
+[2025-09-04 12:33:28] [Rank 0] step:3101/10000 train_time:156336ms step_avg:50.41ms
+[2025-09-04 12:33:28] [Rank 0] step:3101/10000 train_time:156336ms step_avg:50.41ms
+[2025-09-04 12:33:28] [Rank 0] step:3121/10000 train_time:157095ms step_avg:50.33ms
+[2025-09-04 12:33:28] [Rank 0] step:3121/10000 train_time:157095ms step_avg:50.33ms
+[2025-09-04 12:33:29] [Rank 0] step:3141/10000 train_time:157855ms step_avg:50.26ms
+[2025-09-04 12:33:29] [Rank 0] step:3141/10000 train_time:157855ms step_avg:50.26ms
+[2025-09-04 12:33:30] [Rank 0] step:3161/10000 train_time:158615ms step_avg:50.18ms
+[2025-09-04 12:33:30] [Rank 0] step:3161/10000 train_time:158615ms step_avg:50.18ms
+[2025-09-04 12:33:31] [Rank 0] step:3181/10000 train_time:159375ms step_avg:50.10ms
+[2025-09-04 12:33:31] [Rank 0] step:3181/10000 train_time:159375ms step_avg:50.10ms
+[2025-09-04 12:33:32] [Rank 0] step:3201/10000 train_time:160136ms step_avg:50.03ms
+[2025-09-04 12:33:32] [Rank 0] step:3201/10000 train_time:160136ms step_avg:50.03ms
+[2025-09-04 12:33:32] [Rank 0] step:3221/10000 train_time:160895ms step_avg:49.95ms
+[2025-09-04 12:33:32] [Rank 0] step:3221/10000 train_time:160895ms step_avg:49.95ms
+[2025-09-04 12:33:33] [Rank 0] step:3241/10000 train_time:161655ms step_avg:49.88ms
+[2025-09-04 12:33:33] [Rank 0] step:3241/10000 train_time:161655ms step_avg:49.88ms
+[2025-09-04 12:33:34] [Rank 0] step:3261/10000 train_time:162415ms step_avg:49.81ms
+[2025-09-04 12:33:34] [Rank 0] step:3261/10000 train_time:162415ms step_avg:49.81ms
+[2025-09-04 12:33:35] [Rank 0] step:3281/10000 train_time:163174ms step_avg:49.73ms
+[2025-09-04 12:33:35] [Rank 0] step:3281/10000 train_time:163174ms step_avg:49.73ms
+[2025-09-04 12:33:35] [Rank 0] step:3301/10000 train_time:163934ms step_avg:49.66ms
+[2025-09-04 12:33:35] [Rank 0] step:3301/10000 train_time:163934ms step_avg:49.66ms
+[2025-09-04 12:33:36] [Rank 0] step:3321/10000 train_time:164693ms step_avg:49.59ms
+[2025-09-04 12:33:36] [Rank 0] step:3321/10000 train_time:164693ms step_avg:49.59ms
+[2025-09-04 12:33:37] [Rank 0] step:3341/10000 train_time:165452ms step_avg:49.52ms
+[2025-09-04 12:33:37] [Rank 0] step:3341/10000 train_time:165452ms step_avg:49.52ms
+[2025-09-04 12:33:38] [Rank 0] step:3361/10000 train_time:166210ms step_avg:49.45ms
+[2025-09-04 12:33:38] [Rank 0] step:3361/10000 train_time:166210ms step_avg:49.45ms
+[2025-09-04 12:33:38] [Rank 0] step:3381/10000 train_time:166969ms step_avg:49.38ms
+[2025-09-04 12:33:38] [Rank 0] step:3381/10000 train_time:166969ms step_avg:49.38ms
+[2025-09-04 12:33:39] [Rank 0] step:3401/10000 train_time:167727ms step_avg:49.32ms
+[2025-09-04 12:33:39] [Rank 0] step:3401/10000 train_time:167727ms step_avg:49.32ms
+[2025-09-04 12:33:40] [Rank 0] step:3421/10000 train_time:168486ms step_avg:49.25ms
+[2025-09-04 12:33:40] [Rank 0] step:3421/10000 train_time:168486ms step_avg:49.25ms
+[2025-09-04 12:33:41] [Rank 0] step:3441/10000 train_time:169245ms step_avg:49.18ms
+[2025-09-04 12:33:41] [Rank 0] step:3441/10000 train_time:169245ms step_avg:49.18ms
+[2025-09-04 12:33:41] [Rank 0] step:3461/10000 train_time:170004ms step_avg:49.12ms
+[2025-09-04 12:33:41] [Rank 0] step:3461/10000 train_time:170004ms step_avg:49.12ms
+[2025-09-04 12:33:42] [Rank 0] step:3481/10000 train_time:170763ms step_avg:49.06ms
+[2025-09-04 12:33:42] [Rank 0] step:3481/10000 train_time:170763ms step_avg:49.06ms
+[2025-09-04 12:33:43] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:33:43] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:33:43] [Rank 0] PRINT: step:3500/10000 train_loss:0.6898 val_loss:0.6715 train_time:171527ms step_avg:49.01ms
+[2025-09-04 12:33:43] [Rank 0] PRINT: step:3500/10000 train_loss:0.6898 val_loss:0.6715 train_time:171527ms step_avg:49.01ms
+[2025-09-04 12:33:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:33:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:33:44] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:33:44] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:35:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:35:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:35:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:35:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:35:21] [Rank 0]   Total Loss: 4.8172
+[2025-09-04 12:35:21] [Rank 0]   Total Loss: 4.8172
+[2025-09-04 12:35:21] [Rank 0]   Total FTA (Unweighted): 0.8738
+[2025-09-04 12:35:21] [Rank 0]   Total FTA (Unweighted): 0.8738
+[2025-09-04 12:35:21] [Rank 0]   Total FTA (Weighted):   0.8738
+[2025-09-04 12:35:21] [Rank 0]   Total FTA (Weighted):   0.8738
+[2025-09-04 12:35:21] [Rank 0]   Group 0 Loss: 4.7900
+[2025-09-04 12:35:21] [Rank 0]   Group 0 Loss: 4.7900
+[2025-09-04 12:35:21] [Rank 0]   Group 1 Loss: 4.4055
+[2025-09-04 12:35:21] [Rank 0]   Group 1 Loss: 4.4055
+[2025-09-04 12:35:21] [Rank 0]   Group 2 Loss: 4.2523
+[2025-09-04 12:35:21] [Rank 0]   Group 2 Loss: 4.2523
+[2025-09-04 12:35:21] [Rank 0]   Group 3 Loss: 4.7805
+[2025-09-04 12:35:21] [Rank 0]   Group 3 Loss: 4.7805
+[2025-09-04 12:35:21] [Rank 0]   Group 4 Loss: 4.6994
+[2025-09-04 12:35:21] [Rank 0]   Group 4 Loss: 4.6994
+[2025-09-04 12:35:21] [Rank 0]   Group 5 Loss: 4.7357
+[2025-09-04 12:35:21] [Rank 0]   Group 5 Loss: 4.7357
+[2025-09-04 12:35:21] [Rank 0]   Group 6 Loss: 4.6213
+[2025-09-04 12:35:21] [Rank 0]   Group 6 Loss: 4.6213
+[2025-09-04 12:35:21] [Rank 0]   Group 7 Loss: 4.7405
+[2025-09-04 12:35:21] [Rank 0]   Group 7 Loss: 4.7405
+[2025-09-04 12:35:21] [Rank 0]   Group 8 Loss: 4.9339
+[2025-09-04 12:35:21] [Rank 0]   Group 8 Loss: 4.9339
+[2025-09-04 12:35:21] [Rank 0]   Group 9 Loss: 4.8684
+[2025-09-04 12:35:21] [Rank 0]   Group 9 Loss: 4.8684
+[2025-09-04 12:35:21] [Rank 0]   Group 10 Loss: 5.0131
+[2025-09-04 12:35:21] [Rank 0]   Group 10 Loss: 5.0131
+[2025-09-04 12:35:21] [Rank 0]   Group 11 Loss: 5.0382
+[2025-09-04 12:35:21] [Rank 0]   Group 11 Loss: 5.0382
+[2025-09-04 12:35:21] [Rank 0]   Group 12 Loss: 4.9644
+[2025-09-04 12:35:21] [Rank 0]   Group 12 Loss: 4.9644
+[2025-09-04 12:35:21] [Rank 0]   Group 13 Loss: 5.0819
+[2025-09-04 12:35:21] [Rank 0]   Group 13 Loss: 5.0819
+[2025-09-04 12:35:21] [Rank 0]   Group 14 Loss: 5.0452
+[2025-09-04 12:35:21] [Rank 0]   Group 14 Loss: 5.0452
+[2025-09-04 12:35:21] [Rank 0]   Group 15 Loss: 5.1043
+[2025-09-04 12:35:21] [Rank 0]   Group 15 Loss: 5.1043
+[2025-09-04 12:35:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:35:21] [Rank 0]   Group 11 FTA: 0.9700
+[2025-09-04 12:35:21] [Rank 0]   Group 11 FTA: 0.9700
+[2025-09-04 12:35:21] [Rank 0]   Group 12 FTA: 0.9400
+[2025-09-04 12:35:21] [Rank 0]   Group 12 FTA: 0.9400
+[2025-09-04 12:35:21] [Rank 0]   Group 13 FTA: 0.6800
+[2025-09-04 12:35:21] [Rank 0]   Group 13 FTA: 0.6800
+[2025-09-04 12:35:21] [Rank 0]   Group 14 FTA: 0.2600
+[2025-09-04 12:35:21] [Rank 0]   Group 14 FTA: 0.2600
+[2025-09-04 12:35:21] [Rank 0]   Group 15 FTA: 0.1300
+[2025-09-04 12:35:21] [Rank 0]   Group 15 FTA: 0.1300
+[2025-09-04 12:35:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:35:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:35:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:35:22] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:35:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:35:22] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:35:22] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:35:22] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:35:23] [Rank 0] step:3501/10000 train_time:171542ms step_avg:49.00ms
+[2025-09-04 12:35:23] [Rank 0] step:3501/10000 train_time:171542ms step_avg:49.00ms
+[2025-09-04 12:35:23] [Rank 0] step:3521/10000 train_time:172325ms step_avg:48.94ms
+[2025-09-04 12:35:23] [Rank 0] step:3521/10000 train_time:172325ms step_avg:48.94ms
+[2025-09-04 12:35:24] [Rank 0] step:3541/10000 train_time:173083ms step_avg:48.88ms
+[2025-09-04 12:35:24] [Rank 0] step:3541/10000 train_time:173083ms step_avg:48.88ms
+[2025-09-04 12:35:25] [Rank 0] step:3561/10000 train_time:173842ms step_avg:48.82ms
+[2025-09-04 12:35:25] [Rank 0] step:3561/10000 train_time:173842ms step_avg:48.82ms
+[2025-09-04 12:35:26] [Rank 0] step:3581/10000 train_time:174602ms step_avg:48.76ms
+[2025-09-04 12:35:26] [Rank 0] step:3581/10000 train_time:174602ms step_avg:48.76ms
+[2025-09-04 12:35:26] [Rank 0] step:3601/10000 train_time:175361ms step_avg:48.70ms
+[2025-09-04 12:35:26] [Rank 0] step:3601/10000 train_time:175361ms step_avg:48.70ms
+[2025-09-04 12:35:27] [Rank 0] step:3621/10000 train_time:176121ms step_avg:48.64ms
+[2025-09-04 12:35:27] [Rank 0] step:3621/10000 train_time:176121ms step_avg:48.64ms
+[2025-09-04 12:35:28] [Rank 0] step:3641/10000 train_time:177148ms step_avg:48.65ms
+[2025-09-04 12:35:28] [Rank 0] step:3641/10000 train_time:177148ms step_avg:48.65ms
+[2025-09-04 12:35:29] [Rank 0] step:3661/10000 train_time:177909ms step_avg:48.60ms
+[2025-09-04 12:35:29] [Rank 0] step:3661/10000 train_time:177909ms step_avg:48.60ms
+[2025-09-04 12:35:30] [Rank 0] step:3681/10000 train_time:178668ms step_avg:48.54ms
+[2025-09-04 12:35:30] [Rank 0] step:3681/10000 train_time:178668ms step_avg:48.54ms
+[2025-09-04 12:35:30] [Rank 0] step:3701/10000 train_time:179428ms step_avg:48.48ms
+[2025-09-04 12:35:30] [Rank 0] step:3701/10000 train_time:179428ms step_avg:48.48ms
+[2025-09-04 12:35:31] [Rank 0] step:3721/10000 train_time:180187ms step_avg:48.42ms
+[2025-09-04 12:35:31] [Rank 0] step:3721/10000 train_time:180187ms step_avg:48.42ms
+[2025-09-04 12:35:32] [Rank 0] step:3741/10000 train_time:180946ms step_avg:48.37ms
+[2025-09-04 12:35:32] [Rank 0] step:3741/10000 train_time:180946ms step_avg:48.37ms
+[2025-09-04 12:35:33] [Rank 0] step:3761/10000 train_time:181706ms step_avg:48.31ms
+[2025-09-04 12:35:33] [Rank 0] step:3761/10000 train_time:181706ms step_avg:48.31ms
+[2025-09-04 12:35:33] [Rank 0] step:3781/10000 train_time:182465ms step_avg:48.26ms
+[2025-09-04 12:35:33] [Rank 0] step:3781/10000 train_time:182465ms step_avg:48.26ms
+[2025-09-04 12:35:34] [Rank 0] step:3801/10000 train_time:183224ms step_avg:48.20ms
+[2025-09-04 12:35:34] [Rank 0] step:3801/10000 train_time:183224ms step_avg:48.20ms
+[2025-09-04 12:35:35] [Rank 0] step:3821/10000 train_time:183985ms step_avg:48.15ms
+[2025-09-04 12:35:35] [Rank 0] step:3821/10000 train_time:183985ms step_avg:48.15ms
+[2025-09-04 12:35:36] [Rank 0] step:3841/10000 train_time:184745ms step_avg:48.10ms
+[2025-09-04 12:35:36] [Rank 0] step:3841/10000 train_time:184745ms step_avg:48.10ms
+[2025-09-04 12:35:36] [Rank 0] step:3861/10000 train_time:185504ms step_avg:48.05ms
+[2025-09-04 12:35:36] [Rank 0] step:3861/10000 train_time:185504ms step_avg:48.05ms
+[2025-09-04 12:35:37] [Rank 0] step:3881/10000 train_time:186263ms step_avg:47.99ms
+[2025-09-04 12:35:37] [Rank 0] step:3881/10000 train_time:186263ms step_avg:47.99ms
+[2025-09-04 12:35:38] [Rank 0] step:3901/10000 train_time:187023ms step_avg:47.94ms
+[2025-09-04 12:35:38] [Rank 0] step:3901/10000 train_time:187023ms step_avg:47.94ms
+[2025-09-04 12:35:39] [Rank 0] step:3921/10000 train_time:187782ms step_avg:47.89ms
+[2025-09-04 12:35:39] [Rank 0] step:3921/10000 train_time:187782ms step_avg:47.89ms
+[2025-09-04 12:35:40] [Rank 0] step:3941/10000 train_time:188542ms step_avg:47.84ms
+[2025-09-04 12:35:40] [Rank 0] step:3941/10000 train_time:188542ms step_avg:47.84ms
+[2025-09-04 12:35:40] [Rank 0] step:3961/10000 train_time:189302ms step_avg:47.79ms
+[2025-09-04 12:35:40] [Rank 0] step:3961/10000 train_time:189302ms step_avg:47.79ms
+[2025-09-04 12:35:41] [Rank 0] step:3981/10000 train_time:190061ms step_avg:47.74ms
+[2025-09-04 12:35:41] [Rank 0] step:3981/10000 train_time:190061ms step_avg:47.74ms
+[2025-09-04 12:35:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:35:42] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:35:42] [Rank 0] PRINT: step:4000/10000 train_loss:0.6773 val_loss:0.6597 train_time:190825ms step_avg:47.71ms
+[2025-09-04 12:35:42] [Rank 0] PRINT: step:4000/10000 train_loss:0.6773 val_loss:0.6597 train_time:190825ms step_avg:47.71ms
+[2025-09-04 12:35:42] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:35:42] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:35:42] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:35:42] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:37:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:37:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:37:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:37:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:37:20] [Rank 0]   Total Loss: 4.7451
+[2025-09-04 12:37:20] [Rank 0]   Total Loss: 4.7451
+[2025-09-04 12:37:20] [Rank 0]   Total FTA (Unweighted): 0.9006
+[2025-09-04 12:37:20] [Rank 0]   Total FTA (Unweighted): 0.9006
+[2025-09-04 12:37:20] [Rank 0]   Total FTA (Weighted):   0.9006
+[2025-09-04 12:37:20] [Rank 0]   Total FTA (Weighted):   0.9006
+[2025-09-04 12:37:20] [Rank 0]   Group 0 Loss: 4.6930
+[2025-09-04 12:37:20] [Rank 0]   Group 0 Loss: 4.6930
+[2025-09-04 12:37:20] [Rank 0]   Group 1 Loss: 4.3787
+[2025-09-04 12:37:20] [Rank 0]   Group 1 Loss: 4.3787
+[2025-09-04 12:37:20] [Rank 0]   Group 2 Loss: 4.1847
+[2025-09-04 12:37:20] [Rank 0]   Group 2 Loss: 4.1847
+[2025-09-04 12:37:20] [Rank 0]   Group 3 Loss: 4.7276
+[2025-09-04 12:37:20] [Rank 0]   Group 3 Loss: 4.7276
+[2025-09-04 12:37:20] [Rank 0]   Group 4 Loss: 4.6389
+[2025-09-04 12:37:20] [Rank 0]   Group 4 Loss: 4.6389
+[2025-09-04 12:37:20] [Rank 0]   Group 5 Loss: 4.7171
+[2025-09-04 12:37:20] [Rank 0]   Group 5 Loss: 4.7171
+[2025-09-04 12:37:20] [Rank 0]   Group 6 Loss: 4.5905
+[2025-09-04 12:37:20] [Rank 0]   Group 6 Loss: 4.5905
+[2025-09-04 12:37:20] [Rank 0]   Group 7 Loss: 4.6533
+[2025-09-04 12:37:20] [Rank 0]   Group 7 Loss: 4.6533
+[2025-09-04 12:37:20] [Rank 0]   Group 8 Loss: 4.8177
+[2025-09-04 12:37:20] [Rank 0]   Group 8 Loss: 4.8177
+[2025-09-04 12:37:20] [Rank 0]   Group 9 Loss: 4.8336
+[2025-09-04 12:37:20] [Rank 0]   Group 9 Loss: 4.8336
+[2025-09-04 12:37:20] [Rank 0]   Group 10 Loss: 4.9298
+[2025-09-04 12:37:20] [Rank 0]   Group 10 Loss: 4.9298
+[2025-09-04 12:37:20] [Rank 0]   Group 11 Loss: 4.9517
+[2025-09-04 12:37:20] [Rank 0]   Group 11 Loss: 4.9517
+[2025-09-04 12:37:20] [Rank 0]   Group 12 Loss: 4.8972
+[2025-09-04 12:37:20] [Rank 0]   Group 12 Loss: 4.8972
+[2025-09-04 12:37:20] [Rank 0]   Group 13 Loss: 4.9955
+[2025-09-04 12:37:20] [Rank 0]   Group 13 Loss: 4.9955
+[2025-09-04 12:37:20] [Rank 0]   Group 14 Loss: 4.9378
+[2025-09-04 12:37:20] [Rank 0]   Group 14 Loss: 4.9378
+[2025-09-04 12:37:20] [Rank 0]   Group 15 Loss: 4.9752
+[2025-09-04 12:37:20] [Rank 0]   Group 15 Loss: 4.9752
+[2025-09-04 12:37:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:37:20] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:37:20] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:37:20] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:37:20] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:37:20] [Rank 0]   Group 13 FTA: 0.8500
+[2025-09-04 12:37:20] [Rank 0]   Group 13 FTA: 0.8500
+[2025-09-04 12:37:20] [Rank 0]   Group 14 FTA: 0.4000
+[2025-09-04 12:37:20] [Rank 0]   Group 14 FTA: 0.4000
+[2025-09-04 12:37:20] [Rank 0]   Group 15 FTA: 0.1800
+[2025-09-04 12:37:20] [Rank 0]   Group 15 FTA: 0.1800
+[2025-09-04 12:37:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:37:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:37:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:37:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:37:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:37:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:37:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:37:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:37:21] [Rank 0] step:4001/10000 train_time:190841ms step_avg:47.70ms
+[2025-09-04 12:37:21] [Rank 0] step:4001/10000 train_time:190841ms step_avg:47.70ms
+[2025-09-04 12:37:22] [Rank 0] step:4021/10000 train_time:191865ms step_avg:47.72ms
+[2025-09-04 12:37:22] [Rank 0] step:4021/10000 train_time:191865ms step_avg:47.72ms
+[2025-09-04 12:37:23] [Rank 0] step:4041/10000 train_time:192623ms step_avg:47.67ms
+[2025-09-04 12:37:23] [Rank 0] step:4041/10000 train_time:192623ms step_avg:47.67ms
+[2025-09-04 12:37:24] [Rank 0] step:4061/10000 train_time:193382ms step_avg:47.62ms
+[2025-09-04 12:37:24] [Rank 0] step:4061/10000 train_time:193382ms step_avg:47.62ms
+[2025-09-04 12:37:25] [Rank 0] step:4081/10000 train_time:194141ms step_avg:47.57ms
+[2025-09-04 12:37:25] [Rank 0] step:4081/10000 train_time:194141ms step_avg:47.57ms
+[2025-09-04 12:37:25] [Rank 0] step:4101/10000 train_time:194900ms step_avg:47.53ms
+[2025-09-04 12:37:25] [Rank 0] step:4101/10000 train_time:194900ms step_avg:47.53ms
+[2025-09-04 12:37:26] [Rank 0] step:4121/10000 train_time:195659ms step_avg:47.48ms
+[2025-09-04 12:37:26] [Rank 0] step:4121/10000 train_time:195659ms step_avg:47.48ms
+[2025-09-04 12:37:27] [Rank 0] step:4141/10000 train_time:196418ms step_avg:47.43ms
+[2025-09-04 12:37:27] [Rank 0] step:4141/10000 train_time:196418ms step_avg:47.43ms
+[2025-09-04 12:37:28] [Rank 0] step:4161/10000 train_time:197177ms step_avg:47.39ms
+[2025-09-04 12:37:28] [Rank 0] step:4161/10000 train_time:197177ms step_avg:47.39ms
+[2025-09-04 12:37:28] [Rank 0] step:4181/10000 train_time:197935ms step_avg:47.34ms
+[2025-09-04 12:37:28] [Rank 0] step:4181/10000 train_time:197935ms step_avg:47.34ms
+[2025-09-04 12:37:29] [Rank 0] step:4201/10000 train_time:198694ms step_avg:47.30ms
+[2025-09-04 12:37:29] [Rank 0] step:4201/10000 train_time:198694ms step_avg:47.30ms
+[2025-09-04 12:37:30] [Rank 0] step:4221/10000 train_time:199453ms step_avg:47.25ms
+[2025-09-04 12:37:30] [Rank 0] step:4221/10000 train_time:199453ms step_avg:47.25ms
+[2025-09-04 12:37:31] [Rank 0] step:4241/10000 train_time:200212ms step_avg:47.21ms
+[2025-09-04 12:37:31] [Rank 0] step:4241/10000 train_time:200212ms step_avg:47.21ms
+[2025-09-04 12:37:31] [Rank 0] step:4261/10000 train_time:200972ms step_avg:47.17ms
+[2025-09-04 12:37:31] [Rank 0] step:4261/10000 train_time:200972ms step_avg:47.17ms
+[2025-09-04 12:37:32] [Rank 0] step:4281/10000 train_time:201731ms step_avg:47.12ms
+[2025-09-04 12:37:32] [Rank 0] step:4281/10000 train_time:201731ms step_avg:47.12ms
+[2025-09-04 12:37:33] [Rank 0] step:4301/10000 train_time:202490ms step_avg:47.08ms
+[2025-09-04 12:37:33] [Rank 0] step:4301/10000 train_time:202490ms step_avg:47.08ms
+[2025-09-04 12:37:34] [Rank 0] step:4321/10000 train_time:203249ms step_avg:47.04ms
+[2025-09-04 12:37:34] [Rank 0] step:4321/10000 train_time:203249ms step_avg:47.04ms
+[2025-09-04 12:37:34] [Rank 0] step:4341/10000 train_time:204007ms step_avg:47.00ms
+[2025-09-04 12:37:34] [Rank 0] step:4341/10000 train_time:204007ms step_avg:47.00ms
+[2025-09-04 12:37:35] [Rank 0] step:4361/10000 train_time:204766ms step_avg:46.95ms
+[2025-09-04 12:37:35] [Rank 0] step:4361/10000 train_time:204766ms step_avg:46.95ms
+[2025-09-04 12:37:36] [Rank 0] step:4381/10000 train_time:205524ms step_avg:46.91ms
+[2025-09-04 12:37:36] [Rank 0] step:4381/10000 train_time:205524ms step_avg:46.91ms
+[2025-09-04 12:37:37] [Rank 0] step:4401/10000 train_time:206284ms step_avg:46.87ms
+[2025-09-04 12:37:37] [Rank 0] step:4401/10000 train_time:206284ms step_avg:46.87ms
+[2025-09-04 12:37:37] [Rank 0] step:4421/10000 train_time:207042ms step_avg:46.83ms
+[2025-09-04 12:37:37] [Rank 0] step:4421/10000 train_time:207042ms step_avg:46.83ms
+[2025-09-04 12:37:38] [Rank 0] step:4441/10000 train_time:207802ms step_avg:46.79ms
+[2025-09-04 12:37:38] [Rank 0] step:4441/10000 train_time:207802ms step_avg:46.79ms
+[2025-09-04 12:37:39] [Rank 0] step:4461/10000 train_time:208561ms step_avg:46.75ms
+[2025-09-04 12:37:39] [Rank 0] step:4461/10000 train_time:208561ms step_avg:46.75ms
+[2025-09-04 12:37:40] [Rank 0] step:4481/10000 train_time:209320ms step_avg:46.71ms
+[2025-09-04 12:37:40] [Rank 0] step:4481/10000 train_time:209320ms step_avg:46.71ms
+[2025-09-04 12:37:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:37:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:37:41] [Rank 0] PRINT: step:4500/10000 train_loss:0.6671 val_loss:0.6505 train_time:210083ms step_avg:46.69ms
+[2025-09-04 12:37:41] [Rank 0] PRINT: step:4500/10000 train_loss:0.6671 val_loss:0.6505 train_time:210083ms step_avg:46.69ms
+[2025-09-04 12:37:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:37:41] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:37:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:37:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:39:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:39:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:39:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:39:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:39:18] [Rank 0]   Total Loss: 4.8982
+[2025-09-04 12:39:18] [Rank 0]   Total Loss: 4.8982
+[2025-09-04 12:39:18] [Rank 0]   Total FTA (Unweighted): 0.9100
+[2025-09-04 12:39:18] [Rank 0]   Total FTA (Unweighted): 0.9100
+[2025-09-04 12:39:18] [Rank 0]   Total FTA (Weighted):   0.9100
+[2025-09-04 12:39:18] [Rank 0]   Total FTA (Weighted):   0.9100
+[2025-09-04 12:39:18] [Rank 0]   Group 0 Loss: 4.9329
+[2025-09-04 12:39:18] [Rank 0]   Group 0 Loss: 4.9329
+[2025-09-04 12:39:18] [Rank 0]   Group 1 Loss: 4.4186
+[2025-09-04 12:39:18] [Rank 0]   Group 1 Loss: 4.4186
+[2025-09-04 12:39:18] [Rank 0]   Group 2 Loss: 4.3313
+[2025-09-04 12:39:18] [Rank 0]   Group 2 Loss: 4.3313
+[2025-09-04 12:39:18] [Rank 0]   Group 3 Loss: 4.9230
+[2025-09-04 12:39:18] [Rank 0]   Group 3 Loss: 4.9230
+[2025-09-04 12:39:18] [Rank 0]   Group 4 Loss: 4.7981
+[2025-09-04 12:39:18] [Rank 0]   Group 4 Loss: 4.7981
+[2025-09-04 12:39:18] [Rank 0]   Group 5 Loss: 4.7711
+[2025-09-04 12:39:18] [Rank 0]   Group 5 Loss: 4.7711
+[2025-09-04 12:39:18] [Rank 0]   Group 6 Loss: 4.7611
+[2025-09-04 12:39:18] [Rank 0]   Group 6 Loss: 4.7611
+[2025-09-04 12:39:18] [Rank 0]   Group 7 Loss: 4.8282
+[2025-09-04 12:39:18] [Rank 0]   Group 7 Loss: 4.8282
+[2025-09-04 12:39:18] [Rank 0]   Group 8 Loss: 4.9630
+[2025-09-04 12:39:18] [Rank 0]   Group 8 Loss: 4.9630
+[2025-09-04 12:39:18] [Rank 0]   Group 9 Loss: 4.9712
+[2025-09-04 12:39:18] [Rank 0]   Group 9 Loss: 4.9712
+[2025-09-04 12:39:18] [Rank 0]   Group 10 Loss: 5.1123
+[2025-09-04 12:39:18] [Rank 0]   Group 10 Loss: 5.1123
+[2025-09-04 12:39:18] [Rank 0]   Group 11 Loss: 5.1019
+[2025-09-04 12:39:18] [Rank 0]   Group 11 Loss: 5.1019
+[2025-09-04 12:39:18] [Rank 0]   Group 12 Loss: 5.0583
+[2025-09-04 12:39:18] [Rank 0]   Group 12 Loss: 5.0583
+[2025-09-04 12:39:18] [Rank 0]   Group 13 Loss: 5.1667
+[2025-09-04 12:39:18] [Rank 0]   Group 13 Loss: 5.1667
+[2025-09-04 12:39:18] [Rank 0]   Group 14 Loss: 5.0884
+[2025-09-04 12:39:18] [Rank 0]   Group 14 Loss: 5.0884
+[2025-09-04 12:39:18] [Rank 0]   Group 15 Loss: 5.1447
+[2025-09-04 12:39:18] [Rank 0]   Group 15 Loss: 5.1447
+[2025-09-04 12:39:18] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 12:39:18] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 12:39:18] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:39:18] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:39:18] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:39:18] [Rank 0]   Group 13 FTA: 0.9300
+[2025-09-04 12:39:18] [Rank 0]   Group 13 FTA: 0.9300
+[2025-09-04 12:39:18] [Rank 0]   Group 14 FTA: 0.4500
+[2025-09-04 12:39:18] [Rank 0]   Group 14 FTA: 0.4500
+[2025-09-04 12:39:18] [Rank 0]   Group 15 FTA: 0.2000
+[2025-09-04 12:39:18] [Rank 0]   Group 15 FTA: 0.2000
+[2025-09-04 12:39:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:39:19] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:39:19] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:39:19] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:39:19] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:39:19] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:39:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:39:20] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:39:20] [Rank 0] step:4501/10000 train_time:210099ms step_avg:46.68ms
+[2025-09-04 12:39:20] [Rank 0] step:4501/10000 train_time:210099ms step_avg:46.68ms
+[2025-09-04 12:39:20] [Rank 0] step:4521/10000 train_time:210868ms step_avg:46.64ms
+[2025-09-04 12:39:20] [Rank 0] step:4521/10000 train_time:210868ms step_avg:46.64ms
+[2025-09-04 12:39:21] [Rank 0] step:4541/10000 train_time:211626ms step_avg:46.60ms
+[2025-09-04 12:39:21] [Rank 0] step:4541/10000 train_time:211626ms step_avg:46.60ms
+[2025-09-04 12:39:22] [Rank 0] step:4561/10000 train_time:212593ms step_avg:46.61ms
+[2025-09-04 12:39:22] [Rank 0] step:4561/10000 train_time:212593ms step_avg:46.61ms
+[2025-09-04 12:39:23] [Rank 0] step:4581/10000 train_time:213353ms step_avg:46.57ms
+[2025-09-04 12:39:23] [Rank 0] step:4581/10000 train_time:213353ms step_avg:46.57ms
+[2025-09-04 12:39:24] [Rank 0] step:4601/10000 train_time:214112ms step_avg:46.54ms
+[2025-09-04 12:39:24] [Rank 0] step:4601/10000 train_time:214112ms step_avg:46.54ms
+[2025-09-04 12:39:24] [Rank 0] step:4621/10000 train_time:214873ms step_avg:46.50ms
+[2025-09-04 12:39:24] [Rank 0] step:4621/10000 train_time:214873ms step_avg:46.50ms
+[2025-09-04 12:39:25] [Rank 0] step:4641/10000 train_time:215632ms step_avg:46.46ms
+[2025-09-04 12:39:25] [Rank 0] step:4641/10000 train_time:215632ms step_avg:46.46ms
+[2025-09-04 12:39:26] [Rank 0] step:4661/10000 train_time:216392ms step_avg:46.43ms
+[2025-09-04 12:39:26] [Rank 0] step:4661/10000 train_time:216392ms step_avg:46.43ms
+[2025-09-04 12:39:27] [Rank 0] step:4681/10000 train_time:217151ms step_avg:46.39ms
+[2025-09-04 12:39:27] [Rank 0] step:4681/10000 train_time:217151ms step_avg:46.39ms
+[2025-09-04 12:39:27] [Rank 0] step:4701/10000 train_time:217912ms step_avg:46.35ms
+[2025-09-04 12:39:27] [Rank 0] step:4701/10000 train_time:217912ms step_avg:46.35ms
+[2025-09-04 12:39:28] [Rank 0] step:4721/10000 train_time:218669ms step_avg:46.32ms
+[2025-09-04 12:39:28] [Rank 0] step:4721/10000 train_time:218669ms step_avg:46.32ms
+[2025-09-04 12:39:29] [Rank 0] step:4741/10000 train_time:219429ms step_avg:46.28ms
+[2025-09-04 12:39:29] [Rank 0] step:4741/10000 train_time:219429ms step_avg:46.28ms
+[2025-09-04 12:39:30] [Rank 0] step:4761/10000 train_time:220189ms step_avg:46.25ms
+[2025-09-04 12:39:30] [Rank 0] step:4761/10000 train_time:220189ms step_avg:46.25ms
+[2025-09-04 12:39:31] [Rank 0] step:4781/10000 train_time:220948ms step_avg:46.21ms
+[2025-09-04 12:39:31] [Rank 0] step:4781/10000 train_time:220948ms step_avg:46.21ms
+[2025-09-04 12:39:31] [Rank 0] step:4801/10000 train_time:221709ms step_avg:46.18ms
+[2025-09-04 12:39:31] [Rank 0] step:4801/10000 train_time:221709ms step_avg:46.18ms
+[2025-09-04 12:39:32] [Rank 0] step:4821/10000 train_time:222469ms step_avg:46.15ms
+[2025-09-04 12:39:32] [Rank 0] step:4821/10000 train_time:222469ms step_avg:46.15ms
+[2025-09-04 12:39:33] [Rank 0] step:4841/10000 train_time:223540ms step_avg:46.18ms
+[2025-09-04 12:39:33] [Rank 0] step:4841/10000 train_time:223540ms step_avg:46.18ms
+[2025-09-04 12:39:34] [Rank 0] step:4861/10000 train_time:224300ms step_avg:46.14ms
+[2025-09-04 12:39:34] [Rank 0] step:4861/10000 train_time:224300ms step_avg:46.14ms
+[2025-09-04 12:39:35] [Rank 0] step:4881/10000 train_time:225062ms step_avg:46.11ms
+[2025-09-04 12:39:35] [Rank 0] step:4881/10000 train_time:225062ms step_avg:46.11ms
+[2025-09-04 12:39:35] [Rank 0] step:4901/10000 train_time:225820ms step_avg:46.08ms
+[2025-09-04 12:39:35] [Rank 0] step:4901/10000 train_time:225820ms step_avg:46.08ms
+[2025-09-04 12:39:36] [Rank 0] step:4921/10000 train_time:226580ms step_avg:46.04ms
+[2025-09-04 12:39:36] [Rank 0] step:4921/10000 train_time:226580ms step_avg:46.04ms
+[2025-09-04 12:39:37] [Rank 0] step:4941/10000 train_time:227339ms step_avg:46.01ms
+[2025-09-04 12:39:37] [Rank 0] step:4941/10000 train_time:227339ms step_avg:46.01ms
+[2025-09-04 12:39:38] [Rank 0] step:4961/10000 train_time:228098ms step_avg:45.98ms
+[2025-09-04 12:39:38] [Rank 0] step:4961/10000 train_time:228098ms step_avg:45.98ms
+[2025-09-04 12:39:38] [Rank 0] step:4981/10000 train_time:228858ms step_avg:45.95ms
+[2025-09-04 12:39:38] [Rank 0] step:4981/10000 train_time:228858ms step_avg:45.95ms
+[2025-09-04 12:39:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:39:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:39:40] [Rank 0] PRINT: step:5000/10000 train_loss:0.6582 val_loss:0.6429 train_time:229622ms step_avg:45.92ms
+[2025-09-04 12:39:40] [Rank 0] PRINT: step:5000/10000 train_loss:0.6582 val_loss:0.6429 train_time:229622ms step_avg:45.92ms
+[2025-09-04 12:39:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:39:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:39:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:39:40] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:41:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:41:18] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:41:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:41:18] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:41:18] [Rank 0]   Total Loss: 4.8340
+[2025-09-04 12:41:18] [Rank 0]   Total Loss: 4.8340
+[2025-09-04 12:41:18] [Rank 0]   Total FTA (Unweighted): 0.9238
+[2025-09-04 12:41:18] [Rank 0]   Total FTA (Unweighted): 0.9238
+[2025-09-04 12:41:18] [Rank 0]   Total FTA (Weighted):   0.9237
+[2025-09-04 12:41:18] [Rank 0]   Total FTA (Weighted):   0.9237
+[2025-09-04 12:41:18] [Rank 0]   Group 0 Loss: 4.7494
+[2025-09-04 12:41:18] [Rank 0]   Group 0 Loss: 4.7494
+[2025-09-04 12:41:18] [Rank 0]   Group 1 Loss: 4.4182
+[2025-09-04 12:41:18] [Rank 0]   Group 1 Loss: 4.4182
+[2025-09-04 12:41:18] [Rank 0]   Group 2 Loss: 4.3089
+[2025-09-04 12:41:18] [Rank 0]   Group 2 Loss: 4.3089
+[2025-09-04 12:41:18] [Rank 0]   Group 3 Loss: 4.7786
+[2025-09-04 12:41:18] [Rank 0]   Group 3 Loss: 4.7786
+[2025-09-04 12:41:18] [Rank 0]   Group 4 Loss: 4.7662
+[2025-09-04 12:41:18] [Rank 0]   Group 4 Loss: 4.7662
+[2025-09-04 12:41:18] [Rank 0]   Group 5 Loss: 4.7581
+[2025-09-04 12:41:18] [Rank 0]   Group 5 Loss: 4.7581
+[2025-09-04 12:41:18] [Rank 0]   Group 6 Loss: 4.7510
+[2025-09-04 12:41:18] [Rank 0]   Group 6 Loss: 4.7510
+[2025-09-04 12:41:18] [Rank 0]   Group 7 Loss: 4.7300
+[2025-09-04 12:41:18] [Rank 0]   Group 7 Loss: 4.7300
+[2025-09-04 12:41:18] [Rank 0]   Group 8 Loss: 4.9050
+[2025-09-04 12:41:18] [Rank 0]   Group 8 Loss: 4.9050
+[2025-09-04 12:41:18] [Rank 0]   Group 9 Loss: 4.9367
+[2025-09-04 12:41:18] [Rank 0]   Group 9 Loss: 4.9367
+[2025-09-04 12:41:18] [Rank 0]   Group 10 Loss: 5.0296
+[2025-09-04 12:41:18] [Rank 0]   Group 10 Loss: 5.0296
+[2025-09-04 12:41:18] [Rank 0]   Group 11 Loss: 5.0846
+[2025-09-04 12:41:18] [Rank 0]   Group 11 Loss: 5.0846
+[2025-09-04 12:41:18] [Rank 0]   Group 12 Loss: 5.0148
+[2025-09-04 12:41:18] [Rank 0]   Group 12 Loss: 5.0148
+[2025-09-04 12:41:18] [Rank 0]   Group 13 Loss: 5.0945
+[2025-09-04 12:41:18] [Rank 0]   Group 13 Loss: 5.0945
+[2025-09-04 12:41:18] [Rank 0]   Group 14 Loss: 5.0154
+[2025-09-04 12:41:18] [Rank 0]   Group 14 Loss: 5.0154
+[2025-09-04 12:41:18] [Rank 0]   Group 15 Loss: 5.0035
+[2025-09-04 12:41:18] [Rank 0]   Group 15 Loss: 5.0035
+[2025-09-04 12:41:18] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:41:18] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 12:41:18] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 12:41:18] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:41:18] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:41:18] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 12:41:18] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 12:41:18] [Rank 0]   Group 13 FTA: 0.9300
+[2025-09-04 12:41:18] [Rank 0]   Group 13 FTA: 0.9300
+[2025-09-04 12:41:18] [Rank 0]   Group 14 FTA: 0.6500
+[2025-09-04 12:41:18] [Rank 0]   Group 14 FTA: 0.6500
+[2025-09-04 12:41:18] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 12:41:18] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 12:41:18] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:41:18] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:41:19] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:41:19] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:41:19] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:41:19] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:41:19] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:41:19] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:41:19] [Rank 0] step:5001/10000 train_time:229639ms step_avg:45.92ms
+[2025-09-04 12:41:19] [Rank 0] step:5001/10000 train_time:229639ms step_avg:45.92ms
+[2025-09-04 12:41:20] [Rank 0] step:5021/10000 train_time:230415ms step_avg:45.89ms
+[2025-09-04 12:41:20] [Rank 0] step:5021/10000 train_time:230415ms step_avg:45.89ms
+[2025-09-04 12:41:21] [Rank 0] step:5041/10000 train_time:231173ms step_avg:45.86ms
+[2025-09-04 12:41:21] [Rank 0] step:5041/10000 train_time:231173ms step_avg:45.86ms
+[2025-09-04 12:41:21] [Rank 0] step:5061/10000 train_time:231932ms step_avg:45.83ms
+[2025-09-04 12:41:21] [Rank 0] step:5061/10000 train_time:231932ms step_avg:45.83ms
+[2025-09-04 12:41:22] [Rank 0] step:5081/10000 train_time:232690ms step_avg:45.80ms
+[2025-09-04 12:41:22] [Rank 0] step:5081/10000 train_time:232690ms step_avg:45.80ms
+[2025-09-04 12:41:23] [Rank 0] step:5101/10000 train_time:233448ms step_avg:45.77ms
+[2025-09-04 12:41:23] [Rank 0] step:5101/10000 train_time:233448ms step_avg:45.77ms
+[2025-09-04 12:41:24] [Rank 0] step:5121/10000 train_time:234207ms step_avg:45.73ms
+[2025-09-04 12:41:24] [Rank 0] step:5121/10000 train_time:234207ms step_avg:45.73ms
+[2025-09-04 12:41:24] [Rank 0] step:5141/10000 train_time:234965ms step_avg:45.70ms
+[2025-09-04 12:41:24] [Rank 0] step:5141/10000 train_time:234965ms step_avg:45.70ms
+[2025-09-04 12:41:25] [Rank 0] step:5161/10000 train_time:235878ms step_avg:45.70ms
+[2025-09-04 12:41:25] [Rank 0] step:5161/10000 train_time:235878ms step_avg:45.70ms
+[2025-09-04 12:41:26] [Rank 0] step:5181/10000 train_time:236742ms step_avg:45.69ms
+[2025-09-04 12:41:26] [Rank 0] step:5181/10000 train_time:236742ms step_avg:45.69ms
+[2025-09-04 12:41:27] [Rank 0] step:5201/10000 train_time:237501ms step_avg:45.66ms
+[2025-09-04 12:41:27] [Rank 0] step:5201/10000 train_time:237501ms step_avg:45.66ms
+[2025-09-04 12:41:28] [Rank 0] step:5221/10000 train_time:238439ms step_avg:45.67ms
+[2025-09-04 12:41:28] [Rank 0] step:5221/10000 train_time:238439ms step_avg:45.67ms
+[2025-09-04 12:41:29] [Rank 0] step:5241/10000 train_time:239248ms step_avg:45.65ms
+[2025-09-04 12:41:29] [Rank 0] step:5241/10000 train_time:239248ms step_avg:45.65ms
+[2025-09-04 12:41:30] [Rank 0] step:5261/10000 train_time:240007ms step_avg:45.62ms
+[2025-09-04 12:41:30] [Rank 0] step:5261/10000 train_time:240007ms step_avg:45.62ms
+[2025-09-04 12:41:30] [Rank 0] step:5281/10000 train_time:240765ms step_avg:45.59ms
+[2025-09-04 12:41:30] [Rank 0] step:5281/10000 train_time:240765ms step_avg:45.59ms
+[2025-09-04 12:41:31] [Rank 0] step:5301/10000 train_time:241525ms step_avg:45.56ms
+[2025-09-04 12:41:31] [Rank 0] step:5301/10000 train_time:241525ms step_avg:45.56ms
+[2025-09-04 12:41:32] [Rank 0] step:5321/10000 train_time:242284ms step_avg:45.53ms
+[2025-09-04 12:41:32] [Rank 0] step:5321/10000 train_time:242284ms step_avg:45.53ms
+[2025-09-04 12:41:33] [Rank 0] step:5341/10000 train_time:243044ms step_avg:45.51ms
+[2025-09-04 12:41:33] [Rank 0] step:5341/10000 train_time:243044ms step_avg:45.51ms
+[2025-09-04 12:41:33] [Rank 0] step:5361/10000 train_time:243803ms step_avg:45.48ms
+[2025-09-04 12:41:33] [Rank 0] step:5361/10000 train_time:243803ms step_avg:45.48ms
+[2025-09-04 12:41:34] [Rank 0] step:5381/10000 train_time:244562ms step_avg:45.45ms
+[2025-09-04 12:41:34] [Rank 0] step:5381/10000 train_time:244562ms step_avg:45.45ms
+[2025-09-04 12:41:35] [Rank 0] step:5401/10000 train_time:245321ms step_avg:45.42ms
+[2025-09-04 12:41:35] [Rank 0] step:5401/10000 train_time:245321ms step_avg:45.42ms
+[2025-09-04 12:41:36] [Rank 0] step:5421/10000 train_time:246079ms step_avg:45.39ms
+[2025-09-04 12:41:36] [Rank 0] step:5421/10000 train_time:246079ms step_avg:45.39ms
+[2025-09-04 12:41:36] [Rank 0] step:5441/10000 train_time:246839ms step_avg:45.37ms
+[2025-09-04 12:41:36] [Rank 0] step:5441/10000 train_time:246839ms step_avg:45.37ms
+[2025-09-04 12:41:37] [Rank 0] step:5461/10000 train_time:247598ms step_avg:45.34ms
+[2025-09-04 12:41:37] [Rank 0] step:5461/10000 train_time:247598ms step_avg:45.34ms
+[2025-09-04 12:41:38] [Rank 0] step:5481/10000 train_time:248357ms step_avg:45.31ms
+[2025-09-04 12:41:38] [Rank 0] step:5481/10000 train_time:248357ms step_avg:45.31ms
+[2025-09-04 12:41:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:41:39] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:41:39] [Rank 0] PRINT: step:5500/10000 train_loss:0.6506 val_loss:0.6356 train_time:249121ms step_avg:45.29ms
+[2025-09-04 12:41:39] [Rank 0] PRINT: step:5500/10000 train_loss:0.6506 val_loss:0.6356 train_time:249121ms step_avg:45.29ms
+[2025-09-04 12:41:39] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:41:39] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:41:39] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:41:39] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:43:16] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:43:16] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:43:16] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:43:16] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:43:16] [Rank 0]   Total Loss: 4.8730
+[2025-09-04 12:43:16] [Rank 0]   Total Loss: 4.8730
+[2025-09-04 12:43:16] [Rank 0]   Total FTA (Unweighted): 0.9375
+[2025-09-04 12:43:16] [Rank 0]   Total FTA (Unweighted): 0.9375
+[2025-09-04 12:43:16] [Rank 0]   Total FTA (Weighted):   0.9375
+[2025-09-04 12:43:16] [Rank 0]   Total FTA (Weighted):   0.9375
+[2025-09-04 12:43:16] [Rank 0]   Group 0 Loss: 4.8241
+[2025-09-04 12:43:16] [Rank 0]   Group 0 Loss: 4.8241
+[2025-09-04 12:43:17] [Rank 0]   Group 1 Loss: 4.3907
+[2025-09-04 12:43:17] [Rank 0]   Group 1 Loss: 4.3907
+[2025-09-04 12:43:17] [Rank 0]   Group 2 Loss: 4.3140
+[2025-09-04 12:43:17] [Rank 0]   Group 2 Loss: 4.3140
+[2025-09-04 12:43:17] [Rank 0]   Group 3 Loss: 4.8054
+[2025-09-04 12:43:17] [Rank 0]   Group 3 Loss: 4.8054
+[2025-09-04 12:43:17] [Rank 0]   Group 4 Loss: 4.8320
+[2025-09-04 12:43:17] [Rank 0]   Group 4 Loss: 4.8320
+[2025-09-04 12:43:17] [Rank 0]   Group 5 Loss: 4.7840
+[2025-09-04 12:43:17] [Rank 0]   Group 5 Loss: 4.7840
+[2025-09-04 12:43:17] [Rank 0]   Group 6 Loss: 4.7386
+[2025-09-04 12:43:17] [Rank 0]   Group 6 Loss: 4.7386
+[2025-09-04 12:43:17] [Rank 0]   Group 7 Loss: 4.7952
+[2025-09-04 12:43:17] [Rank 0]   Group 7 Loss: 4.7952
+[2025-09-04 12:43:17] [Rank 0]   Group 8 Loss: 4.9601
+[2025-09-04 12:43:17] [Rank 0]   Group 8 Loss: 4.9601
+[2025-09-04 12:43:17] [Rank 0]   Group 9 Loss: 4.9655
+[2025-09-04 12:43:17] [Rank 0]   Group 9 Loss: 4.9655
+[2025-09-04 12:43:17] [Rank 0]   Group 10 Loss: 5.0817
+[2025-09-04 12:43:17] [Rank 0]   Group 10 Loss: 5.0817
+[2025-09-04 12:43:17] [Rank 0]   Group 11 Loss: 5.1321
+[2025-09-04 12:43:17] [Rank 0]   Group 11 Loss: 5.1321
+[2025-09-04 12:43:17] [Rank 0]   Group 12 Loss: 5.0447
+[2025-09-04 12:43:17] [Rank 0]   Group 12 Loss: 5.0447
+[2025-09-04 12:43:17] [Rank 0]   Group 13 Loss: 5.1221
+[2025-09-04 12:43:17] [Rank 0]   Group 13 Loss: 5.1221
+[2025-09-04 12:43:17] [Rank 0]   Group 14 Loss: 5.0662
+[2025-09-04 12:43:17] [Rank 0]   Group 14 Loss: 5.0662
+[2025-09-04 12:43:17] [Rank 0]   Group 15 Loss: 5.1115
+[2025-09-04 12:43:17] [Rank 0]   Group 15 Loss: 5.1115
+[2025-09-04 12:43:17] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:43:17] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 12:43:17] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 12:43:17] [Rank 0]   Group 13 FTA: 0.9600
+[2025-09-04 12:43:17] [Rank 0]   Group 13 FTA: 0.9600
+[2025-09-04 12:43:17] [Rank 0]   Group 14 FTA: 0.7300
+[2025-09-04 12:43:17] [Rank 0]   Group 14 FTA: 0.7300
+[2025-09-04 12:43:17] [Rank 0]   Group 15 FTA: 0.3300
+[2025-09-04 12:43:17] [Rank 0]   Group 15 FTA: 0.3300
+[2025-09-04 12:43:17] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:43:17] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:43:17] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:43:17] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:43:18] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:43:18] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:43:18] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:43:18] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:43:18] [Rank 0] step:5501/10000 train_time:249138ms step_avg:45.29ms
+[2025-09-04 12:43:18] [Rank 0] step:5501/10000 train_time:249138ms step_avg:45.29ms
+[2025-09-04 12:43:19] [Rank 0] step:5521/10000 train_time:249918ms step_avg:45.27ms
+[2025-09-04 12:43:19] [Rank 0] step:5521/10000 train_time:249918ms step_avg:45.27ms
+[2025-09-04 12:43:20] [Rank 0] step:5541/10000 train_time:250677ms step_avg:45.24ms
+[2025-09-04 12:43:20] [Rank 0] step:5541/10000 train_time:250677ms step_avg:45.24ms
+[2025-09-04 12:43:20] [Rank 0] step:5561/10000 train_time:251438ms step_avg:45.21ms
+[2025-09-04 12:43:20] [Rank 0] step:5561/10000 train_time:251438ms step_avg:45.21ms
+[2025-09-04 12:43:21] [Rank 0] step:5581/10000 train_time:252197ms step_avg:45.19ms
+[2025-09-04 12:43:21] [Rank 0] step:5581/10000 train_time:252197ms step_avg:45.19ms
+[2025-09-04 12:43:22] [Rank 0] step:5601/10000 train_time:252957ms step_avg:45.16ms
+[2025-09-04 12:43:22] [Rank 0] step:5601/10000 train_time:252957ms step_avg:45.16ms
+[2025-09-04 12:43:23] [Rank 0] step:5621/10000 train_time:253716ms step_avg:45.14ms
+[2025-09-04 12:43:23] [Rank 0] step:5621/10000 train_time:253716ms step_avg:45.14ms
+[2025-09-04 12:43:23] [Rank 0] step:5641/10000 train_time:254541ms step_avg:45.12ms
+[2025-09-04 12:43:23] [Rank 0] step:5641/10000 train_time:254541ms step_avg:45.12ms
+[2025-09-04 12:43:24] [Rank 0] step:5661/10000 train_time:255300ms step_avg:45.10ms
+[2025-09-04 12:43:24] [Rank 0] step:5661/10000 train_time:255300ms step_avg:45.10ms
+[2025-09-04 12:43:25] [Rank 0] step:5681/10000 train_time:256060ms step_avg:45.07ms
+[2025-09-04 12:43:25] [Rank 0] step:5681/10000 train_time:256060ms step_avg:45.07ms
+[2025-09-04 12:43:26] [Rank 0] step:5701/10000 train_time:256819ms step_avg:45.05ms
+[2025-09-04 12:43:26] [Rank 0] step:5701/10000 train_time:256819ms step_avg:45.05ms
+[2025-09-04 12:43:27] [Rank 0] step:5721/10000 train_time:257579ms step_avg:45.02ms
+[2025-09-04 12:43:27] [Rank 0] step:5721/10000 train_time:257579ms step_avg:45.02ms
+[2025-09-04 12:43:27] [Rank 0] step:5741/10000 train_time:258338ms step_avg:45.00ms
+[2025-09-04 12:43:27] [Rank 0] step:5741/10000 train_time:258338ms step_avg:45.00ms
+[2025-09-04 12:43:28] [Rank 0] step:5761/10000 train_time:259098ms step_avg:44.97ms
+[2025-09-04 12:43:28] [Rank 0] step:5761/10000 train_time:259098ms step_avg:44.97ms
+[2025-09-04 12:43:29] [Rank 0] step:5781/10000 train_time:259857ms step_avg:44.95ms
+[2025-09-04 12:43:29] [Rank 0] step:5781/10000 train_time:259857ms step_avg:44.95ms
+[2025-09-04 12:43:30] [Rank 0] step:5801/10000 train_time:260616ms step_avg:44.93ms
+[2025-09-04 12:43:30] [Rank 0] step:5801/10000 train_time:260616ms step_avg:44.93ms
+[2025-09-04 12:43:30] [Rank 0] step:5821/10000 train_time:261376ms step_avg:44.90ms
+[2025-09-04 12:43:30] [Rank 0] step:5821/10000 train_time:261376ms step_avg:44.90ms
+[2025-09-04 12:43:31] [Rank 0] step:5841/10000 train_time:262135ms step_avg:44.88ms
+[2025-09-04 12:43:31] [Rank 0] step:5841/10000 train_time:262135ms step_avg:44.88ms
+[2025-09-04 12:43:32] [Rank 0] step:5861/10000 train_time:263112ms step_avg:44.89ms
+[2025-09-04 12:43:32] [Rank 0] step:5861/10000 train_time:263112ms step_avg:44.89ms
+[2025-09-04 12:43:33] [Rank 0] step:5881/10000 train_time:263871ms step_avg:44.87ms
+[2025-09-04 12:43:33] [Rank 0] step:5881/10000 train_time:263871ms step_avg:44.87ms
+[2025-09-04 12:43:34] [Rank 0] step:5901/10000 train_time:264633ms step_avg:44.85ms
+[2025-09-04 12:43:34] [Rank 0] step:5901/10000 train_time:264633ms step_avg:44.85ms
+[2025-09-04 12:43:35] [Rank 0] step:5921/10000 train_time:265685ms step_avg:44.87ms
+[2025-09-04 12:43:35] [Rank 0] step:5921/10000 train_time:265685ms step_avg:44.87ms
+[2025-09-04 12:43:35] [Rank 0] step:5941/10000 train_time:266444ms step_avg:44.85ms
+[2025-09-04 12:43:35] [Rank 0] step:5941/10000 train_time:266444ms step_avg:44.85ms
+[2025-09-04 12:43:36] [Rank 0] step:5961/10000 train_time:267203ms step_avg:44.83ms
+[2025-09-04 12:43:36] [Rank 0] step:5961/10000 train_time:267203ms step_avg:44.83ms
+[2025-09-04 12:43:37] [Rank 0] step:5981/10000 train_time:267963ms step_avg:44.80ms
+[2025-09-04 12:43:37] [Rank 0] step:5981/10000 train_time:267963ms step_avg:44.80ms
+[2025-09-04 12:43:38] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:43:38] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:43:38] [Rank 0] PRINT: step:6000/10000 train_loss:0.6435 val_loss:0.6294 train_time:268727ms step_avg:44.79ms
+[2025-09-04 12:43:38] [Rank 0] PRINT: step:6000/10000 train_loss:0.6435 val_loss:0.6294 train_time:268727ms step_avg:44.79ms
+[2025-09-04 12:43:38] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:43:38] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:43:38] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:43:38] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:45:15] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:45:15] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:45:15] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:45:15] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:45:15] [Rank 0]   Total Loss: 4.9391
+[2025-09-04 12:45:15] [Rank 0]   Total Loss: 4.9391
+[2025-09-04 12:45:15] [Rank 0]   Total FTA (Unweighted): 0.9475
+[2025-09-04 12:45:15] [Rank 0]   Total FTA (Unweighted): 0.9475
+[2025-09-04 12:45:15] [Rank 0]   Total FTA (Weighted):   0.9475
+[2025-09-04 12:45:15] [Rank 0]   Total FTA (Weighted):   0.9475
+[2025-09-04 12:45:15] [Rank 0]   Group 0 Loss: 4.9184
+[2025-09-04 12:45:15] [Rank 0]   Group 0 Loss: 4.9184
+[2025-09-04 12:45:15] [Rank 0]   Group 1 Loss: 4.4504
+[2025-09-04 12:45:15] [Rank 0]   Group 1 Loss: 4.4504
+[2025-09-04 12:45:15] [Rank 0]   Group 2 Loss: 4.3882
+[2025-09-04 12:45:15] [Rank 0]   Group 2 Loss: 4.3882
+[2025-09-04 12:45:15] [Rank 0]   Group 3 Loss: 4.8902
+[2025-09-04 12:45:15] [Rank 0]   Group 3 Loss: 4.8902
+[2025-09-04 12:45:15] [Rank 0]   Group 4 Loss: 4.8915
+[2025-09-04 12:45:15] [Rank 0]   Group 4 Loss: 4.8915
+[2025-09-04 12:45:15] [Rank 0]   Group 5 Loss: 4.8909
+[2025-09-04 12:45:15] [Rank 0]   Group 5 Loss: 4.8909
+[2025-09-04 12:45:15] [Rank 0]   Group 6 Loss: 4.8120
+[2025-09-04 12:45:15] [Rank 0]   Group 6 Loss: 4.8120
+[2025-09-04 12:45:15] [Rank 0]   Group 7 Loss: 4.8626
+[2025-09-04 12:45:15] [Rank 0]   Group 7 Loss: 4.8626
+[2025-09-04 12:45:15] [Rank 0]   Group 8 Loss: 5.0063
+[2025-09-04 12:45:15] [Rank 0]   Group 8 Loss: 5.0063
+[2025-09-04 12:45:15] [Rank 0]   Group 9 Loss: 5.0252
+[2025-09-04 12:45:15] [Rank 0]   Group 9 Loss: 5.0252
+[2025-09-04 12:45:15] [Rank 0]   Group 10 Loss: 5.1155
+[2025-09-04 12:45:15] [Rank 0]   Group 10 Loss: 5.1155
+[2025-09-04 12:45:15] [Rank 0]   Group 11 Loss: 5.1689
+[2025-09-04 12:45:15] [Rank 0]   Group 11 Loss: 5.1689
+[2025-09-04 12:45:15] [Rank 0]   Group 12 Loss: 5.0856
+[2025-09-04 12:45:15] [Rank 0]   Group 12 Loss: 5.0856
+[2025-09-04 12:45:15] [Rank 0]   Group 13 Loss: 5.2071
+[2025-09-04 12:45:15] [Rank 0]   Group 13 Loss: 5.2071
+[2025-09-04 12:45:15] [Rank 0]   Group 14 Loss: 5.1680
+[2025-09-04 12:45:15] [Rank 0]   Group 14 Loss: 5.1680
+[2025-09-04 12:45:15] [Rank 0]   Group 15 Loss: 5.1455
+[2025-09-04 12:45:15] [Rank 0]   Group 15 Loss: 5.1455
+[2025-09-04 12:45:15] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:45:15] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:45:15] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 12:45:15] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:45:15] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:45:15] [Rank 0]   Group 13 FTA: 0.9600
+[2025-09-04 12:45:15] [Rank 0]   Group 13 FTA: 0.9600
+[2025-09-04 12:45:15] [Rank 0]   Group 14 FTA: 0.7800
+[2025-09-04 12:45:15] [Rank 0]   Group 14 FTA: 0.7800
+[2025-09-04 12:45:15] [Rank 0]   Group 15 FTA: 0.4400
+[2025-09-04 12:45:15] [Rank 0]   Group 15 FTA: 0.4400
+[2025-09-04 12:45:16] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:45:16] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:45:16] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:45:16] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:45:16] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:45:16] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:45:17] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:45:17] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:45:17] [Rank 0] step:6001/10000 train_time:268743ms step_avg:44.78ms
+[2025-09-04 12:45:17] [Rank 0] step:6001/10000 train_time:268743ms step_avg:44.78ms
+[2025-09-04 12:45:18] [Rank 0] step:6021/10000 train_time:269778ms step_avg:44.81ms
+[2025-09-04 12:45:18] [Rank 0] step:6021/10000 train_time:269778ms step_avg:44.81ms
+[2025-09-04 12:45:18] [Rank 0] step:6041/10000 train_time:270538ms step_avg:44.78ms
+[2025-09-04 12:45:18] [Rank 0] step:6041/10000 train_time:270538ms step_avg:44.78ms
+[2025-09-04 12:45:19] [Rank 0] step:6061/10000 train_time:271297ms step_avg:44.76ms
+[2025-09-04 12:45:19] [Rank 0] step:6061/10000 train_time:271297ms step_avg:44.76ms
+[2025-09-04 12:45:20] [Rank 0] step:6081/10000 train_time:272056ms step_avg:44.74ms
+[2025-09-04 12:45:20] [Rank 0] step:6081/10000 train_time:272056ms step_avg:44.74ms
+[2025-09-04 12:45:21] [Rank 0] step:6101/10000 train_time:272815ms step_avg:44.72ms
+[2025-09-04 12:45:21] [Rank 0] step:6101/10000 train_time:272815ms step_avg:44.72ms
+[2025-09-04 12:45:21] [Rank 0] step:6121/10000 train_time:273575ms step_avg:44.69ms
+[2025-09-04 12:45:21] [Rank 0] step:6121/10000 train_time:273575ms step_avg:44.69ms
+[2025-09-04 12:45:22] [Rank 0] step:6141/10000 train_time:274335ms step_avg:44.67ms
+[2025-09-04 12:45:22] [Rank 0] step:6141/10000 train_time:274335ms step_avg:44.67ms
+[2025-09-04 12:45:23] [Rank 0] step:6161/10000 train_time:275095ms step_avg:44.65ms
+[2025-09-04 12:45:23] [Rank 0] step:6161/10000 train_time:275095ms step_avg:44.65ms
+[2025-09-04 12:45:24] [Rank 0] step:6181/10000 train_time:275855ms step_avg:44.63ms
+[2025-09-04 12:45:24] [Rank 0] step:6181/10000 train_time:275855ms step_avg:44.63ms
+[2025-09-04 12:45:24] [Rank 0] step:6201/10000 train_time:276615ms step_avg:44.61ms
+[2025-09-04 12:45:24] [Rank 0] step:6201/10000 train_time:276615ms step_avg:44.61ms
+[2025-09-04 12:45:25] [Rank 0] step:6221/10000 train_time:277373ms step_avg:44.59ms
+[2025-09-04 12:45:25] [Rank 0] step:6221/10000 train_time:277373ms step_avg:44.59ms
+[2025-09-04 12:45:26] [Rank 0] step:6241/10000 train_time:278132ms step_avg:44.57ms
+[2025-09-04 12:45:26] [Rank 0] step:6241/10000 train_time:278132ms step_avg:44.57ms
+[2025-09-04 12:45:27] [Rank 0] step:6261/10000 train_time:278891ms step_avg:44.54ms
+[2025-09-04 12:45:27] [Rank 0] step:6261/10000 train_time:278891ms step_avg:44.54ms
+[2025-09-04 12:45:28] [Rank 0] step:6281/10000 train_time:279651ms step_avg:44.52ms
+[2025-09-04 12:45:28] [Rank 0] step:6281/10000 train_time:279651ms step_avg:44.52ms
+[2025-09-04 12:45:28] [Rank 0] step:6301/10000 train_time:280410ms step_avg:44.50ms
+[2025-09-04 12:45:28] [Rank 0] step:6301/10000 train_time:280410ms step_avg:44.50ms
+[2025-09-04 12:45:29] [Rank 0] step:6321/10000 train_time:281169ms step_avg:44.48ms
+[2025-09-04 12:45:29] [Rank 0] step:6321/10000 train_time:281169ms step_avg:44.48ms
+[2025-09-04 12:45:30] [Rank 0] step:6341/10000 train_time:281928ms step_avg:44.46ms
+[2025-09-04 12:45:30] [Rank 0] step:6341/10000 train_time:281928ms step_avg:44.46ms
+[2025-09-04 12:45:31] [Rank 0] step:6361/10000 train_time:282687ms step_avg:44.44ms
+[2025-09-04 12:45:31] [Rank 0] step:6361/10000 train_time:282687ms step_avg:44.44ms
+[2025-09-04 12:45:31] [Rank 0] step:6381/10000 train_time:283447ms step_avg:44.42ms
+[2025-09-04 12:45:31] [Rank 0] step:6381/10000 train_time:283447ms step_avg:44.42ms
+[2025-09-04 12:45:32] [Rank 0] step:6401/10000 train_time:284206ms step_avg:44.40ms
+[2025-09-04 12:45:32] [Rank 0] step:6401/10000 train_time:284206ms step_avg:44.40ms
+[2025-09-04 12:45:33] [Rank 0] step:6421/10000 train_time:284966ms step_avg:44.38ms
+[2025-09-04 12:45:33] [Rank 0] step:6421/10000 train_time:284966ms step_avg:44.38ms
+[2025-09-04 12:45:34] [Rank 0] step:6441/10000 train_time:285725ms step_avg:44.36ms
+[2025-09-04 12:45:34] [Rank 0] step:6441/10000 train_time:285725ms step_avg:44.36ms
+[2025-09-04 12:45:34] [Rank 0] step:6461/10000 train_time:286484ms step_avg:44.34ms
+[2025-09-04 12:45:34] [Rank 0] step:6461/10000 train_time:286484ms step_avg:44.34ms
+[2025-09-04 12:45:35] [Rank 0] step:6481/10000 train_time:287243ms step_avg:44.32ms
+[2025-09-04 12:45:35] [Rank 0] step:6481/10000 train_time:287243ms step_avg:44.32ms
+[2025-09-04 12:45:36] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:45:36] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:45:36] [Rank 0] PRINT: step:6500/10000 train_loss:0.6372 val_loss:0.6240 train_time:288008ms step_avg:44.31ms
+[2025-09-04 12:45:36] [Rank 0] PRINT: step:6500/10000 train_loss:0.6372 val_loss:0.6240 train_time:288008ms step_avg:44.31ms
+[2025-09-04 12:45:36] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:45:36] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:45:37] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:45:37] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:47:13] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:47:13] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:47:13] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:47:13] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:47:13] [Rank 0]   Total Loss: 4.9721
+[2025-09-04 12:47:13] [Rank 0]   Total Loss: 4.9721
+[2025-09-04 12:47:13] [Rank 0]   Total FTA (Unweighted): 0.9569
+[2025-09-04 12:47:13] [Rank 0]   Total FTA (Unweighted): 0.9569
+[2025-09-04 12:47:13] [Rank 0]   Total FTA (Weighted):   0.9569
+[2025-09-04 12:47:13] [Rank 0]   Total FTA (Weighted):   0.9569
+[2025-09-04 12:47:13] [Rank 0]   Group 0 Loss: 4.8922
+[2025-09-04 12:47:13] [Rank 0]   Group 0 Loss: 4.8922
+[2025-09-04 12:47:14] [Rank 0]   Group 1 Loss: 4.5321
+[2025-09-04 12:47:14] [Rank 0]   Group 1 Loss: 4.5321
+[2025-09-04 12:47:14] [Rank 0]   Group 2 Loss: 4.3744
+[2025-09-04 12:47:14] [Rank 0]   Group 2 Loss: 4.3744
+[2025-09-04 12:47:14] [Rank 0]   Group 3 Loss: 4.9087
+[2025-09-04 12:47:14] [Rank 0]   Group 3 Loss: 4.9087
+[2025-09-04 12:47:14] [Rank 0]   Group 4 Loss: 4.9256
+[2025-09-04 12:47:14] [Rank 0]   Group 4 Loss: 4.9256
+[2025-09-04 12:47:14] [Rank 0]   Group 5 Loss: 4.8859
+[2025-09-04 12:47:14] [Rank 0]   Group 5 Loss: 4.8859
+[2025-09-04 12:47:14] [Rank 0]   Group 6 Loss: 4.8356
+[2025-09-04 12:47:14] [Rank 0]   Group 6 Loss: 4.8356
+[2025-09-04 12:47:14] [Rank 0]   Group 7 Loss: 4.8919
+[2025-09-04 12:47:14] [Rank 0]   Group 7 Loss: 4.8919
+[2025-09-04 12:47:14] [Rank 0]   Group 8 Loss: 5.0286
+[2025-09-04 12:47:14] [Rank 0]   Group 8 Loss: 5.0286
+[2025-09-04 12:47:14] [Rank 0]   Group 9 Loss: 5.0320
+[2025-09-04 12:47:14] [Rank 0]   Group 9 Loss: 5.0320
+[2025-09-04 12:47:14] [Rank 0]   Group 10 Loss: 5.1604
+[2025-09-04 12:47:14] [Rank 0]   Group 10 Loss: 5.1604
+[2025-09-04 12:47:14] [Rank 0]   Group 11 Loss: 5.2636
+[2025-09-04 12:47:14] [Rank 0]   Group 11 Loss: 5.2636
+[2025-09-04 12:47:14] [Rank 0]   Group 12 Loss: 5.1620
+[2025-09-04 12:47:14] [Rank 0]   Group 12 Loss: 5.1620
+[2025-09-04 12:47:14] [Rank 0]   Group 13 Loss: 5.2537
+[2025-09-04 12:47:14] [Rank 0]   Group 13 Loss: 5.2537
+[2025-09-04 12:47:14] [Rank 0]   Group 14 Loss: 5.2168
+[2025-09-04 12:47:14] [Rank 0]   Group 14 Loss: 5.2168
+[2025-09-04 12:47:14] [Rank 0]   Group 15 Loss: 5.1900
+[2025-09-04 12:47:14] [Rank 0]   Group 15 Loss: 5.1900
+[2025-09-04 12:47:14] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:47:14] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:47:14] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:47:14] [Rank 0]   Group 14 FTA: 0.8600
+[2025-09-04 12:47:14] [Rank 0]   Group 14 FTA: 0.8600
+[2025-09-04 12:47:14] [Rank 0]   Group 15 FTA: 0.4600
+[2025-09-04 12:47:14] [Rank 0]   Group 15 FTA: 0.4600
+[2025-09-04 12:47:14] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:47:14] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:47:15] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:47:15] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:47:15] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:47:15] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:47:15] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:47:15] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:47:15] [Rank 0] step:6501/10000 train_time:288024ms step_avg:44.30ms
+[2025-09-04 12:47:15] [Rank 0] step:6501/10000 train_time:288024ms step_avg:44.30ms
+[2025-09-04 12:47:16] [Rank 0] step:6521/10000 train_time:288796ms step_avg:44.29ms
+[2025-09-04 12:47:16] [Rank 0] step:6521/10000 train_time:288796ms step_avg:44.29ms
+[2025-09-04 12:47:17] [Rank 0] step:6541/10000 train_time:289555ms step_avg:44.27ms
+[2025-09-04 12:47:17] [Rank 0] step:6541/10000 train_time:289555ms step_avg:44.27ms
+[2025-09-04 12:47:17] [Rank 0] step:6561/10000 train_time:290314ms step_avg:44.25ms
+[2025-09-04 12:47:17] [Rank 0] step:6561/10000 train_time:290314ms step_avg:44.25ms
+[2025-09-04 12:47:18] [Rank 0] step:6581/10000 train_time:291073ms step_avg:44.23ms
+[2025-09-04 12:47:18] [Rank 0] step:6581/10000 train_time:291073ms step_avg:44.23ms
+[2025-09-04 12:47:19] [Rank 0] step:6601/10000 train_time:291832ms step_avg:44.21ms
+[2025-09-04 12:47:19] [Rank 0] step:6601/10000 train_time:291832ms step_avg:44.21ms
+[2025-09-04 12:47:20] [Rank 0] step:6621/10000 train_time:292591ms step_avg:44.19ms
+[2025-09-04 12:47:20] [Rank 0] step:6621/10000 train_time:292591ms step_avg:44.19ms
+[2025-09-04 12:47:20] [Rank 0] step:6641/10000 train_time:293350ms step_avg:44.17ms
+[2025-09-04 12:47:20] [Rank 0] step:6641/10000 train_time:293350ms step_avg:44.17ms
+[2025-09-04 12:47:21] [Rank 0] step:6661/10000 train_time:294110ms step_avg:44.15ms
+[2025-09-04 12:47:21] [Rank 0] step:6661/10000 train_time:294110ms step_avg:44.15ms
+[2025-09-04 12:47:22] [Rank 0] step:6681/10000 train_time:294869ms step_avg:44.14ms
+[2025-09-04 12:47:22] [Rank 0] step:6681/10000 train_time:294869ms step_avg:44.14ms
+[2025-09-04 12:47:23] [Rank 0] step:6701/10000 train_time:295628ms step_avg:44.12ms
+[2025-09-04 12:47:23] [Rank 0] step:6701/10000 train_time:295628ms step_avg:44.12ms
+[2025-09-04 12:47:23] [Rank 0] step:6721/10000 train_time:296387ms step_avg:44.10ms
+[2025-09-04 12:47:23] [Rank 0] step:6721/10000 train_time:296387ms step_avg:44.10ms
+[2025-09-04 12:47:24] [Rank 0] step:6741/10000 train_time:297147ms step_avg:44.08ms
+[2025-09-04 12:47:24] [Rank 0] step:6741/10000 train_time:297147ms step_avg:44.08ms
+[2025-09-04 12:47:25] [Rank 0] step:6761/10000 train_time:297906ms step_avg:44.06ms
+[2025-09-04 12:47:25] [Rank 0] step:6761/10000 train_time:297906ms step_avg:44.06ms
+[2025-09-04 12:47:26] [Rank 0] step:6781/10000 train_time:298669ms step_avg:44.04ms
+[2025-09-04 12:47:26] [Rank 0] step:6781/10000 train_time:298669ms step_avg:44.04ms
+[2025-09-04 12:47:27] [Rank 0] step:6801/10000 train_time:299428ms step_avg:44.03ms
+[2025-09-04 12:47:27] [Rank 0] step:6801/10000 train_time:299428ms step_avg:44.03ms
+[2025-09-04 12:47:27] [Rank 0] step:6821/10000 train_time:300188ms step_avg:44.01ms
+[2025-09-04 12:47:27] [Rank 0] step:6821/10000 train_time:300188ms step_avg:44.01ms
+[2025-09-04 12:47:28] [Rank 0] step:6841/10000 train_time:301218ms step_avg:44.03ms
+[2025-09-04 12:47:28] [Rank 0] step:6841/10000 train_time:301218ms step_avg:44.03ms
+[2025-09-04 12:47:29] [Rank 0] step:6861/10000 train_time:301983ms step_avg:44.01ms
+[2025-09-04 12:47:29] [Rank 0] step:6861/10000 train_time:301983ms step_avg:44.01ms
+[2025-09-04 12:47:30] [Rank 0] step:6881/10000 train_time:302742ms step_avg:44.00ms
+[2025-09-04 12:47:30] [Rank 0] step:6881/10000 train_time:302742ms step_avg:44.00ms
+[2025-09-04 12:47:31] [Rank 0] step:6901/10000 train_time:303501ms step_avg:43.98ms
+[2025-09-04 12:47:31] [Rank 0] step:6901/10000 train_time:303501ms step_avg:43.98ms
+[2025-09-04 12:47:31] [Rank 0] step:6921/10000 train_time:304261ms step_avg:43.96ms
+[2025-09-04 12:47:31] [Rank 0] step:6921/10000 train_time:304261ms step_avg:43.96ms
+[2025-09-04 12:47:32] [Rank 0] step:6941/10000 train_time:305019ms step_avg:43.94ms
+[2025-09-04 12:47:32] [Rank 0] step:6941/10000 train_time:305019ms step_avg:43.94ms
+[2025-09-04 12:47:33] [Rank 0] step:6961/10000 train_time:305779ms step_avg:43.93ms
+[2025-09-04 12:47:33] [Rank 0] step:6961/10000 train_time:305779ms step_avg:43.93ms
+[2025-09-04 12:47:34] [Rank 0] step:6981/10000 train_time:306541ms step_avg:43.91ms
+[2025-09-04 12:47:34] [Rank 0] step:6981/10000 train_time:306541ms step_avg:43.91ms
+[2025-09-04 12:47:34] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:47:34] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:47:35] [Rank 0] PRINT: step:7000/10000 train_loss:0.6310 val_loss:0.6194 train_time:307304ms step_avg:43.90ms
+[2025-09-04 12:47:35] [Rank 0] PRINT: step:7000/10000 train_loss:0.6310 val_loss:0.6194 train_time:307304ms step_avg:43.90ms
+[2025-09-04 12:47:35] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:47:35] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:47:35] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:47:35] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:49:12] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:49:12] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:49:12] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:49:12] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:49:12] [Rank 0]   Total Loss: 4.9564
+[2025-09-04 12:49:12] [Rank 0]   Total Loss: 4.9564
+[2025-09-04 12:49:12] [Rank 0]   Total FTA (Unweighted): 0.9656
+[2025-09-04 12:49:12] [Rank 0]   Total FTA (Unweighted): 0.9656
+[2025-09-04 12:49:12] [Rank 0]   Total FTA (Weighted):   0.9656
+[2025-09-04 12:49:12] [Rank 0]   Total FTA (Weighted):   0.9656
+[2025-09-04 12:49:12] [Rank 0]   Group 0 Loss: 4.8715
+[2025-09-04 12:49:12] [Rank 0]   Group 0 Loss: 4.8715
+[2025-09-04 12:49:12] [Rank 0]   Group 1 Loss: 4.5105
+[2025-09-04 12:49:12] [Rank 0]   Group 1 Loss: 4.5105
+[2025-09-04 12:49:12] [Rank 0]   Group 2 Loss: 4.4174
+[2025-09-04 12:49:12] [Rank 0]   Group 2 Loss: 4.4174
+[2025-09-04 12:49:12] [Rank 0]   Group 3 Loss: 4.8974
+[2025-09-04 12:49:12] [Rank 0]   Group 3 Loss: 4.8974
+[2025-09-04 12:49:12] [Rank 0]   Group 4 Loss: 4.8821
+[2025-09-04 12:49:12] [Rank 0]   Group 4 Loss: 4.8821
+[2025-09-04 12:49:12] [Rank 0]   Group 5 Loss: 4.8842
+[2025-09-04 12:49:12] [Rank 0]   Group 5 Loss: 4.8842
+[2025-09-04 12:49:12] [Rank 0]   Group 6 Loss: 4.8398
+[2025-09-04 12:49:12] [Rank 0]   Group 6 Loss: 4.8398
+[2025-09-04 12:49:12] [Rank 0]   Group 7 Loss: 4.8742
+[2025-09-04 12:49:12] [Rank 0]   Group 7 Loss: 4.8742
+[2025-09-04 12:49:12] [Rank 0]   Group 8 Loss: 5.0037
+[2025-09-04 12:49:12] [Rank 0]   Group 8 Loss: 5.0037
+[2025-09-04 12:49:12] [Rank 0]   Group 9 Loss: 5.0339
+[2025-09-04 12:49:12] [Rank 0]   Group 9 Loss: 5.0339
+[2025-09-04 12:49:12] [Rank 0]   Group 10 Loss: 5.1556
+[2025-09-04 12:49:12] [Rank 0]   Group 10 Loss: 5.1556
+[2025-09-04 12:49:12] [Rank 0]   Group 11 Loss: 5.2092
+[2025-09-04 12:49:12] [Rank 0]   Group 11 Loss: 5.2092
+[2025-09-04 12:49:12] [Rank 0]   Group 12 Loss: 5.1247
+[2025-09-04 12:49:12] [Rank 0]   Group 12 Loss: 5.1247
+[2025-09-04 12:49:12] [Rank 0]   Group 13 Loss: 5.2412
+[2025-09-04 12:49:12] [Rank 0]   Group 13 Loss: 5.2412
+[2025-09-04 12:49:12] [Rank 0]   Group 14 Loss: 5.1956
+[2025-09-04 12:49:12] [Rank 0]   Group 14 Loss: 5.1956
+[2025-09-04 12:49:12] [Rank 0]   Group 15 Loss: 5.1621
+[2025-09-04 12:49:12] [Rank 0]   Group 15 Loss: 5.1621
+[2025-09-04 12:49:12] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:49:12] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 12:49:12] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 12:49:12] [Rank 0]   Group 14 FTA: 0.8900
+[2025-09-04 12:49:12] [Rank 0]   Group 14 FTA: 0.8900
+[2025-09-04 12:49:12] [Rank 0]   Group 15 FTA: 0.5700
+[2025-09-04 12:49:12] [Rank 0]   Group 15 FTA: 0.5700
+[2025-09-04 12:49:12] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:49:12] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:49:13] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:49:13] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:49:13] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:49:13] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:49:13] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:49:13] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:49:13] [Rank 0] step:7001/10000 train_time:307321ms step_avg:43.90ms
+[2025-09-04 12:49:13] [Rank 0] step:7001/10000 train_time:307321ms step_avg:43.90ms
+[2025-09-04 12:49:14] [Rank 0] step:7021/10000 train_time:308091ms step_avg:43.88ms
+[2025-09-04 12:49:14] [Rank 0] step:7021/10000 train_time:308091ms step_avg:43.88ms
+[2025-09-04 12:49:15] [Rank 0] step:7041/10000 train_time:308849ms step_avg:43.86ms
+[2025-09-04 12:49:15] [Rank 0] step:7041/10000 train_time:308849ms step_avg:43.86ms
+[2025-09-04 12:49:16] [Rank 0] step:7061/10000 train_time:309608ms step_avg:43.85ms
+[2025-09-04 12:49:16] [Rank 0] step:7061/10000 train_time:309608ms step_avg:43.85ms
+[2025-09-04 12:49:16] [Rank 0] step:7081/10000 train_time:310367ms step_avg:43.83ms
+[2025-09-04 12:49:16] [Rank 0] step:7081/10000 train_time:310367ms step_avg:43.83ms
+[2025-09-04 12:49:17] [Rank 0] step:7101/10000 train_time:311125ms step_avg:43.81ms
+[2025-09-04 12:49:17] [Rank 0] step:7101/10000 train_time:311125ms step_avg:43.81ms
+[2025-09-04 12:49:18] [Rank 0] step:7121/10000 train_time:311884ms step_avg:43.80ms
+[2025-09-04 12:49:18] [Rank 0] step:7121/10000 train_time:311884ms step_avg:43.80ms
+[2025-09-04 12:49:19] [Rank 0] step:7141/10000 train_time:312643ms step_avg:43.78ms
+[2025-09-04 12:49:19] [Rank 0] step:7141/10000 train_time:312643ms step_avg:43.78ms
+[2025-09-04 12:49:19] [Rank 0] step:7161/10000 train_time:313402ms step_avg:43.77ms
+[2025-09-04 12:49:19] [Rank 0] step:7161/10000 train_time:313402ms step_avg:43.77ms
+[2025-09-04 12:49:20] [Rank 0] step:7181/10000 train_time:314162ms step_avg:43.75ms
+[2025-09-04 12:49:20] [Rank 0] step:7181/10000 train_time:314162ms step_avg:43.75ms
+[2025-09-04 12:49:21] [Rank 0] step:7201/10000 train_time:314921ms step_avg:43.73ms
+[2025-09-04 12:49:21] [Rank 0] step:7201/10000 train_time:314921ms step_avg:43.73ms
+[2025-09-04 12:49:22] [Rank 0] step:7221/10000 train_time:315680ms step_avg:43.72ms
+[2025-09-04 12:49:22] [Rank 0] step:7221/10000 train_time:315680ms step_avg:43.72ms
+[2025-09-04 12:49:22] [Rank 0] step:7241/10000 train_time:316439ms step_avg:43.70ms
+[2025-09-04 12:49:22] [Rank 0] step:7241/10000 train_time:316439ms step_avg:43.70ms
+[2025-09-04 12:49:23] [Rank 0] step:7261/10000 train_time:317198ms step_avg:43.69ms
+[2025-09-04 12:49:23] [Rank 0] step:7261/10000 train_time:317198ms step_avg:43.69ms
+[2025-09-04 12:49:24] [Rank 0] step:7281/10000 train_time:317958ms step_avg:43.67ms
+[2025-09-04 12:49:24] [Rank 0] step:7281/10000 train_time:317958ms step_avg:43.67ms
+[2025-09-04 12:49:25] [Rank 0] step:7301/10000 train_time:318717ms step_avg:43.65ms
+[2025-09-04 12:49:25] [Rank 0] step:7301/10000 train_time:318717ms step_avg:43.65ms
+[2025-09-04 12:49:25] [Rank 0] step:7321/10000 train_time:319476ms step_avg:43.64ms
+[2025-09-04 12:49:25] [Rank 0] step:7321/10000 train_time:319476ms step_avg:43.64ms
+[2025-09-04 12:49:26] [Rank 0] step:7341/10000 train_time:320235ms step_avg:43.62ms
+[2025-09-04 12:49:26] [Rank 0] step:7341/10000 train_time:320235ms step_avg:43.62ms
+[2025-09-04 12:49:27] [Rank 0] step:7361/10000 train_time:320993ms step_avg:43.61ms
+[2025-09-04 12:49:27] [Rank 0] step:7361/10000 train_time:320993ms step_avg:43.61ms
+[2025-09-04 12:49:28] [Rank 0] step:7381/10000 train_time:321752ms step_avg:43.59ms
+[2025-09-04 12:49:28] [Rank 0] step:7381/10000 train_time:321752ms step_avg:43.59ms
+[2025-09-04 12:49:29] [Rank 0] step:7401/10000 train_time:322511ms step_avg:43.58ms
+[2025-09-04 12:49:29] [Rank 0] step:7401/10000 train_time:322511ms step_avg:43.58ms
+[2025-09-04 12:49:29] [Rank 0] step:7421/10000 train_time:323270ms step_avg:43.56ms
+[2025-09-04 12:49:29] [Rank 0] step:7421/10000 train_time:323270ms step_avg:43.56ms
+[2025-09-04 12:49:30] [Rank 0] step:7441/10000 train_time:324028ms step_avg:43.55ms
+[2025-09-04 12:49:30] [Rank 0] step:7441/10000 train_time:324028ms step_avg:43.55ms
+[2025-09-04 12:49:31] [Rank 0] step:7461/10000 train_time:324787ms step_avg:43.53ms
+[2025-09-04 12:49:31] [Rank 0] step:7461/10000 train_time:324787ms step_avg:43.53ms
+[2025-09-04 12:49:32] [Rank 0] step:7481/10000 train_time:325545ms step_avg:43.52ms
+[2025-09-04 12:49:32] [Rank 0] step:7481/10000 train_time:325545ms step_avg:43.52ms
+[2025-09-04 12:49:32] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:49:32] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:49:33] [Rank 0] PRINT: step:7500/10000 train_loss:0.6257 val_loss:0.6156 train_time:326308ms step_avg:43.51ms
+[2025-09-04 12:49:33] [Rank 0] PRINT: step:7500/10000 train_loss:0.6257 val_loss:0.6156 train_time:326308ms step_avg:43.51ms
+[2025-09-04 12:49:33] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:49:33] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:49:33] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:49:33] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:51:10] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:51:10] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:51:10] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:51:10] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:51:10] [Rank 0]   Total Loss: 5.0044
+[2025-09-04 12:51:10] [Rank 0]   Total Loss: 5.0044
+[2025-09-04 12:51:10] [Rank 0]   Total FTA (Unweighted): 0.9681
+[2025-09-04 12:51:10] [Rank 0]   Total FTA (Unweighted): 0.9681
+[2025-09-04 12:51:10] [Rank 0]   Total FTA (Weighted):   0.9681
+[2025-09-04 12:51:10] [Rank 0]   Total FTA (Weighted):   0.9681
+[2025-09-04 12:51:10] [Rank 0]   Group 0 Loss: 4.9098
+[2025-09-04 12:51:10] [Rank 0]   Group 0 Loss: 4.9098
+[2025-09-04 12:51:10] [Rank 0]   Group 1 Loss: 4.5175
+[2025-09-04 12:51:10] [Rank 0]   Group 1 Loss: 4.5175
+[2025-09-04 12:51:10] [Rank 0]   Group 2 Loss: 4.4660
+[2025-09-04 12:51:10] [Rank 0]   Group 2 Loss: 4.4660
+[2025-09-04 12:51:10] [Rank 0]   Group 3 Loss: 4.9276
+[2025-09-04 12:51:10] [Rank 0]   Group 3 Loss: 4.9276
+[2025-09-04 12:51:10] [Rank 0]   Group 4 Loss: 4.9375
+[2025-09-04 12:51:10] [Rank 0]   Group 4 Loss: 4.9375
+[2025-09-04 12:51:10] [Rank 0]   Group 5 Loss: 4.9225
+[2025-09-04 12:51:10] [Rank 0]   Group 5 Loss: 4.9225
+[2025-09-04 12:51:10] [Rank 0]   Group 6 Loss: 4.8417
+[2025-09-04 12:51:10] [Rank 0]   Group 6 Loss: 4.8417
+[2025-09-04 12:51:10] [Rank 0]   Group 7 Loss: 4.9252
+[2025-09-04 12:51:10] [Rank 0]   Group 7 Loss: 4.9252
+[2025-09-04 12:51:10] [Rank 0]   Group 8 Loss: 5.0777
+[2025-09-04 12:51:10] [Rank 0]   Group 8 Loss: 5.0777
+[2025-09-04 12:51:10] [Rank 0]   Group 9 Loss: 5.0716
+[2025-09-04 12:51:10] [Rank 0]   Group 9 Loss: 5.0716
+[2025-09-04 12:51:10] [Rank 0]   Group 10 Loss: 5.2333
+[2025-09-04 12:51:10] [Rank 0]   Group 10 Loss: 5.2333
+[2025-09-04 12:51:10] [Rank 0]   Group 11 Loss: 5.2674
+[2025-09-04 12:51:10] [Rank 0]   Group 11 Loss: 5.2674
+[2025-09-04 12:51:10] [Rank 0]   Group 12 Loss: 5.2088
+[2025-09-04 12:51:10] [Rank 0]   Group 12 Loss: 5.2088
+[2025-09-04 12:51:10] [Rank 0]   Group 13 Loss: 5.2840
+[2025-09-04 12:51:10] [Rank 0]   Group 13 Loss: 5.2840
+[2025-09-04 12:51:10] [Rank 0]   Group 14 Loss: 5.2298
+[2025-09-04 12:51:10] [Rank 0]   Group 14 Loss: 5.2298
+[2025-09-04 12:51:10] [Rank 0]   Group 15 Loss: 5.2499
+[2025-09-04 12:51:10] [Rank 0]   Group 15 Loss: 5.2499
+[2025-09-04 12:51:10] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:51:10] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:51:10] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 12:51:10] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 12:51:10] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 12:51:10] [Rank 0]   Group 14 FTA: 0.9400
+[2025-09-04 12:51:10] [Rank 0]   Group 14 FTA: 0.9400
+[2025-09-04 12:51:10] [Rank 0]   Group 15 FTA: 0.5700
+[2025-09-04 12:51:10] [Rank 0]   Group 15 FTA: 0.5700
+[2025-09-04 12:51:11] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:51:11] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:51:11] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:51:11] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:51:11] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:51:11] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:51:12] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:51:12] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:51:12] [Rank 0] step:7501/10000 train_time:326323ms step_avg:43.50ms
+[2025-09-04 12:51:12] [Rank 0] step:7501/10000 train_time:326323ms step_avg:43.50ms
+[2025-09-04 12:51:13] [Rank 0] step:7521/10000 train_time:327095ms step_avg:43.49ms
+[2025-09-04 12:51:13] [Rank 0] step:7521/10000 train_time:327095ms step_avg:43.49ms
+[2025-09-04 12:51:13] [Rank 0] step:7541/10000 train_time:327854ms step_avg:43.48ms
+[2025-09-04 12:51:13] [Rank 0] step:7541/10000 train_time:327854ms step_avg:43.48ms
+[2025-09-04 12:51:14] [Rank 0] step:7561/10000 train_time:328613ms step_avg:43.46ms
+[2025-09-04 12:51:14] [Rank 0] step:7561/10000 train_time:328613ms step_avg:43.46ms
+[2025-09-04 12:51:15] [Rank 0] step:7581/10000 train_time:329374ms step_avg:43.45ms
+[2025-09-04 12:51:15] [Rank 0] step:7581/10000 train_time:329374ms step_avg:43.45ms
+[2025-09-04 12:51:16] [Rank 0] step:7601/10000 train_time:330131ms step_avg:43.43ms
+[2025-09-04 12:51:16] [Rank 0] step:7601/10000 train_time:330131ms step_avg:43.43ms
+[2025-09-04 12:51:16] [Rank 0] step:7621/10000 train_time:330891ms step_avg:43.42ms
+[2025-09-04 12:51:16] [Rank 0] step:7621/10000 train_time:330891ms step_avg:43.42ms
+[2025-09-04 12:51:18] [Rank 0] step:7641/10000 train_time:332329ms step_avg:43.49ms
+[2025-09-04 12:51:18] [Rank 0] step:7641/10000 train_time:332329ms step_avg:43.49ms
+[2025-09-04 12:51:19] [Rank 0] step:7661/10000 train_time:333088ms step_avg:43.48ms
+[2025-09-04 12:51:19] [Rank 0] step:7661/10000 train_time:333088ms step_avg:43.48ms
+[2025-09-04 12:51:19] [Rank 0] step:7681/10000 train_time:333847ms step_avg:43.46ms
+[2025-09-04 12:51:19] [Rank 0] step:7681/10000 train_time:333847ms step_avg:43.46ms
+[2025-09-04 12:51:20] [Rank 0] step:7701/10000 train_time:334605ms step_avg:43.45ms
+[2025-09-04 12:51:20] [Rank 0] step:7701/10000 train_time:334605ms step_avg:43.45ms
+[2025-09-04 12:51:21] [Rank 0] step:7721/10000 train_time:335364ms step_avg:43.44ms
+[2025-09-04 12:51:21] [Rank 0] step:7721/10000 train_time:335364ms step_avg:43.44ms
+[2025-09-04 12:51:22] [Rank 0] step:7741/10000 train_time:336123ms step_avg:43.42ms
+[2025-09-04 12:51:22] [Rank 0] step:7741/10000 train_time:336123ms step_avg:43.42ms
+[2025-09-04 12:51:22] [Rank 0] step:7761/10000 train_time:336882ms step_avg:43.41ms
+[2025-09-04 12:51:22] [Rank 0] step:7761/10000 train_time:336882ms step_avg:43.41ms
+[2025-09-04 12:51:23] [Rank 0] step:7781/10000 train_time:337641ms step_avg:43.39ms
+[2025-09-04 12:51:23] [Rank 0] step:7781/10000 train_time:337641ms step_avg:43.39ms
+[2025-09-04 12:51:24] [Rank 0] step:7801/10000 train_time:338400ms step_avg:43.38ms
+[2025-09-04 12:51:24] [Rank 0] step:7801/10000 train_time:338400ms step_avg:43.38ms
+[2025-09-04 12:51:25] [Rank 0] step:7821/10000 train_time:339159ms step_avg:43.37ms
+[2025-09-04 12:51:25] [Rank 0] step:7821/10000 train_time:339159ms step_avg:43.37ms
+[2025-09-04 12:51:25] [Rank 0] step:7841/10000 train_time:339918ms step_avg:43.35ms
+[2025-09-04 12:51:25] [Rank 0] step:7841/10000 train_time:339918ms step_avg:43.35ms
+[2025-09-04 12:51:26] [Rank 0] step:7861/10000 train_time:340677ms step_avg:43.34ms
+[2025-09-04 12:51:26] [Rank 0] step:7861/10000 train_time:340677ms step_avg:43.34ms
+[2025-09-04 12:51:27] [Rank 0] step:7881/10000 train_time:341436ms step_avg:43.32ms
+[2025-09-04 12:51:27] [Rank 0] step:7881/10000 train_time:341436ms step_avg:43.32ms
+[2025-09-04 12:51:28] [Rank 0] step:7901/10000 train_time:342195ms step_avg:43.31ms
+[2025-09-04 12:51:28] [Rank 0] step:7901/10000 train_time:342195ms step_avg:43.31ms
+[2025-09-04 12:51:28] [Rank 0] step:7921/10000 train_time:342955ms step_avg:43.30ms
+[2025-09-04 12:51:28] [Rank 0] step:7921/10000 train_time:342955ms step_avg:43.30ms
+[2025-09-04 12:51:29] [Rank 0] step:7941/10000 train_time:343714ms step_avg:43.28ms
+[2025-09-04 12:51:29] [Rank 0] step:7941/10000 train_time:343714ms step_avg:43.28ms
+[2025-09-04 12:51:30] [Rank 0] step:7961/10000 train_time:344474ms step_avg:43.27ms
+[2025-09-04 12:51:30] [Rank 0] step:7961/10000 train_time:344474ms step_avg:43.27ms
+[2025-09-04 12:51:31] [Rank 0] step:7981/10000 train_time:345233ms step_avg:43.26ms
+[2025-09-04 12:51:31] [Rank 0] step:7981/10000 train_time:345233ms step_avg:43.26ms
+[2025-09-04 12:51:31] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:51:31] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:51:32] [Rank 0] PRINT: step:8000/10000 train_loss:0.6210 val_loss:0.6124 train_time:345997ms step_avg:43.25ms
+[2025-09-04 12:51:32] [Rank 0] PRINT: step:8000/10000 train_loss:0.6210 val_loss:0.6124 train_time:345997ms step_avg:43.25ms
+[2025-09-04 12:51:32] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:51:32] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:51:32] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:51:32] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:53:09] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:53:09] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:53:09] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:53:09] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:53:09] [Rank 0]   Total Loss: 4.9700
+[2025-09-04 12:53:09] [Rank 0]   Total Loss: 4.9700
+[2025-09-04 12:53:09] [Rank 0]   Total FTA (Unweighted): 0.9813
+[2025-09-04 12:53:09] [Rank 0]   Total FTA (Unweighted): 0.9813
+[2025-09-04 12:53:09] [Rank 0]   Total FTA (Weighted):   0.9812
+[2025-09-04 12:53:09] [Rank 0]   Total FTA (Weighted):   0.9812
+[2025-09-04 12:53:09] [Rank 0]   Group 0 Loss: 4.8923
+[2025-09-04 12:53:09] [Rank 0]   Group 0 Loss: 4.8923
+[2025-09-04 12:53:09] [Rank 0]   Group 1 Loss: 4.4808
+[2025-09-04 12:53:09] [Rank 0]   Group 1 Loss: 4.4808
+[2025-09-04 12:53:09] [Rank 0]   Group 2 Loss: 4.4214
+[2025-09-04 12:53:09] [Rank 0]   Group 2 Loss: 4.4214
+[2025-09-04 12:53:09] [Rank 0]   Group 3 Loss: 4.8844
+[2025-09-04 12:53:09] [Rank 0]   Group 3 Loss: 4.8844
+[2025-09-04 12:53:09] [Rank 0]   Group 4 Loss: 4.9292
+[2025-09-04 12:53:09] [Rank 0]   Group 4 Loss: 4.9292
+[2025-09-04 12:53:09] [Rank 0]   Group 5 Loss: 4.9061
+[2025-09-04 12:53:09] [Rank 0]   Group 5 Loss: 4.9061
+[2025-09-04 12:53:09] [Rank 0]   Group 6 Loss: 4.8226
+[2025-09-04 12:53:09] [Rank 0]   Group 6 Loss: 4.8226
+[2025-09-04 12:53:09] [Rank 0]   Group 7 Loss: 4.8921
+[2025-09-04 12:53:09] [Rank 0]   Group 7 Loss: 4.8921
+[2025-09-04 12:53:09] [Rank 0]   Group 8 Loss: 5.0567
+[2025-09-04 12:53:09] [Rank 0]   Group 8 Loss: 5.0567
+[2025-09-04 12:53:09] [Rank 0]   Group 9 Loss: 5.0536
+[2025-09-04 12:53:09] [Rank 0]   Group 9 Loss: 5.0536
+[2025-09-04 12:53:09] [Rank 0]   Group 10 Loss: 5.1868
+[2025-09-04 12:53:09] [Rank 0]   Group 10 Loss: 5.1868
+[2025-09-04 12:53:09] [Rank 0]   Group 11 Loss: 5.2082
+[2025-09-04 12:53:09] [Rank 0]   Group 11 Loss: 5.2082
+[2025-09-04 12:53:09] [Rank 0]   Group 12 Loss: 5.1570
+[2025-09-04 12:53:09] [Rank 0]   Group 12 Loss: 5.1570
+[2025-09-04 12:53:09] [Rank 0]   Group 13 Loss: 5.2387
+[2025-09-04 12:53:09] [Rank 0]   Group 13 Loss: 5.2387
+[2025-09-04 12:53:09] [Rank 0]   Group 14 Loss: 5.2087
+[2025-09-04 12:53:09] [Rank 0]   Group 14 Loss: 5.2087
+[2025-09-04 12:53:09] [Rank 0]   Group 15 Loss: 5.1822
+[2025-09-04 12:53:09] [Rank 0]   Group 15 Loss: 5.1822
+[2025-09-04 12:53:09] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:53:09] [Rank 0]   Group 14 FTA: 0.9800
+[2025-09-04 12:53:09] [Rank 0]   Group 14 FTA: 0.9800
+[2025-09-04 12:53:09] [Rank 0]   Group 15 FTA: 0.7200
+[2025-09-04 12:53:09] [Rank 0]   Group 15 FTA: 0.7200
+[2025-09-04 12:53:09] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:53:09] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:53:10] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:53:10] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:53:10] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:53:10] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:53:10] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:53:10] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:53:10] [Rank 0] step:8001/10000 train_time:346014ms step_avg:43.25ms
+[2025-09-04 12:53:10] [Rank 0] step:8001/10000 train_time:346014ms step_avg:43.25ms
+[2025-09-04 12:53:11] [Rank 0] step:8021/10000 train_time:347050ms step_avg:43.27ms
+[2025-09-04 12:53:11] [Rank 0] step:8021/10000 train_time:347050ms step_avg:43.27ms
+[2025-09-04 12:53:12] [Rank 0] step:8041/10000 train_time:347808ms step_avg:43.25ms
+[2025-09-04 12:53:12] [Rank 0] step:8041/10000 train_time:347808ms step_avg:43.25ms
+[2025-09-04 12:53:13] [Rank 0] step:8061/10000 train_time:348567ms step_avg:43.24ms
+[2025-09-04 12:53:13] [Rank 0] step:8061/10000 train_time:348567ms step_avg:43.24ms
+[2025-09-04 12:53:14] [Rank 0] step:8081/10000 train_time:349325ms step_avg:43.23ms
+[2025-09-04 12:53:14] [Rank 0] step:8081/10000 train_time:349325ms step_avg:43.23ms
+[2025-09-04 12:53:14] [Rank 0] step:8101/10000 train_time:350083ms step_avg:43.21ms
+[2025-09-04 12:53:14] [Rank 0] step:8101/10000 train_time:350083ms step_avg:43.21ms
+[2025-09-04 12:53:15] [Rank 0] step:8121/10000 train_time:350842ms step_avg:43.20ms
+[2025-09-04 12:53:15] [Rank 0] step:8121/10000 train_time:350842ms step_avg:43.20ms
+[2025-09-04 12:53:16] [Rank 0] step:8141/10000 train_time:351600ms step_avg:43.19ms
+[2025-09-04 12:53:16] [Rank 0] step:8141/10000 train_time:351600ms step_avg:43.19ms
+[2025-09-04 12:53:17] [Rank 0] step:8161/10000 train_time:352358ms step_avg:43.18ms
+[2025-09-04 12:53:17] [Rank 0] step:8161/10000 train_time:352358ms step_avg:43.18ms
+[2025-09-04 12:53:17] [Rank 0] step:8181/10000 train_time:353117ms step_avg:43.16ms
+[2025-09-04 12:53:17] [Rank 0] step:8181/10000 train_time:353117ms step_avg:43.16ms
+[2025-09-04 12:53:18] [Rank 0] step:8201/10000 train_time:353876ms step_avg:43.15ms
+[2025-09-04 12:53:18] [Rank 0] step:8201/10000 train_time:353876ms step_avg:43.15ms
+[2025-09-04 12:53:19] [Rank 0] step:8221/10000 train_time:354635ms step_avg:43.14ms
+[2025-09-04 12:53:19] [Rank 0] step:8221/10000 train_time:354635ms step_avg:43.14ms
+[2025-09-04 12:53:20] [Rank 0] step:8241/10000 train_time:355393ms step_avg:43.13ms
+[2025-09-04 12:53:20] [Rank 0] step:8241/10000 train_time:355393ms step_avg:43.13ms
+[2025-09-04 12:53:20] [Rank 0] step:8261/10000 train_time:356152ms step_avg:43.11ms
+[2025-09-04 12:53:20] [Rank 0] step:8261/10000 train_time:356152ms step_avg:43.11ms
+[2025-09-04 12:53:21] [Rank 0] step:8281/10000 train_time:356911ms step_avg:43.10ms
+[2025-09-04 12:53:21] [Rank 0] step:8281/10000 train_time:356911ms step_avg:43.10ms
+[2025-09-04 12:53:22] [Rank 0] step:8301/10000 train_time:357671ms step_avg:43.09ms
+[2025-09-04 12:53:22] [Rank 0] step:8301/10000 train_time:357671ms step_avg:43.09ms
+[2025-09-04 12:53:23] [Rank 0] step:8321/10000 train_time:358430ms step_avg:43.08ms
+[2025-09-04 12:53:23] [Rank 0] step:8321/10000 train_time:358430ms step_avg:43.08ms
+[2025-09-04 12:53:23] [Rank 0] step:8341/10000 train_time:359189ms step_avg:43.06ms
+[2025-09-04 12:53:23] [Rank 0] step:8341/10000 train_time:359189ms step_avg:43.06ms
+[2025-09-04 12:53:24] [Rank 0] step:8361/10000 train_time:359948ms step_avg:43.05ms
+[2025-09-04 12:53:24] [Rank 0] step:8361/10000 train_time:359948ms step_avg:43.05ms
+[2025-09-04 12:53:25] [Rank 0] step:8381/10000 train_time:360706ms step_avg:43.04ms
+[2025-09-04 12:53:25] [Rank 0] step:8381/10000 train_time:360706ms step_avg:43.04ms
+[2025-09-04 12:53:26] [Rank 0] step:8401/10000 train_time:361465ms step_avg:43.03ms
+[2025-09-04 12:53:26] [Rank 0] step:8401/10000 train_time:361465ms step_avg:43.03ms
+[2025-09-04 12:53:27] [Rank 0] step:8421/10000 train_time:362224ms step_avg:43.01ms
+[2025-09-04 12:53:27] [Rank 0] step:8421/10000 train_time:362224ms step_avg:43.01ms
+[2025-09-04 12:53:27] [Rank 0] step:8441/10000 train_time:362984ms step_avg:43.00ms
+[2025-09-04 12:53:27] [Rank 0] step:8441/10000 train_time:362984ms step_avg:43.00ms
+[2025-09-04 12:53:28] [Rank 0] step:8461/10000 train_time:363742ms step_avg:42.99ms
+[2025-09-04 12:53:28] [Rank 0] step:8461/10000 train_time:363742ms step_avg:42.99ms
+[2025-09-04 12:53:29] [Rank 0] step:8481/10000 train_time:364501ms step_avg:42.98ms
+[2025-09-04 12:53:29] [Rank 0] step:8481/10000 train_time:364501ms step_avg:42.98ms
+[2025-09-04 12:53:30] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:53:30] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:53:30] [Rank 0] PRINT: step:8500/10000 train_loss:0.6170 val_loss:0.6096 train_time:365265ms step_avg:42.97ms
+[2025-09-04 12:53:30] [Rank 0] PRINT: step:8500/10000 train_loss:0.6170 val_loss:0.6096 train_time:365265ms step_avg:42.97ms
+[2025-09-04 12:53:30] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:53:30] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:53:30] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:53:30] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:55:07] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:55:07] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:55:07] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:55:07] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:55:07] [Rank 0]   Total Loss: 5.0073
+[2025-09-04 12:55:07] [Rank 0]   Total Loss: 5.0073
+[2025-09-04 12:55:07] [Rank 0]   Total FTA (Unweighted): 0.9844
+[2025-09-04 12:55:07] [Rank 0]   Total FTA (Unweighted): 0.9844
+[2025-09-04 12:55:07] [Rank 0]   Total FTA (Weighted):   0.9844
+[2025-09-04 12:55:07] [Rank 0]   Total FTA (Weighted):   0.9844
+[2025-09-04 12:55:07] [Rank 0]   Group 0 Loss: 4.8916
+[2025-09-04 12:55:07] [Rank 0]   Group 0 Loss: 4.8916
+[2025-09-04 12:55:07] [Rank 0]   Group 1 Loss: 4.5191
+[2025-09-04 12:55:07] [Rank 0]   Group 1 Loss: 4.5191
+[2025-09-04 12:55:07] [Rank 0]   Group 2 Loss: 4.4244
+[2025-09-04 12:55:07] [Rank 0]   Group 2 Loss: 4.4244
+[2025-09-04 12:55:07] [Rank 0]   Group 3 Loss: 4.9356
+[2025-09-04 12:55:07] [Rank 0]   Group 3 Loss: 4.9356
+[2025-09-04 12:55:07] [Rank 0]   Group 4 Loss: 4.9751
+[2025-09-04 12:55:07] [Rank 0]   Group 4 Loss: 4.9751
+[2025-09-04 12:55:07] [Rank 0]   Group 5 Loss: 4.9261
+[2025-09-04 12:55:07] [Rank 0]   Group 5 Loss: 4.9261
+[2025-09-04 12:55:07] [Rank 0]   Group 6 Loss: 4.8523
+[2025-09-04 12:55:07] [Rank 0]   Group 6 Loss: 4.8523
+[2025-09-04 12:55:07] [Rank 0]   Group 7 Loss: 4.9306
+[2025-09-04 12:55:07] [Rank 0]   Group 7 Loss: 4.9306
+[2025-09-04 12:55:07] [Rank 0]   Group 8 Loss: 5.0867
+[2025-09-04 12:55:07] [Rank 0]   Group 8 Loss: 5.0867
+[2025-09-04 12:55:07] [Rank 0]   Group 9 Loss: 5.0757
+[2025-09-04 12:55:07] [Rank 0]   Group 9 Loss: 5.0757
+[2025-09-04 12:55:07] [Rank 0]   Group 10 Loss: 5.2292
+[2025-09-04 12:55:07] [Rank 0]   Group 10 Loss: 5.2292
+[2025-09-04 12:55:07] [Rank 0]   Group 11 Loss: 5.2702
+[2025-09-04 12:55:07] [Rank 0]   Group 11 Loss: 5.2702
+[2025-09-04 12:55:07] [Rank 0]   Group 12 Loss: 5.2086
+[2025-09-04 12:55:07] [Rank 0]   Group 12 Loss: 5.2086
+[2025-09-04 12:55:07] [Rank 0]   Group 13 Loss: 5.3046
+[2025-09-04 12:55:07] [Rank 0]   Group 13 Loss: 5.3046
+[2025-09-04 12:55:07] [Rank 0]   Group 14 Loss: 5.2544
+[2025-09-04 12:55:07] [Rank 0]   Group 14 Loss: 5.2544
+[2025-09-04 12:55:07] [Rank 0]   Group 15 Loss: 5.2324
+[2025-09-04 12:55:07] [Rank 0]   Group 15 Loss: 5.2324
+[2025-09-04 12:55:07] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 6 FTA: 0.9800
+[2025-09-04 12:55:07] [Rank 0]   Group 6 FTA: 0.9800
+[2025-09-04 12:55:07] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:55:07] [Rank 0]   Group 14 FTA: 0.9800
+[2025-09-04 12:55:07] [Rank 0]   Group 14 FTA: 0.9800
+[2025-09-04 12:55:07] [Rank 0]   Group 15 FTA: 0.7900
+[2025-09-04 12:55:07] [Rank 0]   Group 15 FTA: 0.7900
+[2025-09-04 12:55:07] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:55:07] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:55:08] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:55:08] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:55:08] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:55:08] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:55:08] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:55:08] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:55:08] [Rank 0] step:8501/10000 train_time:365285ms step_avg:42.97ms
+[2025-09-04 12:55:08] [Rank 0] step:8501/10000 train_time:365285ms step_avg:42.97ms
+[2025-09-04 12:55:09] [Rank 0] step:8521/10000 train_time:366050ms step_avg:42.96ms
+[2025-09-04 12:55:09] [Rank 0] step:8521/10000 train_time:366050ms step_avg:42.96ms
+[2025-09-04 12:55:10] [Rank 0] step:8541/10000 train_time:367092ms step_avg:42.98ms
+[2025-09-04 12:55:10] [Rank 0] step:8541/10000 train_time:367092ms step_avg:42.98ms
+[2025-09-04 12:55:11] [Rank 0] step:8561/10000 train_time:367851ms step_avg:42.97ms
+[2025-09-04 12:55:11] [Rank 0] step:8561/10000 train_time:367851ms step_avg:42.97ms
+[2025-09-04 12:55:12] [Rank 0] step:8581/10000 train_time:368609ms step_avg:42.96ms
+[2025-09-04 12:55:12] [Rank 0] step:8581/10000 train_time:368609ms step_avg:42.96ms
+[2025-09-04 12:55:13] [Rank 0] step:8601/10000 train_time:369368ms step_avg:42.94ms
+[2025-09-04 12:55:13] [Rank 0] step:8601/10000 train_time:369368ms step_avg:42.94ms
+[2025-09-04 12:55:13] [Rank 0] step:8621/10000 train_time:370128ms step_avg:42.93ms
+[2025-09-04 12:55:13] [Rank 0] step:8621/10000 train_time:370128ms step_avg:42.93ms
+[2025-09-04 12:55:14] [Rank 0] step:8641/10000 train_time:370886ms step_avg:42.92ms
+[2025-09-04 12:55:14] [Rank 0] step:8641/10000 train_time:370886ms step_avg:42.92ms
+[2025-09-04 12:55:15] [Rank 0] step:8661/10000 train_time:371644ms step_avg:42.91ms
+[2025-09-04 12:55:15] [Rank 0] step:8661/10000 train_time:371644ms step_avg:42.91ms
+[2025-09-04 12:55:16] [Rank 0] step:8681/10000 train_time:372404ms step_avg:42.90ms
+[2025-09-04 12:55:16] [Rank 0] step:8681/10000 train_time:372404ms step_avg:42.90ms
+[2025-09-04 12:55:16] [Rank 0] step:8701/10000 train_time:373162ms step_avg:42.89ms
+[2025-09-04 12:55:16] [Rank 0] step:8701/10000 train_time:373162ms step_avg:42.89ms
+[2025-09-04 12:55:17] [Rank 0] step:8721/10000 train_time:373922ms step_avg:42.88ms
+[2025-09-04 12:55:17] [Rank 0] step:8721/10000 train_time:373922ms step_avg:42.88ms
+[2025-09-04 12:55:18] [Rank 0] step:8741/10000 train_time:374680ms step_avg:42.86ms
+[2025-09-04 12:55:18] [Rank 0] step:8741/10000 train_time:374680ms step_avg:42.86ms
+[2025-09-04 12:55:19] [Rank 0] step:8761/10000 train_time:375440ms step_avg:42.85ms
+[2025-09-04 12:55:19] [Rank 0] step:8761/10000 train_time:375440ms step_avg:42.85ms
+[2025-09-04 12:55:19] [Rank 0] step:8781/10000 train_time:376200ms step_avg:42.84ms
+[2025-09-04 12:55:19] [Rank 0] step:8781/10000 train_time:376200ms step_avg:42.84ms
+[2025-09-04 12:55:20] [Rank 0] step:8801/10000 train_time:376960ms step_avg:42.83ms
+[2025-09-04 12:55:20] [Rank 0] step:8801/10000 train_time:376960ms step_avg:42.83ms
+[2025-09-04 12:55:21] [Rank 0] step:8821/10000 train_time:377723ms step_avg:42.82ms
+[2025-09-04 12:55:21] [Rank 0] step:8821/10000 train_time:377723ms step_avg:42.82ms
+[2025-09-04 12:55:22] [Rank 0] step:8841/10000 train_time:378755ms step_avg:42.84ms
+[2025-09-04 12:55:22] [Rank 0] step:8841/10000 train_time:378755ms step_avg:42.84ms
+[2025-09-04 12:55:23] [Rank 0] step:8861/10000 train_time:379515ms step_avg:42.83ms
+[2025-09-04 12:55:23] [Rank 0] step:8861/10000 train_time:379515ms step_avg:42.83ms
+[2025-09-04 12:55:23] [Rank 0] step:8881/10000 train_time:380276ms step_avg:42.82ms
+[2025-09-04 12:55:23] [Rank 0] step:8881/10000 train_time:380276ms step_avg:42.82ms
+[2025-09-04 12:55:24] [Rank 0] step:8901/10000 train_time:381036ms step_avg:42.81ms
+[2025-09-04 12:55:24] [Rank 0] step:8901/10000 train_time:381036ms step_avg:42.81ms
+[2025-09-04 12:55:25] [Rank 0] step:8921/10000 train_time:381797ms step_avg:42.80ms
+[2025-09-04 12:55:25] [Rank 0] step:8921/10000 train_time:381797ms step_avg:42.80ms
+[2025-09-04 12:55:26] [Rank 0] step:8941/10000 train_time:382557ms step_avg:42.79ms
+[2025-09-04 12:55:26] [Rank 0] step:8941/10000 train_time:382557ms step_avg:42.79ms
+[2025-09-04 12:55:27] [Rank 0] step:8961/10000 train_time:383317ms step_avg:42.78ms
+[2025-09-04 12:55:27] [Rank 0] step:8961/10000 train_time:383317ms step_avg:42.78ms
+[2025-09-04 12:55:27] [Rank 0] step:8981/10000 train_time:384076ms step_avg:42.77ms
+[2025-09-04 12:55:27] [Rank 0] step:8981/10000 train_time:384076ms step_avg:42.77ms
+[2025-09-04 12:55:28] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:55:28] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:55:28] [Rank 0] PRINT: step:9000/10000 train_loss:0.6135 val_loss:0.6077 train_time:384841ms step_avg:42.76ms
+[2025-09-04 12:55:28] [Rank 0] PRINT: step:9000/10000 train_loss:0.6135 val_loss:0.6077 train_time:384841ms step_avg:42.76ms
+[2025-09-04 12:55:29] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:55:29] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:55:29] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:55:29] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:57:06] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:57:06] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:57:06] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:57:06] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:57:06] [Rank 0]   Total Loss: 5.0159
+[2025-09-04 12:57:06] [Rank 0]   Total Loss: 5.0159
+[2025-09-04 12:57:06] [Rank 0]   Total FTA (Unweighted): 0.9925
+[2025-09-04 12:57:06] [Rank 0]   Total FTA (Unweighted): 0.9925
+[2025-09-04 12:57:06] [Rank 0]   Total FTA (Weighted):   0.9925
+[2025-09-04 12:57:06] [Rank 0]   Total FTA (Weighted):   0.9925
+[2025-09-04 12:57:06] [Rank 0]   Group 0 Loss: 4.8954
+[2025-09-04 12:57:06] [Rank 0]   Group 0 Loss: 4.8954
+[2025-09-04 12:57:06] [Rank 0]   Group 1 Loss: 4.5195
+[2025-09-04 12:57:06] [Rank 0]   Group 1 Loss: 4.5195
+[2025-09-04 12:57:06] [Rank 0]   Group 2 Loss: 4.4295
+[2025-09-04 12:57:06] [Rank 0]   Group 2 Loss: 4.4295
+[2025-09-04 12:57:06] [Rank 0]   Group 3 Loss: 4.9565
+[2025-09-04 12:57:06] [Rank 0]   Group 3 Loss: 4.9565
+[2025-09-04 12:57:06] [Rank 0]   Group 4 Loss: 4.9732
+[2025-09-04 12:57:06] [Rank 0]   Group 4 Loss: 4.9732
+[2025-09-04 12:57:06] [Rank 0]   Group 5 Loss: 4.9319
+[2025-09-04 12:57:06] [Rank 0]   Group 5 Loss: 4.9319
+[2025-09-04 12:57:06] [Rank 0]   Group 6 Loss: 4.8676
+[2025-09-04 12:57:06] [Rank 0]   Group 6 Loss: 4.8676
+[2025-09-04 12:57:06] [Rank 0]   Group 7 Loss: 4.9378
+[2025-09-04 12:57:06] [Rank 0]   Group 7 Loss: 4.9378
+[2025-09-04 12:57:06] [Rank 0]   Group 8 Loss: 5.0955
+[2025-09-04 12:57:06] [Rank 0]   Group 8 Loss: 5.0955
+[2025-09-04 12:57:06] [Rank 0]   Group 9 Loss: 5.1042
+[2025-09-04 12:57:06] [Rank 0]   Group 9 Loss: 5.1042
+[2025-09-04 12:57:06] [Rank 0]   Group 10 Loss: 5.2386
+[2025-09-04 12:57:06] [Rank 0]   Group 10 Loss: 5.2386
+[2025-09-04 12:57:06] [Rank 0]   Group 11 Loss: 5.2897
+[2025-09-04 12:57:06] [Rank 0]   Group 11 Loss: 5.2897
+[2025-09-04 12:57:06] [Rank 0]   Group 12 Loss: 5.1884
+[2025-09-04 12:57:06] [Rank 0]   Group 12 Loss: 5.1884
+[2025-09-04 12:57:06] [Rank 0]   Group 13 Loss: 5.3079
+[2025-09-04 12:57:06] [Rank 0]   Group 13 Loss: 5.3079
+[2025-09-04 12:57:06] [Rank 0]   Group 14 Loss: 5.2751
+[2025-09-04 12:57:06] [Rank 0]   Group 14 Loss: 5.2751
+[2025-09-04 12:57:06] [Rank 0]   Group 15 Loss: 5.2428
+[2025-09-04 12:57:06] [Rank 0]   Group 15 Loss: 5.2428
+[2025-09-04 12:57:06] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:57:06] [Rank 0]   Group 15 FTA: 0.8800
+[2025-09-04 12:57:06] [Rank 0]   Group 15 FTA: 0.8800
+[2025-09-04 12:57:06] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:57:06] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:57:07] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:57:07] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:57:07] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:57:07] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:57:07] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:57:07] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:57:07] [Rank 0] step:9001/10000 train_time:384857ms step_avg:42.76ms
+[2025-09-04 12:57:07] [Rank 0] step:9001/10000 train_time:384857ms step_avg:42.76ms
+[2025-09-04 12:57:08] [Rank 0] step:9021/10000 train_time:385637ms step_avg:42.75ms
+[2025-09-04 12:57:08] [Rank 0] step:9021/10000 train_time:385637ms step_avg:42.75ms
+[2025-09-04 12:57:09] [Rank 0] step:9041/10000 train_time:386397ms step_avg:42.74ms
+[2025-09-04 12:57:09] [Rank 0] step:9041/10000 train_time:386397ms step_avg:42.74ms
+[2025-09-04 12:57:10] [Rank 0] step:9061/10000 train_time:387156ms step_avg:42.73ms
+[2025-09-04 12:57:10] [Rank 0] step:9061/10000 train_time:387156ms step_avg:42.73ms
+[2025-09-04 12:57:10] [Rank 0] step:9081/10000 train_time:387916ms step_avg:42.72ms
+[2025-09-04 12:57:10] [Rank 0] step:9081/10000 train_time:387916ms step_avg:42.72ms
+[2025-09-04 12:57:11] [Rank 0] step:9101/10000 train_time:388674ms step_avg:42.71ms
+[2025-09-04 12:57:11] [Rank 0] step:9101/10000 train_time:388674ms step_avg:42.71ms
+[2025-09-04 12:57:12] [Rank 0] step:9121/10000 train_time:389433ms step_avg:42.70ms
+[2025-09-04 12:57:12] [Rank 0] step:9121/10000 train_time:389433ms step_avg:42.70ms
+[2025-09-04 12:57:13] [Rank 0] step:9141/10000 train_time:390191ms step_avg:42.69ms
+[2025-09-04 12:57:13] [Rank 0] step:9141/10000 train_time:390191ms step_avg:42.69ms
+[2025-09-04 12:57:13] [Rank 0] step:9161/10000 train_time:390949ms step_avg:42.68ms
+[2025-09-04 12:57:13] [Rank 0] step:9161/10000 train_time:390949ms step_avg:42.68ms
+[2025-09-04 12:57:15] [Rank 0] step:9181/10000 train_time:391999ms step_avg:42.70ms
+[2025-09-04 12:57:15] [Rank 0] step:9181/10000 train_time:391999ms step_avg:42.70ms
+[2025-09-04 12:57:15] [Rank 0] step:9201/10000 train_time:392758ms step_avg:42.69ms
+[2025-09-04 12:57:15] [Rank 0] step:9201/10000 train_time:392758ms step_avg:42.69ms
+[2025-09-04 12:57:16] [Rank 0] step:9221/10000 train_time:393516ms step_avg:42.68ms
+[2025-09-04 12:57:16] [Rank 0] step:9221/10000 train_time:393516ms step_avg:42.68ms
+[2025-09-04 12:57:17] [Rank 0] step:9241/10000 train_time:394569ms step_avg:42.70ms
+[2025-09-04 12:57:17] [Rank 0] step:9241/10000 train_time:394569ms step_avg:42.70ms
+[2025-09-04 12:57:18] [Rank 0] step:9261/10000 train_time:395329ms step_avg:42.69ms
+[2025-09-04 12:57:18] [Rank 0] step:9261/10000 train_time:395329ms step_avg:42.69ms
+[2025-09-04 12:57:19] [Rank 0] step:9281/10000 train_time:396088ms step_avg:42.68ms
+[2025-09-04 12:57:19] [Rank 0] step:9281/10000 train_time:396088ms step_avg:42.68ms
+[2025-09-04 12:57:19] [Rank 0] step:9301/10000 train_time:396846ms step_avg:42.67ms
+[2025-09-04 12:57:19] [Rank 0] step:9301/10000 train_time:396846ms step_avg:42.67ms
+[2025-09-04 12:57:20] [Rank 0] step:9321/10000 train_time:397605ms step_avg:42.66ms
+[2025-09-04 12:57:20] [Rank 0] step:9321/10000 train_time:397605ms step_avg:42.66ms
+[2025-09-04 12:57:21] [Rank 0] step:9341/10000 train_time:398365ms step_avg:42.65ms
+[2025-09-04 12:57:21] [Rank 0] step:9341/10000 train_time:398365ms step_avg:42.65ms
+[2025-09-04 12:57:22] [Rank 0] step:9361/10000 train_time:399124ms step_avg:42.64ms
+[2025-09-04 12:57:22] [Rank 0] step:9361/10000 train_time:399124ms step_avg:42.64ms
+[2025-09-04 12:57:22] [Rank 0] step:9381/10000 train_time:399882ms step_avg:42.63ms
+[2025-09-04 12:57:22] [Rank 0] step:9381/10000 train_time:399882ms step_avg:42.63ms
+[2025-09-04 12:57:23] [Rank 0] step:9401/10000 train_time:400646ms step_avg:42.62ms
+[2025-09-04 12:57:23] [Rank 0] step:9401/10000 train_time:400646ms step_avg:42.62ms
+[2025-09-04 12:57:24] [Rank 0] step:9421/10000 train_time:401405ms step_avg:42.61ms
+[2025-09-04 12:57:24] [Rank 0] step:9421/10000 train_time:401405ms step_avg:42.61ms
+[2025-09-04 12:57:25] [Rank 0] step:9441/10000 train_time:402165ms step_avg:42.60ms
+[2025-09-04 12:57:25] [Rank 0] step:9441/10000 train_time:402165ms step_avg:42.60ms
+[2025-09-04 12:57:25] [Rank 0] step:9461/10000 train_time:402924ms step_avg:42.59ms
+[2025-09-04 12:57:25] [Rank 0] step:9461/10000 train_time:402924ms step_avg:42.59ms
+[2025-09-04 12:57:26] [Rank 0] step:9481/10000 train_time:403683ms step_avg:42.58ms
+[2025-09-04 12:57:26] [Rank 0] step:9481/10000 train_time:403683ms step_avg:42.58ms
+[2025-09-04 12:57:27] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:57:27] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:57:27] [Rank 0] PRINT: step:9500/10000 train_loss:0.6109 val_loss:0.6064 train_time:404447ms step_avg:42.57ms
+[2025-09-04 12:57:27] [Rank 0] PRINT: step:9500/10000 train_loss:0.6109 val_loss:0.6064 train_time:404447ms step_avg:42.57ms
+[2025-09-04 12:57:27] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:57:27] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:57:28] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:57:28] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:59:04] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:59:04] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 12:59:04] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:59:04] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 12:59:04] [Rank 0]   Total Loss: 5.0322
+[2025-09-04 12:59:04] [Rank 0]   Total Loss: 5.0322
+[2025-09-04 12:59:04] [Rank 0]   Total FTA (Unweighted): 0.9950
+[2025-09-04 12:59:04] [Rank 0]   Total FTA (Unweighted): 0.9950
+[2025-09-04 12:59:04] [Rank 0]   Total FTA (Weighted):   0.9950
+[2025-09-04 12:59:04] [Rank 0]   Total FTA (Weighted):   0.9950
+[2025-09-04 12:59:04] [Rank 0]   Group 0 Loss: 4.9864
+[2025-09-04 12:59:04] [Rank 0]   Group 0 Loss: 4.9864
+[2025-09-04 12:59:04] [Rank 0]   Group 1 Loss: 4.5206
+[2025-09-04 12:59:04] [Rank 0]   Group 1 Loss: 4.5206
+[2025-09-04 12:59:04] [Rank 0]   Group 2 Loss: 4.4811
+[2025-09-04 12:59:04] [Rank 0]   Group 2 Loss: 4.4811
+[2025-09-04 12:59:04] [Rank 0]   Group 3 Loss: 4.9421
+[2025-09-04 12:59:04] [Rank 0]   Group 3 Loss: 4.9421
+[2025-09-04 12:59:04] [Rank 0]   Group 4 Loss: 4.9890
+[2025-09-04 12:59:04] [Rank 0]   Group 4 Loss: 4.9890
+[2025-09-04 12:59:04] [Rank 0]   Group 5 Loss: 4.9451
+[2025-09-04 12:59:04] [Rank 0]   Group 5 Loss: 4.9451
+[2025-09-04 12:59:04] [Rank 0]   Group 6 Loss: 4.8532
+[2025-09-04 12:59:04] [Rank 0]   Group 6 Loss: 4.8532
+[2025-09-04 12:59:04] [Rank 0]   Group 7 Loss: 4.9558
+[2025-09-04 12:59:04] [Rank 0]   Group 7 Loss: 4.9558
+[2025-09-04 12:59:04] [Rank 0]   Group 8 Loss: 5.0945
+[2025-09-04 12:59:04] [Rank 0]   Group 8 Loss: 5.0945
+[2025-09-04 12:59:04] [Rank 0]   Group 9 Loss: 5.0990
+[2025-09-04 12:59:04] [Rank 0]   Group 9 Loss: 5.0990
+[2025-09-04 12:59:04] [Rank 0]   Group 10 Loss: 5.2478
+[2025-09-04 12:59:04] [Rank 0]   Group 10 Loss: 5.2478
+[2025-09-04 12:59:04] [Rank 0]   Group 11 Loss: 5.2907
+[2025-09-04 12:59:04] [Rank 0]   Group 11 Loss: 5.2907
+[2025-09-04 12:59:04] [Rank 0]   Group 12 Loss: 5.2207
+[2025-09-04 12:59:04] [Rank 0]   Group 12 Loss: 5.2207
+[2025-09-04 12:59:04] [Rank 0]   Group 13 Loss: 5.3203
+[2025-09-04 12:59:04] [Rank 0]   Group 13 Loss: 5.3203
+[2025-09-04 12:59:04] [Rank 0]   Group 14 Loss: 5.2823
+[2025-09-04 12:59:04] [Rank 0]   Group 14 Loss: 5.2823
+[2025-09-04 12:59:04] [Rank 0]   Group 15 Loss: 5.2859
+[2025-09-04 12:59:04] [Rank 0]   Group 15 Loss: 5.2859
+[2025-09-04 12:59:04] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 12:59:04] [Rank 0]   Group 15 FTA: 0.9200
+[2025-09-04 12:59:04] [Rank 0]   Group 15 FTA: 0.9200
+[2025-09-04 12:59:05] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:59:05] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 12:59:05] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:59:05] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 12:59:05] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:59:05] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 12:59:05] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:59:05] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 12:59:06] [Rank 0] step:9501/10000 train_time:404463ms step_avg:42.57ms
+[2025-09-04 12:59:06] [Rank 0] step:9501/10000 train_time:404463ms step_avg:42.57ms
+[2025-09-04 12:59:06] [Rank 0] step:9521/10000 train_time:405244ms step_avg:42.56ms
+[2025-09-04 12:59:06] [Rank 0] step:9521/10000 train_time:405244ms step_avg:42.56ms
+[2025-09-04 12:59:07] [Rank 0] step:9541/10000 train_time:406004ms step_avg:42.55ms
+[2025-09-04 12:59:07] [Rank 0] step:9541/10000 train_time:406004ms step_avg:42.55ms
+[2025-09-04 12:59:08] [Rank 0] step:9561/10000 train_time:406763ms step_avg:42.54ms
+[2025-09-04 12:59:08] [Rank 0] step:9561/10000 train_time:406763ms step_avg:42.54ms
+[2025-09-04 12:59:09] [Rank 0] step:9581/10000 train_time:407522ms step_avg:42.53ms
+[2025-09-04 12:59:09] [Rank 0] step:9581/10000 train_time:407522ms step_avg:42.53ms
+[2025-09-04 12:59:09] [Rank 0] step:9601/10000 train_time:408281ms step_avg:42.52ms
+[2025-09-04 12:59:09] [Rank 0] step:9601/10000 train_time:408281ms step_avg:42.52ms
+[2025-09-04 12:59:10] [Rank 0] step:9621/10000 train_time:409040ms step_avg:42.52ms
+[2025-09-04 12:59:10] [Rank 0] step:9621/10000 train_time:409040ms step_avg:42.52ms
+[2025-09-04 12:59:11] [Rank 0] step:9641/10000 train_time:409800ms step_avg:42.51ms
+[2025-09-04 12:59:11] [Rank 0] step:9641/10000 train_time:409800ms step_avg:42.51ms
+[2025-09-04 12:59:12] [Rank 0] step:9661/10000 train_time:410840ms step_avg:42.53ms
+[2025-09-04 12:59:12] [Rank 0] step:9661/10000 train_time:410840ms step_avg:42.53ms
+[2025-09-04 12:59:13] [Rank 0] step:9681/10000 train_time:411599ms step_avg:42.52ms
+[2025-09-04 12:59:13] [Rank 0] step:9681/10000 train_time:411599ms step_avg:42.52ms
+[2025-09-04 12:59:13] [Rank 0] step:9701/10000 train_time:412357ms step_avg:42.51ms
+[2025-09-04 12:59:13] [Rank 0] step:9701/10000 train_time:412357ms step_avg:42.51ms
+[2025-09-04 12:59:14] [Rank 0] step:9721/10000 train_time:413116ms step_avg:42.50ms
+[2025-09-04 12:59:14] [Rank 0] step:9721/10000 train_time:413116ms step_avg:42.50ms
+[2025-09-04 12:59:15] [Rank 0] step:9741/10000 train_time:413876ms step_avg:42.49ms
+[2025-09-04 12:59:15] [Rank 0] step:9741/10000 train_time:413876ms step_avg:42.49ms
+[2025-09-04 12:59:16] [Rank 0] step:9761/10000 train_time:414636ms step_avg:42.48ms
+[2025-09-04 12:59:16] [Rank 0] step:9761/10000 train_time:414636ms step_avg:42.48ms
+[2025-09-04 12:59:16] [Rank 0] step:9781/10000 train_time:415395ms step_avg:42.47ms
+[2025-09-04 12:59:16] [Rank 0] step:9781/10000 train_time:415395ms step_avg:42.47ms
+[2025-09-04 12:59:17] [Rank 0] step:9801/10000 train_time:416154ms step_avg:42.46ms
+[2025-09-04 12:59:17] [Rank 0] step:9801/10000 train_time:416154ms step_avg:42.46ms
+[2025-09-04 12:59:18] [Rank 0] step:9821/10000 train_time:416913ms step_avg:42.45ms
+[2025-09-04 12:59:18] [Rank 0] step:9821/10000 train_time:416913ms step_avg:42.45ms
+[2025-09-04 12:59:19] [Rank 0] step:9841/10000 train_time:417672ms step_avg:42.44ms
+[2025-09-04 12:59:19] [Rank 0] step:9841/10000 train_time:417672ms step_avg:42.44ms
+[2025-09-04 12:59:20] [Rank 0] step:9861/10000 train_time:418430ms step_avg:42.43ms
+[2025-09-04 12:59:20] [Rank 0] step:9861/10000 train_time:418430ms step_avg:42.43ms
+[2025-09-04 12:59:21] [Rank 0] step:9881/10000 train_time:419413ms step_avg:42.45ms
+[2025-09-04 12:59:21] [Rank 0] step:9881/10000 train_time:419413ms step_avg:42.45ms
+[2025-09-04 12:59:21] [Rank 0] step:9901/10000 train_time:420173ms step_avg:42.44ms
+[2025-09-04 12:59:21] [Rank 0] step:9901/10000 train_time:420173ms step_avg:42.44ms
+[2025-09-04 12:59:22] [Rank 0] step:9921/10000 train_time:420932ms step_avg:42.43ms
+[2025-09-04 12:59:22] [Rank 0] step:9921/10000 train_time:420932ms step_avg:42.43ms
+[2025-09-04 12:59:23] [Rank 0] step:9941/10000 train_time:421982ms step_avg:42.45ms
+[2025-09-04 12:59:23] [Rank 0] step:9941/10000 train_time:421982ms step_avg:42.45ms
+[2025-09-04 12:59:24] [Rank 0] step:9961/10000 train_time:422743ms step_avg:42.44ms
+[2025-09-04 12:59:24] [Rank 0] step:9961/10000 train_time:422743ms step_avg:42.44ms
+[2025-09-04 12:59:25] [Rank 0] step:9981/10000 train_time:423502ms step_avg:42.43ms
+[2025-09-04 12:59:25] [Rank 0] step:9981/10000 train_time:423502ms step_avg:42.43ms
+[2025-09-04 12:59:25] [Rank 0] step:10000/10000 train_time:424223ms step_avg:42.42ms
+[2025-09-04 12:59:25] [Rank 0] step:10000/10000 train_time:424223ms step_avg:42.42ms
+[2025-09-04 12:59:25] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:59:25] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 12:59:26] [Rank 0] PRINT: step:10000/10000 train_loss:0.6088 val_loss:0.6052 train_time:424273ms step_avg:42.43ms
+[2025-09-04 12:59:26] [Rank 0] PRINT: step:10000/10000 train_loss:0.6088 val_loss:0.6052 train_time:424273ms step_avg:42.43ms
+[2025-09-04 12:59:26] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:59:26] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 12:59:26] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 12:59:26] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:01:02] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:01:02] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:01:02] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:01:02] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:01:02] [Rank 0]   Total Loss: 5.0112
+[2025-09-04 13:01:02] [Rank 0]   Total Loss: 5.0112
+[2025-09-04 13:01:02] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 13:01:02] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 13:01:02] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 13:01:02] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 13:01:02] [Rank 0]   Group 0 Loss: 4.9122
+[2025-09-04 13:01:02] [Rank 0]   Group 0 Loss: 4.9122
+[2025-09-04 13:01:02] [Rank 0]   Group 1 Loss: 4.5181
+[2025-09-04 13:01:02] [Rank 0]   Group 1 Loss: 4.5181
+[2025-09-04 13:01:02] [Rank 0]   Group 2 Loss: 4.4301
+[2025-09-04 13:01:02] [Rank 0]   Group 2 Loss: 4.4301
+[2025-09-04 13:01:02] [Rank 0]   Group 3 Loss: 4.9253
+[2025-09-04 13:01:02] [Rank 0]   Group 3 Loss: 4.9253
+[2025-09-04 13:01:02] [Rank 0]   Group 4 Loss: 4.9629
+[2025-09-04 13:01:02] [Rank 0]   Group 4 Loss: 4.9629
+[2025-09-04 13:01:02] [Rank 0]   Group 5 Loss: 4.9308
+[2025-09-04 13:01:02] [Rank 0]   Group 5 Loss: 4.9308
+[2025-09-04 13:01:02] [Rank 0]   Group 6 Loss: 4.8493
+[2025-09-04 13:01:02] [Rank 0]   Group 6 Loss: 4.8493
+[2025-09-04 13:01:02] [Rank 0]   Group 7 Loss: 4.9110
+[2025-09-04 13:01:02] [Rank 0]   Group 7 Loss: 4.9110
+[2025-09-04 13:01:02] [Rank 0]   Group 8 Loss: 5.0854
+[2025-09-04 13:01:02] [Rank 0]   Group 8 Loss: 5.0854
+[2025-09-04 13:01:02] [Rank 0]   Group 9 Loss: 5.0976
+[2025-09-04 13:01:02] [Rank 0]   Group 9 Loss: 5.0976
+[2025-09-04 13:01:02] [Rank 0]   Group 10 Loss: 5.2382
+[2025-09-04 13:01:02] [Rank 0]   Group 10 Loss: 5.2382
+[2025-09-04 13:01:02] [Rank 0]   Group 11 Loss: 5.2826
+[2025-09-04 13:01:02] [Rank 0]   Group 11 Loss: 5.2826
+[2025-09-04 13:01:02] [Rank 0]   Group 12 Loss: 5.1859
+[2025-09-04 13:01:02] [Rank 0]   Group 12 Loss: 5.1859
+[2025-09-04 13:01:02] [Rank 0]   Group 13 Loss: 5.3000
+[2025-09-04 13:01:02] [Rank 0]   Group 13 Loss: 5.3000
+[2025-09-04 13:01:02] [Rank 0]   Group 14 Loss: 5.2729
+[2025-09-04 13:01:02] [Rank 0]   Group 14 Loss: 5.2729
+[2025-09-04 13:01:02] [Rank 0]   Group 15 Loss: 5.2776
+[2025-09-04 13:01:02] [Rank 0]   Group 15 Loss: 5.2776
+[2025-09-04 13:01:02] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:01:02] [Rank 0]   Group 15 FTA: 0.9300
+[2025-09-04 13:01:02] [Rank 0]   Group 15 FTA: 0.9300
+[2025-09-04 13:01:03] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 13:01:03] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_loss_curves.png
+[2025-09-04 13:01:03] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 13:01:03] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/per_class_acc_curves.png
+[2025-09-04 13:01:04] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 13:01:04] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_loss_curve.png
+[2025-09-04 13:01:04] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 13:01:04] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_45/total_acc_curve.png
+[2025-09-04 13:01:04] [Rank 0] step:10001/10000 train_time:424288ms step_avg:42.42ms
+[2025-09-04 13:01:04] [Rank 0] step:10001/10000 train_time:424288ms step_avg:42.42ms
+[2025-09-04 13:01:04] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 13:01:04 2025 ---
+[2025-09-04 13:01:04] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 13:01:04 2025 ---
+[2025-09-04 13:01:04] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB
+[2025-09-04 13:01:04] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/config.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..b1c57c91f735bcdf732691cbeb46a87b0f672391
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/config.json
@@ -0,0 +1,29 @@
+{
+    "cli_args": {
+        "unet": false,
+        "seed": 46,
+        "optimizer_mode": 10,
+        "model_parameterization": "qkvo",
+        "per_group_k": 100,
+        "muon_lr": 0.002,
+        "adam_lr": 0.002,
+        "base_dir": "logs_qa_muon/diff_modes",
+        "sgd_lr": 0.01,
+        "m_val": 15,
+        "qa_jsonl_path": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl"
+    },
+    "hyperparameters": {
+        "train_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin",
+        "val_files": "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin",
+        "val_tokens": 491520,
+        "train_seq_len": 3072,
+        "val_seq_len": 16384,
+        "num_iterations": 10000,
+        "cooldown_frac": 0.8,
+        "vocab_size": 50257,
+        "val_loss_every": 500,
+        "save_checkpoint": false
+    },
+    "run_uuid_for_log": "49633f62-3f6e-4f99-aa92-a1a5f4280a83",
+    "script_code_logged_at_start": true
+}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/fixed_eval_indices.json b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/fixed_eval_indices.json
new file mode 100644
index 0000000000000000000000000000000000000000..a823775225c5e592eb10700e5e0319b0491b1eb6
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/fixed_eval_indices.json
@@ -0,0 +1 @@
+{"1": [1238956, 182074, 1437575, 1061037, 383150, 1176376, 926, 823011, 832520, 1266421, 512738, 144357, 848076, 890204, 213997, 95146, 261767, 467731, 832231, 217985, 913168, 107253, 1361828, 61314, 1230420, 1133619, 146690, 429587, 419151, 58695, 1579770, 503799, 1421284, 882534, 1022637, 785343, 1154604, 67783, 1325109, 243941, 1213240, 438111, 460295, 269373, 538055, 1347006, 71775, 255496, 299906, 1227973, 815402, 190082, 1304077, 1023347, 613801, 983830, 1284420, 389321, 1625224, 717538, 1172273, 992184, 1181312, 1014039, 885952, 1538489, 158933, 1667270, 1250445, 958097, 1458224, 1306495, 62945, 733843, 1360200, 540493, 762461, 501460, 1208142, 1180559, 1333588, 690481, 355756, 618511, 733586, 650301, 799437, 165533, 1238977, 323078, 1485080, 609610, 1212241, 606952, 1253407, 1420922, 327112, 701, 777907, 1626516], "0": [1390189, 1220977, 1312259, 1201125, 1235379, 1272843, 344142, 1119560, 856330, 766456, 1145928, 550624, 684681, 1309879, 54620, 1126124, 1278077, 910610, 1532044, 73645, 460165, 462723, 457959, 476877, 320873, 1408316, 989156, 1276325, 338260, 577743, 1136972, 25115, 211215, 1296818, 968705, 436781, 318224, 385325, 1516441, 533464, 1628693, 703399, 579670, 1518978, 305154, 1625960, 1400284, 713841, 1652150, 811009, 336920, 881015, 512030, 1347531, 1005706, 412167, 173312, 1136816, 343017, 537797, 1489267, 935475, 703854, 1570584, 1504269, 1458356, 1310700, 1242505, 509976, 22096, 788120, 1616850, 1591198, 1151224, 797820, 823995, 1153364, 1060220, 727674, 580729, 148912, 334290, 686098, 1633151, 1130523, 1012440, 1374480, 953410, 1381730, 1376118, 932096, 739115, 739014, 263875, 1400308, 556177, 1341771, 348626, 575350, 254846], "10": [748592, 515936, 1645691, 329828, 1104748, 1453439, 406461, 1557507, 1146644, 1398421, 163667, 228310, 407739, 585218, 785725, 1371077, 134834, 15431, 1303784, 273099, 1530332, 1320219, 1329247, 753603, 664548, 501873, 1512467, 583734, 1355643, 600440, 428998, 1426876, 121341, 1456507, 892281, 1242339, 1304508, 968336, 1593981, 634471, 1253799, 593156, 1486078, 712425, 441285, 618217, 1024661, 1395408, 1423117, 1463918, 121752, 800727, 52402, 1185143, 1651644, 1583572, 1302889, 1040220, 1234122, 333457, 1552652, 1567291, 1405501, 222708, 1469496, 875611, 894032, 783951, 934281, 109533, 198744, 643064, 197926, 224405, 1169998, 956511, 1096712, 1213224, 828780, 204659, 564225, 257079, 754287, 1598723, 821179, 1156648, 60674, 290993, 1215838, 1556010, 1025931, 687671, 252775, 771448, 574110, 578615, 279536, 1503737, 1089787, 2338], "14": [1602125, 1388977, 709154, 666138, 1446614, 572103, 99632, 25794, 1487658, 15077, 1600205, 682155, 679850, 1462049, 970283, 81336, 378196, 1627374, 832926, 1217832, 740140, 1336941, 793290, 1441127, 297494, 303372, 339524, 1363322, 525184, 593339, 61055, 1468219, 1110310, 1061855, 1046762, 410655, 449257, 530144, 1362373, 241552, 926068, 1618007, 629406, 327589, 263223, 1525290, 621372, 606355, 534601, 126046, 934024, 1181343, 1234759, 295773, 838035, 1447427, 1395837, 1568116, 486923, 465204, 98192, 1145989, 271529, 720487, 1091877, 653313, 872526, 13890, 111971, 671086, 301528, 50698, 455373, 786098, 329199, 599178, 644737, 1404606, 33829, 279256, 161938, 309352, 1537653, 21723, 309413, 478184, 340938, 1080124, 1375458, 1301528, 331423, 212589, 1220931, 112707, 454022, 1466562, 1238127, 341561, 1497264, 247245], "3": [143966, 715765, 455823, 1414629, 814361, 278667, 700211, 1286614, 1601013, 923359, 1105960, 549680, 371512, 732263, 1236795, 432942, 1379421, 488467, 679369, 1200930, 479010, 683138, 700149, 1422714, 348019, 1263383, 1044275, 820031, 969830, 75153, 1410929, 394116, 1486865, 299414, 264214, 1416493, 428688, 107076, 753671, 674417, 1456072, 307080, 1259105, 1563344, 1198484, 456551, 1061513, 1128111, 922541, 1040277, 1360208, 1411429, 259514, 883801, 716781, 1004987, 285196, 266162, 781115, 385241, 252691, 1398963, 1440582, 1300653, 756663, 231617, 1319447, 42576, 1434337, 20658, 581284, 178420, 1626533, 694079, 1531414, 1631177, 1478512, 647215, 706568, 303323, 10540, 1010611, 1150313, 159016, 1614581, 1257722, 470491, 1333309, 94699, 756222, 1456446, 296132, 1186414, 1591257, 1155050, 789439, 1472373, 1403759, 1451276, 745193], "4": [1227524, 563788, 436964, 1380176, 1648232, 267090, 273744, 1031517, 1580586, 416846, 1309866, 114170, 1649739, 1161604, 690741, 1432068, 224570, 322835, 1559902, 479718, 709757, 410142, 1668701, 1143278, 968258, 406289, 884233, 1080971, 269480, 631196, 421463, 374502, 430292, 1309826, 1422997, 1439354, 131354, 975504, 84660, 1590390, 288064, 1419944, 695652, 320713, 260376, 222416, 937739, 1104508, 1323311, 962301, 1060820, 1117857, 280792, 692359, 656651, 358878, 828406, 598701, 127434, 1511259, 1332605, 1178325, 1281187, 203589, 792701, 684176, 543034, 980077, 982509, 640126, 1252418, 1529627, 519645, 359068, 992779, 550843, 1493488, 447632, 1512599, 361760, 626672, 89042, 604248, 574827, 87735, 1432602, 473061, 668609, 925422, 311417, 71495, 682067, 221270, 600053, 673169, 1408933, 719302, 788676, 998101, 243902], "15": [1444113, 1102795, 1149896, 1181364, 349047, 1076908, 816006, 945310, 197585, 276574, 946546, 1505399, 740223, 1154813, 1312891, 1261018, 249986, 1504592, 647427, 915615, 1245136, 1288301, 802239, 410533, 805664, 1013403, 498363, 777073, 1508176, 18111, 1147340, 880289, 215773, 924760, 1301970, 318135, 1524489, 103475, 872748, 1568512, 722849, 1364157, 838956, 1451408, 970208, 1350756, 1498244, 980461, 1098677, 1141812, 463859, 95672, 1220365, 554451, 524014, 1161048, 17478, 113548, 1027741, 1657643, 1462809, 1279104, 85588, 241936, 510707, 151817, 1191227, 574942, 735662, 1046248, 1056729, 96279, 781378, 1042711, 1602339, 1503514, 299087, 898968, 553537, 293167, 897565, 572125, 425215, 916907, 1210531, 1013275, 201532, 883447, 1641610, 1185682, 425791, 831133, 1291227, 1099064, 934983, 588125, 1284591, 1163873, 524656, 1230762], "13": [1122510, 535854, 845319, 585144, 884184, 441130, 934654, 1046629, 1512962, 1190830, 1462685, 90530, 894615, 1527783, 415323, 90852, 1556905, 1002406, 366356, 1386022, 846952, 928360, 3138, 406205, 78588, 1208032, 1210206, 257177, 137047, 711380, 693628, 105893, 392053, 1301525, 677934, 1549630, 1134520, 447564, 1194061, 672969, 740755, 1045536, 633455, 1336343, 613147, 1161931, 234095, 1246643, 643373, 834221, 1351442, 1502069, 1401555, 759927, 1320337, 666063, 291222, 956633, 400583, 963313, 683474, 534882, 1101312, 1519551, 1513194, 752124, 1028953, 1383362, 169732, 785294, 160568, 1053580, 774534, 372048, 292049, 791470, 1318537, 141966, 510573, 1440830, 930544, 1352026, 463299, 1245306, 159755, 1566413, 316835, 1135200, 266962, 112892, 253603, 1328380, 703630, 21798, 268463, 7112, 1232576, 374978, 240715, 1299198], "2": [463910, 845243, 985993, 402000, 962033, 473711, 108588, 1074932, 539045, 1162793, 1086008, 778638, 751876, 53650, 465364, 803834, 820717, 9490, 473708, 26912, 970677, 1441139, 231484, 749139, 286603, 1346834, 1641988, 699767, 816768, 1220504, 978431, 1331006, 611334, 829473, 271532, 566760, 332292, 1326050, 887731, 597530, 940973, 491159, 1150471, 122010, 225465, 382413, 1207486, 1568920, 992793, 896409, 1479386, 1539347, 759024, 293232, 1345256, 1282360, 956014, 371580, 1202087, 1638038, 892501, 38073, 1651323, 1103388, 1315239, 79005, 1045045, 1230831, 1206253, 1376976, 140037, 320204, 274836, 663525, 397448, 666367, 1501713, 570453, 836424, 694423, 157979, 279074, 1235605, 536170, 301830, 778556, 1069897, 181207, 1445408, 1082031, 969552, 1083538, 1395871, 168187, 5476, 514815, 194418, 1038889, 1288748, 1535767], "8": [258910, 518286, 80108, 811775, 937854, 1623550, 1592007, 661848, 1484674, 300, 104226, 1075215, 1190104, 8885, 1062956, 88002, 1020623, 189243, 1265153, 1472857, 301412, 21371, 441401, 1078954, 1648642, 894192, 1083738, 1182805, 1347655, 316568, 1063971, 255177, 186613, 550617, 102173, 563694, 465966, 965224, 372481, 929520, 1423441, 6575, 862309, 144894, 891435, 123628, 653301, 1378790, 1247156, 1129087, 1381991, 410876, 391895, 202136, 868149, 233421, 564464, 1174972, 1385073, 238326, 483488, 507729, 1329666, 955266, 479623, 902401, 1343890, 827390, 371073, 1480060, 1241650, 1670957, 644144, 1371709, 1324243, 670126, 23715, 1566067, 70858, 522312, 558100, 301714, 359605, 1401913, 1060057, 1027855, 987700, 720084, 324937, 562646, 390910, 1398541, 1097298, 1626769, 1461630, 1430950, 650075, 1025475, 274735, 650932], "7": [181761, 629797, 231148, 365747, 1229714, 500408, 198146, 500703, 1250728, 1578820, 753672, 84476, 1379367, 355424, 1591855, 1400949, 1194404, 912060, 884197, 320928, 1096692, 714201, 331003, 1345763, 1620998, 457231, 1542074, 1108477, 1602113, 641593, 542514, 159345, 360278, 35579, 506, 470465, 1606024, 1349783, 731258, 382125, 883800, 1652746, 382071, 99241, 20216, 1398849, 245155, 1601928, 1544592, 1632383, 1597152, 1266758, 1032571, 88287, 521379, 1423306, 579207, 1152743, 271926, 418168, 8827, 1358019, 429579, 380995, 1040420, 1067138, 283335, 1150469, 280745, 1184794, 468245, 948227, 357499, 527525, 1067827, 816862, 360563, 433490, 258517, 581805, 1089641, 1537081, 473494, 1480784, 60701, 686908, 614323, 143069, 738219, 1234467, 927103, 64088, 658420, 1353941, 1517049, 753334, 198407, 434613, 850731, 1211609], "5": [1250448, 711876, 285881, 342500, 987883, 195047, 810508, 356140, 999602, 284542, 730936, 736191, 661291, 11822, 551928, 617679, 384856, 414906, 238071, 473605, 868498, 1269087, 558407, 245083, 810169, 1456559, 1657433, 70018, 865238, 416303, 1389734, 1561961, 534163, 1098802, 1494214, 1285484, 599936, 296819, 696611, 307176, 29697, 731071, 257656, 1465441, 1668863, 166995, 1668377, 1457645, 1400335, 1525120, 33323, 437285, 459069, 620176, 231704, 846424, 1391144, 364636, 63552, 1093702, 67741, 285428, 454922, 719128, 1393391, 87651, 1418561, 1312176, 652985, 1446762, 71280, 1594253, 89848, 628629, 1117344, 882598, 1014283, 1601256, 739004, 508965, 1273914, 44245, 1267971, 1430470, 611385, 930900, 1439585, 1545682, 1300966, 1331569, 364367, 27900, 107616, 1579070, 94723, 818780, 1183546, 384575, 142627, 431847], "11": [12772, 1091320, 1136723, 560663, 715094, 1097800, 1016547, 1299126, 961155, 725788, 999768, 1277062, 650665, 1323964, 1024796, 1255124, 215111, 268904, 442949, 358645, 621958, 1666210, 1154349, 375517, 614870, 1166618, 1418680, 795635, 184985, 129300, 129366, 1343289, 1151524, 193531, 634856, 1618906, 43721, 1601844, 591772, 1322095, 1522284, 431856, 574972, 1544663, 1389257, 1113893, 1515857, 614345, 809163, 132406, 95327, 997226, 423104, 910705, 261584, 765030, 1216397, 439297, 1565563, 1595785, 1622521, 975588, 1314965, 1205327, 1501612, 1642358, 656493, 1369509, 195325, 1554913, 1661675, 790468, 511387, 1227059, 1219143, 1118001, 788451, 697651, 1575073, 1345104, 167862, 1409843, 138942, 984075, 1382364, 1299538, 214601, 1541475, 459076, 353219, 1310728, 757670, 904351, 334019, 1003259, 670104, 10386, 13556, 1097931, 1526407], "6": [295849, 644897, 1207883, 1648549, 1478750, 1486750, 1156499, 621047, 1387936, 662848, 1362507, 891859, 1188350, 1298200, 1044010, 767288, 1191230, 683760, 533545, 498960, 1421819, 366846, 54389, 1382045, 1002942, 1285004, 457209, 1074763, 823805, 1508295, 719970, 837720, 19881, 1189285, 1226048, 1078656, 1433699, 14805, 254682, 614235, 805087, 900964, 343832, 561709, 243288, 465867, 497085, 1153456, 730115, 42050, 878511, 1382508, 667820, 712856, 715963, 1499782, 1516002, 444399, 726168, 1204910, 591639, 986194, 107160, 1596752, 1239812, 591854, 536489, 1347905, 1597670, 122237, 620494, 354940, 1561646, 1352115, 74785, 163277, 590349, 1429584, 1440950, 848184, 1580723, 1631128, 979479, 758861, 84655, 1079672, 673692, 1315324, 167985, 1353485, 21815, 1175739, 1360408, 1062086, 1369999, 733863, 1371964, 61279, 1152056, 220910], "9": [780278, 1156801, 827882, 1095960, 874763, 1330006, 979412, 1154745, 1399441, 515095, 1505401, 1454104, 1563117, 1316054, 57098, 43367, 1175749, 1257252, 865038, 624410, 186727, 907758, 126168, 1353431, 167806, 1021404, 1585615, 846939, 552247, 1400471, 68525, 190495, 1556857, 181528, 47677, 136468, 189446, 269049, 262467, 503841, 1367168, 382007, 486946, 621235, 622683, 469399, 1537192, 707594, 1003091, 56021, 1354302, 799290, 1618221, 696794, 836658, 1591562, 251096, 1321886, 412122, 1352217, 946767, 1267324, 213504, 747174, 1403653, 1589967, 849743, 549269, 663800, 1315189, 451416, 1330881, 1352197, 1517840, 827169, 1228725, 1097725, 955944, 1375509, 1590213, 266039, 1558465, 1498493, 604161, 608723, 1547013, 1484549, 1530909, 1325336, 1584866, 630737, 1023780, 227549, 961233, 987591, 159267, 608424, 636687, 745397, 460687], "12": [1586709, 357745, 1274743, 202991, 1250469, 837216, 757864, 185235, 493474, 621351, 22706, 1096491, 186138, 1482718, 534525, 1157131, 1350934, 1167912, 1366603, 756644, 311945, 72417, 706444, 1464146, 1236814, 74555, 983713, 1182807, 1082263, 804850, 1028788, 1405036, 932746, 742978, 1589076, 723766, 898380, 906214, 913629, 33386, 256107, 1245317, 360574, 836359, 565572, 400905, 1269191, 1042959, 886332, 1415571, 68870, 470376, 1531, 1333957, 834102, 674685, 839119, 255119, 552181, 1239034, 490552, 468715, 963242, 1612174, 21991, 1157603, 975554, 1342832, 883818, 1475260, 1666539, 1252155, 986632, 906429, 1495805, 561464, 1219097, 567704, 1463586, 1571831, 689821, 481310, 1407520, 97442, 512704, 962995, 876670, 956570, 1292534, 914981, 1655195, 870222, 196978, 617479, 1610748, 128227, 213356, 1590232, 230140, 942835]}
\ No newline at end of file
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..98e3d5ca3f5e195005c1c25eeac4b9bb4d22af2e
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a16b44c6b021168c0f693f912a5ee5e4b70f36af4117310fd02e24270d9f1c05
+size 353036
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..7bb380be1d5187b2b3cbd8d851c12b850ffcaebe
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b93630acabb9dbd3a0ff6cd548f21ecb0a93e52472e96aee9944ca7e27dcbeb7
+size 411545
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..81d2db7289ddd86cd00a97c286ef2d6767a8c32f
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfd62e8d38f6934be868c1e88b52ab285fa24a02410997b13be3ad1b8b748eca
+size 97001
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..3db63657cb3c7c512a069ebe93e00f9664209c18
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51060cf1b97515be97b1c3768fb736907bb2c5125989a5dcb2d847edf1c29bb6
+size 107123
diff --git a/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/training_log_49633f62-3f6e-4f99-aa92-a1a5f4280a83.txt b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/training_log_49633f62-3f6e-4f99-aa92-a1a5f4280a83.txt
new file mode 100644
index 0000000000000000000000000000000000000000..23163ab62a0a88eace6c95a6a52c4495554c142b
--- /dev/null
+++ b/logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/training_log_49633f62-3f6e-4f99-aa92-a1a5f4280a83.txt
@@ -0,0 +1,5236 @@
+[2025-09-04 13:01:25] [Rank 0] PRINT: --- Script Start: Thu Sep  4 13:01:25 2025 ---
+[2025-09-04 13:01:25] [Rank 0] PRINT: --- Script Start: Thu Sep  4 13:01:25 2025 ---
+[2025-09-04 13:01:25] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=46, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 13:01:25] [Rank 0] PRINT: Parsed CLI args: Namespace(unet=False, seed=46, optimizer_mode=10, model_parameterization='qkvo', per_group_k=100, muon_lr=0.002, adam_lr=0.002, base_dir='logs_qa_muon/diff_modes', sgd_lr=0.01, m_val=15, qa_jsonl_path='/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl')
+[2025-09-04 13:01:25] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 13:01:25] [Rank 0] PRINT: Hyperparameters: Hyperparameters()
+[2025-09-04 13:01:26] [Rank 0] PRINT: Using fixed seed: 46
+[2025-09-04 13:01:26] [Rank 0] PRINT: Using fixed seed: 46
+[2025-09-04 13:01:26] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46
+[2025-09-04 13:01:26] [Rank 0] PRINT: Run directory: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46
+[2025-09-04 13:01:26] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 13:01:26] [Rank 0] import os
+import sys
+with open(sys.argv[0]) as f:
+    code = f.read() # read the code of this file ASAP, for logging
+import uuid
+import time
+import copy
+import glob
+import math
+from dataclasses import dataclass, asdict
+from functools import lru_cache
+from pathlib import Path
+import argparse # Keep argparse for --unet and potentially --optimizer_mode
+import json
+import random 
+import numpy as np 
+import itertools
+from itertools import cycle
+from transformers import GPT2Tokenizer
+from collections import defaultdict
+import matplotlib.pyplot as plt
+from matplotlib.colors import Normalize
+from tqdm import tqdm
+import re
+
+
+# 
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import torch
+torch.empty(1, device="cuda", requires_grad=True).backward() # prevents a bug on some systems
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.distributed as dist
+# use of FlexAttention contributed by @KoszarskyB
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+sys.path.append("/home/aiops/zhangfz/MUON_theory_copy/MUON_theory/modded-nanogpt") # Already present
+from optimizers.MUON import Muon
+from utils.float_compute import mm_op, backward as mm_backward_custom, setup_context as mm_setup_context_custom # Renamed
+
+#from kn_util.utils import setup_debugpy
+#torch._inductor.config.coordinate_descent_tuning = True
+
+# -----------------------------------------------------------------------------
+
+mm_op.register_autograd(mm_backward_custom, setup_context=mm_setup_context_custom) # Use renamed imports
+
+# -----------------------------------------------------------------------------
+# Seeding Function
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+    print(f"PRINT: Set seed to {seed}", flush=True) # Print immediately for all ranks
+
+
+
+# -----------------------------------------------------------------------------
+# Our own simple Distributed Data Loader (KEEP AS IS)
+def _load_data_shard(file: Path):
+    header = torch.from_file(str(file), False, 256, dtype=torch.int32)
+    assert header[0] == 20240520, "magic number mismatch in the data .bin file"
+    assert header[1] == 1, "unsupported version"
+    num_tokens = int(header[2])
+    with file.open("rb", buffering=0) as f:
+        tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
+        f.seek(256 * 4)
+        nbytes = f.readinto(tokens.numpy())
+        assert nbytes == 2 * num_tokens, "number of tokens read does not match header"
+    return tokens
+
+def distributed_data_generator(filename_pattern: str, batch_size: int, rank : int, world_size : int):
+    files = [Path(file) for file in sorted(glob.glob(filename_pattern))]
+    assert batch_size % world_size == 0
+    local_batch_size = batch_size // world_size
+    file_iter = cycle(files) # use itertools.cycle(files) instead if you want to do multi-epoch training
+    tokens, pos = _load_data_shard(next(file_iter)), 0
+    while True:
+        if pos + batch_size + 1 >= len(tokens):
+            tokens, pos = _load_data_shard(next(file_iter)), 0
+        buf = tokens[pos + rank * local_batch_size:][:local_batch_size + 1]
+        inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # no sync on host side;
+        targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # H2D in another stream isn't helpful.
+        pos += batch_size
+        yield inputs, targets
+
+
+
+
+
+# -----------------------------------------------------------------------------
+# int main
+parser = argparse.ArgumentParser(description="NanoGPT Training Script with Muon")
+parser.add_argument("--unet", action="store_true", help="Use U-net architecture")
+parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
+# --- MODIFICATION: Add optimizer_mode as a CLI argument ---
+parser.add_argument("--optimizer_mode", type=int, default=0,
+                    help="Defines how Muon is applied. "
+                         "0: Muon(All Hidden Attn+MLP - original); "
+                         "1: Muon(QK Attn)/Adam(VO Attn,MLP); "
+                         "2: Muon(VO Attn)/Adam(QK Attn,MLP); "
+                         "3: Muon(All Attn)/Adam(MLP); "
+                         "4: Muon(MLP)/Adam(All Attn)"
+                         "5: All Adam (No Muon, all applicable matrices to Adam)."
+                         "6: Muon(W_2 MLP)/Adam(attn, W_1 MLP)."
+                         "7: Muon(VO Attn, MLP)/Adam(QK Attn)."
+                         "8: Muon(VO Attn, W_2 MLP)/Adam(QK Attn, W_1 MLP)."
+                         )
+parser.add_argument("--model_parameterization", type=str, default="whole",choices=["whole","qkvo"])
+parser.add_argument("--per_group_k", type=int, default=100, help="Number of samples per group")
+parser.add_argument("--muon_lr", type=float, default=0.01, help="Learning rate for Muon optimizer.")
+parser.add_argument("--adam_lr", type=float, default=1e-3, help="Base learning rate for Adam optimizer groups.")
+parser.add_argument("--base_dir", type=str, default="logs_all_0821/gated", help="Base directory for logs")
+parser.add_argument("--sgd_lr", type=float, default=0.01, help="Learning rate for SGD optimizer (used in mode 9).")
+parser.add_argument("--m_val", type=int, default=15,
+                    help="Power-law exponent m used by the dataset generator.")
+parser.add_argument("--qa_jsonl_path", type=str,
+                    default="/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15.jsonl",
+                    help="Path to the QA jsonl used for evaluation (fixed eval set).")
+
+
+exp_args = parser.parse_args()
+set_seed(exp_args.seed)
+
+M_FOR_POWERLAW: int = exp_args.m_val
+QA_JSONL_PATH: str = exp_args.qa_jsonl_path
+PER_GROUP_K: int = exp_args.per_group_k
+
+# --- MODIFICATION: Import correct GPT model based on --unet flag ---
+if exp_args.unet:
+    print("Using U-net architecture")
+    from models.nano_GPT_unet import GPT
+elif exp_args.model_parameterization == "qkvo":
+    print("Using architecture (models.nano_gpt_qkvo) with CausalSelfAttention having q_w, k_w, v_w")
+    # This MUST be the nano_GPT.py file where CausalSelfAttention has q_w, k_w, v_w
+    from models.nano_GPT_qkvo import GPT
+elif exp_args.model_parameterization == "whole":
+    print("Using original architecture")
+    from models.nano_GPT import GPT
+    
+@dataclass
+class Hyperparameters:
+    # data
+    #train_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_train_*.bin"
+    #val_files = "/home/aiops/zhangfz/MUON_theory/modded-nanogpt/data/fineweb10B/fineweb_val_*.bin"
+    train_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/train_data/train_*.bin"
+    val_files = "/home/aiops/zhangfz/MUON_theory_copy/qa_m15/qa_tail_m15_bin/val_data/val_*.bin"
+    #val_tokens = 1966080 
+    #val_tokens = 10485760
+    #train_seq_len = 12*1024
+    #val_seq_len = 4*16*1024
+    #train_seq_len = 48*1024 # FlexAttention sequence length
+    #train_seq_len = 12*1024 # FlexAttention sequence length
+    #val_seq_len = 4*64*1024 # FlexAttention sequence length for validation
+    #lr_warmup_steps = 1000      
+    #learning_rate = 0.001       
+    #min_learning_rate = 0.0001  
+
+    val_tokens = 491520 
+    train_seq_len = 3*1024
+    val_seq_len = 4*4*1024
+    #train_seq_len = 512
+    #val_seq_len = 512
+    # optimization
+    num_iterations = 10000 #1770 # Original: 1770
+    cooldown_frac = 0.8
+    # architecture
+    vocab_size = 50257
+    #vocab_size = 7
+    # evaluation and logging
+    val_loss_every = 500 # Original: 125
+    save_checkpoint = False # Original: False
+args = Hyperparameters()
+
+# DDP setup (KEEP AS IS, but ensure rank and world_size are correctly used)
+rank = int(os.environ.get("RANK", 0))
+local_rank = int(os.environ.get("LOCAL_RANK", 0)) # Used for device setting
+world_size = int(os.environ.get("WORLD_SIZE", 1))
+
+# print(f"[Rank {rank}] Global Rank: {rank}, Local Rank: {local_rank}, World Size: {world_size}", flush=True) # Debug
+
+assert torch.cuda.is_available()
+device = torch.device("cuda", local_rank) # Use local_rank for device
+torch.cuda.set_device(device)
+
+if not dist.is_initialized(): # Ensure DDP is initialized only once
+    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size) # Pass rank and world_size
+dist.barrier()
+master_process = (rank == 0)
+
+# Logging setup (KEEP AS IS, but maybe add optimizer_mode to filename)
+logfile = None
+# --- MODIFICATION: Add optimizer_mode to log file name and specify new dir ---
+#log_dir = "modded-nanogpt/logs_detailed_attn_minimal_changes"
+#if master_process:
+#    run_id = uuid.uuid4()
+#    os.makedirs(log_dir, exist_ok=True) # Create new log directory
+#    logfile = f"{log_dir}/exp_mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_{run_id}.txt"
+#    print(f"Logging to: {logfile}")
+
+logfile = None
+# run_dir_path_str = f"/home/wangshuche/MUON_theory/modded-nanogpt/logs_bios/qa/mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+# run_dir_path = Path(run_dir_path_str)
+run_dir_path_str = None
+base_log_dir = Path(exp_args.base_dir)
+# Base log directory for bioS mixed training
+
+if master_process:
+    # Set seed again specifically for master process for operations like dir creation, config saving
+    set_seed(exp_args.seed)
+
+    # Construct folder name based on config and seed
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.adam_lr}_seed_{exp_args.seed}"
+    # run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_lr_{exp_args.sgd_lr}_seed_{exp_args.seed}"
+    run_folder_name = f"mode_{exp_args.optimizer_mode}_param_{exp_args.model_parameterization}_seed_{exp_args.seed}"
+    run_dir_path = base_log_dir / run_folder_name
+    run_dir_path.mkdir(parents=True, exist_ok=True)
+    run_dir_path_str = str(run_dir_path)
+
+    run_uuid = uuid.uuid4() 
+    logfile = run_dir_path / f"training_log_{run_uuid}.txt" 
+    print(f"Logging to: {logfile}")
+
+    # Save configuration
+    config_to_save = {
+        "cli_args": vars(exp_args),
+        "hyperparameters": {k: v for k, v in args.__class__.__dict__.items() if not k.startswith('__') and not callable(v)}, 
+        "run_uuid_for_log": str(run_uuid),
+        "script_code_logged_at_start": True
+    }
+    config_file_path = run_dir_path / "config.json"
+    with open(config_file_path, "w") as f:
+        json.dump(config_to_save, f, indent=4)
+    print(f"Saved configuration to: {config_file_path}")
+
+def print0(s, console=False):
+    if master_process:
+        # Add timestamp and rank for better log readability
+        timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
+        log_message = f"[{timestamp}] [Rank {rank}] {s}"
+        
+        # Print to console if requested or if it's a specific "PRINT:" message
+        if console or s.startswith("PRINT:"):
+            actual_s = s[6:] if s.startswith("PRINT:") else s
+            print(actual_s) # Print to stdout for master process
+
+        if logfile:
+            with open(logfile, "a") as f:
+                f.write(log_message + "\n")
+
+        with open(logfile, "a") as f:
+            f.write(log_message + "\n")
+
+
+print0(f"PRINT: --- Script Start: {time.ctime()} ---", console=True) 
+print0(f"PRINT: Parsed CLI args: {exp_args}", console=True)
+print0(f"PRINT: Hyperparameters: {args}", console=True) 
+print0(f"PRINT: Using fixed seed: {exp_args.seed}", console=True) 
+if master_process: 
+     print0(f"PRINT: Run directory: {run_dir_path_str}", console=True)
+print0(code) # Log the code
+# ... (other initial logs)
+
+
+
+# -----------------------------------------------------------------------------
+
+def generate_powerlaw_selection_counts(m: int):
+    """Construct class sample counts to match the paper's distribution."""
+    selection_counts = {}
+    class_groups = []
+    class_id = 0
+    for group_id in range(m + 1):
+        if group_id == 0: num_classes = 1
+        else: num_classes = 2 ** (group_id - 1)
+        samples_per_class = 2 ** (m - group_id)
+        if samples_per_class < 1: continue
+        for _ in range(num_classes):
+            selection_counts[class_id] = samples_per_class
+            class_groups.append(group_id)
+            class_id += 1
+    return selection_counts, class_groups
+
+
+def run_detailed_evaluation(model, tokenizer, qa_data_path, device, m_val, class_to_group_map, fixed_indices=None):
+    """
+    In a single evaluation, compute Per-Class Loss, Per-Class FTA, Total Loss, and Total FTA.
+    """
+    print0("\n--- Starting Detailed Evaluation (Loss & FTA) ---", console=True)
+    model.eval()
+
+    # 1. Load and sample data
+    #with open(qa_data_path, 'r', encoding='utf-8') as f:
+    #    qa_data = [json.loads(line) for line in f]
+    
+    #if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+    #    print0(f"Using stratified sampling to extract ~{num_samples} samples for detailed evaluation...", console=True)
+    #    data_by_class = defaultdict(list)
+    #    for item in qa_data: data_by_class[item['class_id']].append(item)
+    #    sample_ratio = num_samples / len(qa_data)
+    #    stratified_sample_data = []
+    #    for class_id, items in data_by_class.items():
+    #        num_to_sample = max(1, int(len(items) * sample_ratio))
+    #        sampled_items = random.sample(items, min(len(items), num_to_sample))
+    #        stratified_sample_data.extend(sampled_items)
+    #    qa_data = stratified_sample_data
+    #    print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+
+    qa_data = []
+    if fixed_indices is not None:
+        needed = set()
+        for arr in fixed_indices.values():
+            needed.update(arr)
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if idx in needed:
+                    try:
+                        qa_data.append(json.loads(line))
+                    except Exception:
+                        continue
+        print0(f"PRINT: Fixed-eval set loaded with {len(qa_data)} samples.", console=True)
+    else:
+        with open(qa_data_path, 'r', encoding='utf-8') as f:
+            qa_data = [json.loads(line) for line in f]
+        print0(f"PRINT: WARNING: fixed_indices is None; using all {len(qa_data)} samples (may reintroduce jitter).", console=True)
+
+
+    # 2. Initialize counters
+    group_losses = defaultdict(float)
+    group_loss_counts = defaultdict(int)  # For loss sample count
+    group_correct = defaultdict(int)
+    group_total_fta = defaultdict(int)    # For FTA sample count
+
+    # 3. Evaluation loop
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=(not master_process)):
+            if not item or 'text' not in item or not item['text']: continue
+            
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            # --- Data prep for Loss ---
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            original_len = len(tokens)
+            if original_len < 2: continue
+            
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+            
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            # --- Data prep for FTA ---
+            match = re.search(r'^(.*?\?)\s*Answer\s*:\s*(.*)$', item['text'], re.IGNORECASE)
+            if not match: continue
+            prompt, answer = match.groups()
+            prompt, answer = prompt.strip(), answer.strip()
+            if not answer: continue
+            
+            try:
+                expected_token = tokenizer.encode(' ' + answer, add_special_tokens=False)[0]
+            except IndexError:
+                continue
+
+            # --- Model call (once only) ---
+            logits = model(input_seq, target_seq=None, sliding_window_num_blocks=window_blocks)
+            if isinstance(logits, tuple): logits = logits[0]
+
+            # --- Compute Loss ---
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target_seq.view(-1), ignore_index=-100)
+            if not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_loss_counts[group_id] += 1
+            
+            # --- Compute FTA ---
+            prompt_tokens_len = len(tokenizer.encode(prompt, add_special_tokens=False))
+            if prompt_tokens_len > 0 and prompt_tokens_len <= padded_len:
+                last_token_logits = logits.squeeze(0)[prompt_tokens_len - 1, :]
+                predicted_token = torch.argmax(last_token_logits).item()
+                
+                if predicted_token == expected_token:
+                    group_correct[group_id] += 1
+                group_total_fta[group_id] += 1
+
+    # 4. Aggregate results
+    avg_group_loss = {str(g): group_losses[g] / group_loss_counts[g] for g in group_loss_counts if group_loss_counts[g] > 0}
+    avg_group_acc = {str(g): group_correct[g] / group_total_fta[g] for g in group_total_fta if group_total_fta[g] > 0}
+    
+    total_loss = sum(group_losses.values()) / sum(group_loss_counts.values()) if sum(group_loss_counts.values()) > 0 else 0
+    
+    # Two methods for calculating total accuracy
+    total_acc_weighted = sum(group_correct.values()) / sum(group_total_fta.values()) if sum(group_total_fta.values()) > 0 else 0  # Original method: weighted by samples
+    total_acc_unweighted = sum(avg_group_acc.values()) / len(avg_group_acc) if avg_group_acc else 0  # New method: simple average across groups
+    
+    print0("--- Detailed Evaluation Complete ---", console=True)
+    return {
+        'per_class_loss': avg_group_loss,
+        'per_class_acc': avg_group_acc,
+        'total_loss': total_loss,
+        'total_acc_weighted': total_acc_weighted,      # Sample-weighted total accuracy
+        'total_acc_unweighted': total_acc_unweighted,  # Simple average total accuracy across groups
+        'total_acc': total_acc_unweighted              # Primarily use simple average method
+    }
+
+def plot_curves(history, output_path, title, y_label, y_lim=None):
+    """Generic plotting function"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not history:
+        print0(f"Warning: No history data for {y_label}, cannot plot.", console=True)
+        plt.close()
+        return
+    
+    is_per_class = isinstance(next(iter(history.values())), dict)
+    
+    if is_per_class:
+        group_ids = sorted([int(g) for g in history.keys()])
+        cmap = plt.get_cmap("viridis")
+        norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+        for group_id_int in group_ids:
+            group_id_str = str(group_id_int)
+            epoch_data = history[group_id_str]
+            epochs = sorted([int(e) for e in epoch_data.keys()])
+            values = [epoch_data[str(e)] for e in epochs]
+            ax.plot(epochs, values, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+        ax.legend(title="Class Group", bbox_to_anchor=(1.05, 1), loc='upper left')
+    else:
+        epochs = sorted([int(e) for e in history.keys()])
+        values = [history[str(e)] for e in epochs]
+        ax.plot(epochs, values, linewidth=2.5)
+
+    ax.set_xlabel("Epoch", fontsize=14)
+    ax.set_ylabel(y_label, fontsize=14)
+    ax.set_title(title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    
+    if y_lim:
+        ax.set_ylim(y_lim)
+    else:
+        all_values = []
+        if is_per_class:
+            for group_data in history.values(): all_values.extend(group_data.values())
+        else:
+            all_values = list(history.values())
+        if all_values:
+            min_val, max_val = min(all_values), max(all_values)
+            ax.set_ylim(min_val * 0.95, max_val * 1.05)
+            
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"[✓] {title} curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+def evaluate_per_class_loss(model, tokenizer, qa_data_path, device, m_val, num_samples=None):
+    """
+    Internal evaluation on original QA data for per-class loss.
+    (Final fixed version: NameError resolved)
+    """
+    print0("\n--- Starting Per-Class Loss Evaluation (Final Fixed Version) ---", console=True)
+    model.eval()
+
+    with open(qa_data_path, 'r', encoding='utf-8') as f:
+        qa_data = [json.loads(line) for line in f]
+    
+    if num_samples is not None and num_samples > 0 and len(qa_data) > num_samples:
+        print0(f"Using stratified sampling to extract ~{num_samples} samples for evaluation...", console=True)
+        data_by_class = defaultdict(list)
+        for item in qa_data:
+            data_by_class[item['class_id']].append(item)
+        sample_ratio = num_samples / len(qa_data)
+        stratified_sample_data = []
+        for class_id, items in data_by_class.items():
+            num_to_sample = max(1, int(len(items) * sample_ratio))
+            sampled_items = random.sample(items, min(len(items), num_to_sample))
+            stratified_sample_data.extend(sampled_items)
+        qa_data = stratified_sample_data
+        print0(f"Evaluation set size after sampling: {len(qa_data)}", console=True)
+    # =================================================================
+
+    # 3. Create mapping
+    selection_counts, class_groups = generate_powerlaw_selection_counts(m_val)
+    class_to_group_map = {class_id: group_id for class_id, group_id in zip(selection_counts.keys(), class_groups)}
+
+    group_losses = defaultdict(float)
+    group_counts = defaultdict(int)
+
+    with torch.no_grad():
+        for item in tqdm(qa_data, desc="Detailed Evaluation", disable=not master_process):
+            if not item or 'text' not in item or not item['text']: continue
+            group_id = class_to_group_map.get(item['class_id'])
+            if group_id is None: continue
+
+            tokens = tokenizer.encode(item['text'], add_special_tokens=False)
+            tokens.append(tokenizer.eos_token_id)
+            
+            original_len = len(tokens)
+            if original_len < 2: continue
+
+            BLOCK_SIZE = 128
+            padded_len = ((original_len + BLOCK_SIZE - 1) // BLOCK_SIZE) * BLOCK_SIZE
+            max_eval_len = 4096
+            padded_len = min(padded_len, max_eval_len)
+            
+            final_tokens = tokens[:padded_len]
+            pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
+            padded_input = final_tokens + [pad_token_id] * (padded_len - len(final_tokens))
+            
+            input_seq = torch.tensor(padded_input, dtype=torch.long, device=device)
+            
+            target_seq_list = (tokens[1:] + [pad_token_id])[:padded_len]
+            target_seq_list += [-100] * (padded_len - len(target_seq_list))
+            target_seq = torch.tensor(target_seq_list, dtype=torch.long, device=device)
+
+            window_blocks = torch.tensor(padded_len // BLOCK_SIZE, device=device, dtype=torch.int32)
+
+            loss = model(input_seq, target_seq, window_blocks)
+            
+            if loss is not None and not torch.isnan(loss):
+                group_losses[group_id] += loss.item()
+                group_counts[group_id] += 1
+
+    avg_group_losses = {str(group): group_losses[group] / group_counts[group]
+                        for group in group_losses if group_counts[group] > 0}
+    
+    print0("--- Per-Class Loss Evaluation Complete ---", console=True)
+    return avg_group_losses
+
+def plot_loss_curves(loss_history, output_path, plot_title="Per-Class Loss"):
+    """Plot loss curve from aggregated history data"""
+    plt.style.use('seaborn-v0_8-whitegrid')
+    fig, ax = plt.subplots(figsize=(8, 6))
+    if not loss_history:
+        print0("Warning: Loss history is empty. Cannot plot.", console=True)
+        plt.close()
+        return
+    group_ids = sorted([int(g) for g in loss_history.keys()])
+    cmap = plt.get_cmap("viridis")
+    norm = Normalize(vmin=min(group_ids) if group_ids else 0, vmax=max(group_ids) if group_ids else 1)
+    for group_id_int in group_ids:
+        group_id_str = str(group_id_int)
+        epoch_data = loss_history[group_id_str]
+        epochs = sorted([int(e) for e in epoch_data.keys()])
+        losses = [epoch_data[str(e)] for e in epochs]
+        ax.plot(epochs, losses, color=cmap(norm(group_id_int)), linewidth=2.0, label=f'Group {group_id_int}')
+    ax.set_xlabel("Step", fontsize=14)
+    ax.set_ylabel("Per-Class Loss", fontsize=14)
+    ax.set_title(plot_title, fontsize=16)
+    ax.tick_params(axis='both', which='major', labelsize=12)
+    all_losses = [loss for group_data in loss_history.values() for loss in group_data.values()]
+    if all_losses:
+        min_loss, max_loss = min(all_losses), max(all_losses)
+        ax.set_ylim(min_loss * 0.95, max_loss * 1.05)
+    ax.legend(title="Class Group")
+    ax.grid(True)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=300)
+    print0(f"Per-Class Loss curve updated and saved to: {output_path}", console=True)
+    plt.close()
+
+
+
+
+
+
+########################################
+#    Construct model and optimizer     #
+########################################
+
+print0("PRINT: Constructing model...", console=True)
+model: nn.Module = GPT(vocab_size=args.vocab_size, num_layers=12, num_heads=6, model_dim=768,
+                       max_seq_len=max(args.train_seq_len, args.val_seq_len)).cuda()
+for m in model.modules():
+    if isinstance(m, nn.Embedding):
+        m.bfloat16()
+print0("PRINT: Broadcasting model parameters...", console=True)
+for param in model.parameters():
+    dist.broadcast(param.detach(), 0)
+print0("PRINT: Model constructed and broadcasted.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model forward function:", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)
+        model.train()
+        
+        print0(f"PRINT: Model test - Result type: {type(result)}", console=True)
+        if isinstance(result, tuple):
+            print0(f"PRINT: Model test - Tuple length: {len(result)}", console=True)
+            if len(result) >= 2:
+                print0(f"PRINT: Model test - First element (loss): {result[0]}", console=True)
+                print0(f"PRINT: Model test - Second element shape (logits): {result[1].shape if hasattr(result[1], 'shape') else 'No shape'}", console=True)
+        else:
+            print0(f"PRINT: Model test - Single result shape: {result.shape if hasattr(result, 'shape') else 'No shape'}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test failed: {e}", console=True)
+
+
+model_for_inference = model  
+print0("PRINT: Saved original model reference for inference.", console=True)
+
+
+if master_process:
+    print0("PRINT: Testing model with target_seq=None...", console=True)
+    try:
+        test_input = torch.randint(0, 1000, (128,), device=device, dtype=torch.int32)
+        test_blocks = torch.tensor(1, device=device)
+        model.eval()
+        with torch.no_grad():
+            result = model(test_input, None, test_blocks)  # target_seq=None
+        model.train()
+        
+        if isinstance(result, tuple) and len(result) == 2:
+            loss, logits = result
+            print0(f"PRINT: SUCCESS! Model returns (loss={loss}, logits.shape={logits.shape})", console=True)
+        else:
+            print0(f"PRINT: Model returns: {type(result)}", console=True)
+    except Exception as e:
+        print0(f"PRINT: Model test still fails: {e}", console=True)
+
+
+
+# --- START MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+if exp_args.model_parameterization == "qkvo":
+    print0("PRINT: Collecting parameters for optimizers...", console=True)
+    head_params = [model.lm_head.weight]
+    embed_params = [model.embed.weight] + [ve.weight for ve in model.value_embeds]
+
+    # Granular collection for attention and MLP parts
+    attn_q_params = []
+    attn_k_params = []
+    attn_v_params = []
+    attn_o_params = [] # W_O from c_proj
+    mlp_fc_params = []
+    mlp_proj_params = []
+
+    for block_module in model.blocks:
+        if block_module.attn is not None:
+            # These attributes (q_w, k_w, v_w) MUST exist in your CausalSelfAttention class
+            if hasattr(block_module.attn, 'q_w'): attn_q_params.append(block_module.attn.q_w)
+            else: print0(f"PRINT: Warning: q_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'k_w'): attn_k_params.append(block_module.attn.k_w)
+            else: print0(f"PRINT: Warning: k_w not found in attn module of a block.", console=True)
+            if hasattr(block_module.attn, 'v_w'): attn_v_params.append(block_module.attn.v_w)
+            else: print0(f"PRINT: Warning: v_w not found in attn module of a block.", console=True)
+            attn_o_params.append(block_module.attn.c_proj.weight)
+        if block_module.mlp is not None:
+            mlp_fc_params.append(block_module.mlp.c_fc.weight)
+            mlp_proj_params.append(block_module.mlp.c_proj.weight)
+
+    # Combine into logical groups for experiments
+    attn_qk_group = attn_q_params + attn_k_params
+    attn_vo_group = attn_v_params + attn_o_params
+    all_attn_matrices = attn_qk_group + attn_vo_group
+    mlp_w1_group = mlp_fc_params
+    mlp_w2_group = mlp_proj_params
+    all_mlp_matrices = mlp_fc_params + mlp_proj_params
+
+    # Scalar parameters (all others not explicitly grouped as matrices)
+    matrix_params_for_scalar_check = set(head_params + embed_params + all_attn_matrices + all_mlp_matrices)
+    scalar_params = [p for n, p in model.named_parameters() if p not in matrix_params_for_scalar_check]
+    for p_scalar in scalar_params: # Sanity check
+        if p_scalar.ndim >=2:
+            print0(f"PRINT: Warning - Parameter {p_scalar.shape} ended up in scalar_params but has ndim >= 2. Check grouping.", console=True)
+
+
+    # Determine parameter distribution based on optimizer_mode
+    muon_params_target_list = []
+    adam_matrix_target_list = [] # Matrices that Adam will handle specifically
+    adam_matrix_lr = exp_args.adam_lr  # LR for matrices if Adam handles them (can be tuned)
+    muon_lr = exp_args.muon_lr
+
+    current_optimizer_mode = exp_args.optimizer_mode
+    print0(f"PRINT: Configuring optimizers for EXPERIMENT_MODE = {current_optimizer_mode}", console=True)
+
+    if current_optimizer_mode == 0: # Original behavior: Muon on all "hidden_matrix_params"
+        print0(f"PRINT: Mode 0: Muon on ALL Attention (QKVO) and ALL MLP matrices.", console=True)
+        muon_params_target_list = all_attn_matrices + all_mlp_matrices
+        # Adam handles embeds, head, scalars by default. No extra matrices for Adam here.
+    elif current_optimizer_mode == 1: # Muon on QK, Adam on VO and MLP
+        print0(f"PRINT: Mode 1: Muon on QK Attn. Adam on VO Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_qk_group
+        adam_matrix_target_list = attn_vo_group + all_mlp_matrices
+    elif current_optimizer_mode == 2: # Muon on VO, Adam on QK and MLP
+        print0(f"PRINT: Mode 2: Muon on VO Attn. Adam on QK Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group
+        adam_matrix_target_list = attn_qk_group + all_mlp_matrices
+    elif current_optimizer_mode == 3: # Muon on All Attn (QKVO), Adam on MLP
+        print0(f"PRINT: Mode 3: Muon on ALL Attn (QKVO). Adam on MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_attn_matrices
+        adam_matrix_target_list = all_mlp_matrices
+    elif current_optimizer_mode == 4: # Muon on MLP, Adam on All Attn (QKVO)
+        print0(f"PRINT: Mode 4: Muon on MLP. Adam on ALL Attn (QKVO) (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = all_mlp_matrices
+        adam_matrix_target_list = all_attn_matrices
+    elif current_optimizer_mode == 5: # NEW MODE 5 - All Adam
+        print0(f"PRINT: Mode 5: All Adam. All Attn and MLP matrices to Adam (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = [] 
+        adam_matrix_target_list = all_attn_matrices + all_mlp_matrices # All matrices to Adam
+    elif current_optimizer_mode == 6: # Muon on W_2 MLP, Adam on attn, W_1 MLP
+        print0(f"PRINT: Mode 6: Muon on W_2 MLP. Adam on attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = mlp_w2_group
+        adam_matrix_target_list = all_attn_matrices + mlp_w1_group
+    elif current_optimizer_mode == 7: # Muon on VO Attn, MLP, Adam on QK Attn
+        print0(f"PRINT: Mode 7: Muon on VO Attn, MLP. Adam on QK Attn (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + all_mlp_matrices
+        adam_matrix_target_list = attn_qk_group
+    elif current_optimizer_mode == 8: # Muon on VO Attn, W_2 MLP, Adam on QK Attn, W_1 MLP
+        print0(f"PRINT: Mode 8: Muon on VO Attn, W_2 MLP. Adam on QK Attn, W_1 MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_vo_group + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + mlp_w1_group
+    elif current_optimizer_mode == 9: # sgd + momentum
+        # This mode uses SGD with momentum for all parameters, no Muon or Adam
+        print0(f"PRINT: Mode 9: Using pure SGD+Momentum (lr={exp_args.sgd_lr}).", console=True)
+        all_params = list(model.parameters())
+        sgd_lr = exp_args.sgd_lr  # Use learning rate from command line argument
+        optimizer1 = torch.optim.SGD(all_params, lr=sgd_lr, momentum=0.9, weight_decay=1e-4)
+        optimizer2 = None
+        optimizers = [optimizer1]
+    elif current_optimizer_mode == 10: # Muon on O Attn, MLP
+        print0(f"PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + all_mlp_matrices
+        adam_matrix_target_list = attn_v_params + attn_qk_group
+    elif current_optimizer_mode == 13:
+        print0(f"PRINT: Mode 32: Muon on W_2, W_O. Adam on V Attn, QK Attn, W_1 (Adam LR: {adam_matrix_lr}).", console=True)
+        muon_params_target_list = attn_o_params + mlp_w2_group
+        adam_matrix_target_list = attn_qk_group + attn_v_params + mlp_w1_group 
+    else:
+        raise ValueError(f"Unsupported EXPERIMENT_MODE: {current_optimizer_mode}")
+
+    # Skip Adam and Muon setup for SGD mode (9)
+    if current_optimizer_mode != 9:
+        # Adam optimizer setup
+        adam_param_groups_config = [
+            #dict(params=head_params, lr=0.22),
+            #dict(params=embed_params, lr=0.6),
+            #dict(params=scalar_params, lr=0.04) # Scalar params always go to Adam
+            dict(params=head_params, lr=exp_args.adam_lr ),
+            dict(params=embed_params, lr=exp_args.adam_lr ),
+            dict(params=scalar_params, lr=exp_args.adam_lr ) # Scalar params always go to Adam
+        ]
+        # Add matrices specifically assigned to Adam for this experiment mode
+        if adam_matrix_target_list:
+            # Ensure adam_matrix_target_list is flat and contains Parameters
+            flat_adam_matrices = [p for sublist_or_p in adam_matrix_target_list for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]) if p is not None]
+            if flat_adam_matrices: # Only add group if there are params
+                adam_param_groups_config.append(dict(params=flat_adam_matrices, lr=adam_matrix_lr))
+
+        # Filter out any Adam groups that might be empty (e.g., if scalar_params was empty)
+        adam_param_groups_config = [g for g in adam_param_groups_config if g['params']]
+        optimizer1 = torch.optim.Adam(adam_param_groups_config, betas=(0.8, 0.95), eps=1e-10, fused=True)#add weight_decay=0.01 to Adam
+        optimizers = [optimizer1] # Start with Adam
+
+        # Muon optimizer setup
+        if muon_params_target_list:
+            # Ensure muon_params_target_list is flat, unique, and contains Parameters
+            flat_unique_muon_params = []
+            seen_muon_ids = set()
+            for sublist_or_p in muon_params_target_list:
+                for p in (sublist_or_p if isinstance(sublist_or_p, list) else [sublist_or_p]):
+                    if p is not None and id(p) not in seen_muon_ids:
+                        flat_unique_muon_params.append(p)
+                        seen_muon_ids.add(id(p))
+            
+            if flat_unique_muon_params: # Only create Muon if it has parameters
+                optimizer2 = Muon(flat_unique_muon_params, lr=muon_lr, momentum=0.95, nesterov=False, ns_steps=5, rank=rank, world_size=world_size) # Pass nesterov, ns_steps
+                optimizers.append(optimizer2)
+            else:
+                print0("PRINT: Muon optimizer not created as its target parameter list was empty.", console=True)
+                optimizer2 = None # Explicitly set to None if not created
+        else:
+            print0("PRINT: Muon optimizer not created as muon_params_target_list was empty (e.g. mode where Adam handles all matrices).", console=True)
+            optimizer2 = None # Explicitly set to None
+
+    print0(f"PRINT: Optimizers configured. Total optimizers: {len(optimizers)}", console=True)
+    if optimizer2:
+        print0(f"PRINT: Muon optimizer is active with {len(flat_unique_muon_params)} parameters.", console=True)
+    # --- END MODIFIED PARAMETER COLLECTION AND OPTIMIZER SETUP ---
+elif exp_args.model_parameterization == "whole":
+    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
+    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
+    scalar_params = [p for p in model.parameters() if p.ndim < 2]
+    head_params = [model.lm_head.weight]
+
+    # init the optimizer(s)
+    adam_params = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
+    # small adam epsilon by @YouJiacheng. this is an alternate method of fixing the world_size dependence
+    # discovered by @fernbear.bsky.social https://x.com/hi_tysam/status/1879692937589875094
+    optimizer1 = torch.optim.Adam(adam_params, betas=(0.8, 0.95), eps=1e-10, fused=True)
+    optimizer2 = Muon(hidden_matrix_params, lr=0.05, momentum=0.95, rank=rank, world_size=world_size)
+    optimizers = [optimizer1, optimizer2]
+
+for opt in optimizers:
+    for group in opt.param_groups:
+        group["initial_lr"] = group["lr"]
+
+# learning rate schedule: stable then decay (KEEP AS IS, but check assert)
+def get_lr(step: int):
+    x = step / args.num_iterations # progress in training
+    # assert 0 <= x < 1 # Original assert, might fail on last step if step == num_iterations
+    # --- MODIFICATION: Adjust assert for LR schedule ---
+    if not (0 <= x <= 1): # Allow x=1 for the last step
+        x = min(max(x, 0.0), 1.0) # Clamp x if step goes beyond num_iterations
+        # print0(f"LR schedule x = {x:.4f} (step={step}) was clamped.", console=False) # Optional log
+
+    if x < 1 - args.cooldown_frac:
+        return 1.0
+    else:
+        # Ensure cooldown_frac is not zero to avoid division by zero
+        w = (1 - x) / max(args.cooldown_frac, 1e-9) 
+        return w * 1.0 + (1 - w) * 0.1
+
+
+# attention window size schedule (KEEP AS IS)
+def next_multiple_of_n(v: float | int, *, n: int):
+    return next(x for x in range(n, int(v) + 1 + n, n) if x >= v)
+@lru_cache(1)
+def get_window_size_blocks_helper(window_size: int):
+    return torch.tensor(window_size // 128, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
+def get_window_size_blocks(step: int):
+    x = step / args.num_iterations # progress in training
+    # --- MODIFICATION: Adjust assert for window size schedule ---
+    if not (0 <= x <= 1):
+        x = min(max(x, 0.0), 1.0) # Clamp x
+
+    # Ensure window_size is at least 128
+    window_size = max(128, next_multiple_of_n(1728 * x, n=128))
+    return get_window_size_blocks_helper(window_size)
+
+print0("PRINT: Compiling model with TorchInductor...", console=True)
+# Use 'model' for compilation, not 'model_compiled' before it's defined
+
+model_compiled: nn.Module = torch.compile(model, dynamic=False, mode="max-autotune")
+print0("PRINT: Model compilation complete.", console=True)
+
+########################################
+# Warmup kernels
+########################################
+print0("PRINT: Starting warmup...", console=True)
+warmup_steps = 10
+initial_state = dict(
+    model=copy.deepcopy(model_compiled.state_dict()),
+    optimizers=[copy.deepcopy(opt.state_dict()) for opt in optimizers]
+)
+
+for i in range(warmup_steps):
+    inputs = targets = torch.randint(0, args.vocab_size, size=(args.train_seq_len,), device="cuda")
+    loss = model_compiled(inputs.to(torch.int32), targets, get_window_size_blocks(0))
+    loss.backward()
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+    # Add gradient clipping for SGD mode in warmup too
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+    for opt in optimizers:
+        opt.step()
+    model_compiled.zero_grad(set_to_none=True)
+    model_compiled.load_state_dict(initial_state["model"])
+    for opt, opt_state in zip(optimizers, initial_state["optimizers"]):
+        opt.load_state_dict(opt_state)
+
+del initial_state
+print0("PRINT: Warmup complete.", console=True)
+torch.cuda.synchronize()
+
+########################################
+# Training and validation
+########################################
+print0("PRINT: Starting training...", console=True)
+train_loader = distributed_data_generator(args.train_files, world_size * args.train_seq_len, rank, world_size)
+train_loss_sum = torch.zeros(1, device=device)
+train_step_count = torch.zeros(1, device=device)
+training_time_ms = 0
+torch.cuda.synchronize()
+t0 = time.perf_counter()
+train_steps = args.num_iterations
+
+
+
+if master_process:
+    tokenizer_for_eval = GPT2Tokenizer.from_pretrained('gpt2')
+    
+    history = {
+        'per_class_loss': defaultdict(dict),
+        'per_class_acc': defaultdict(dict),
+        'total_loss': {},
+        'total_acc': {}
+    }
+
+
+        # ===== [ADD] Fixed eval set (per-group equal sampling) =====
+    FIXED_VAL_INDEX_PATH = run_dir_path / "fixed_eval_indices.json"
+    #PER_GROUP_K = 100  # Number of samples per group
+
+    def _is_valid_qa_text_for_fta(text: str) -> bool:
+        # Quick filtering for building fixed eval set, ensure parseable "?" + "Answer:"
+        if not isinstance(text, str): 
+            return False
+        return re.search(r'^(.*?\?)\s*Answer\s*:\s*(.+)$', text, re.IGNORECASE) is not None
+
+    def build_fixed_eval_indices(jsonl_path, class_to_group_map, per_group_k, seed=2025):
+        rng = random.Random(seed)
+        # Build buckets by group_id for each line, but only collect samples that can be parsed for FTA
+        buckets = defaultdict(list)  # gid -> [line_idx, ...]
+        with open(jsonl_path, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f):
+                try:
+                    item = json.loads(line)
+                except Exception:
+                    continue
+                gid = class_to_group_map.get(item.get("class_id"))
+                if gid is None:
+                    continue
+                if not _is_valid_qa_text_for_fta(item.get("text", "")):
+                    continue
+                buckets[gid].append(i)
+
+        fixed = {}
+        for gid, arr in buckets.items():
+            if len(arr) <= per_group_k:
+                fixed[str(gid)] = arr[:]  # Take all if fewer than K samples
+            else:
+                fixed[str(gid)] = rng.sample(arr, per_group_k)
+        return fixed
+
+    # You already have: QA_JSONL_PATH / M_FOR_POWERLAW
+    selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+    class_to_group_map_global = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+
+    if not FIXED_VAL_INDEX_PATH.exists():
+        fixed_idx = build_fixed_eval_indices(QA_JSONL_PATH, class_to_group_map_global, PER_GROUP_K)
+        with open(FIXED_VAL_INDEX_PATH, "w") as f:
+            json.dump(fixed_idx, f)
+        print0(f"PRINT: Built fixed eval set. Saved to {FIXED_VAL_INDEX_PATH}", console=True)
+    else:
+        print0(f"PRINT: Using existing fixed eval set: {FIXED_VAL_INDEX_PATH}", console=True)
+        # --- FIX: Load the indices if the file already exists ---
+        with open(FIXED_VAL_INDEX_PATH, "r") as f:
+            fixed_idx = json.load(f)
+    # ===== [END ADD] =====
+
+    # ------------------------------------
+    #QA_JSONL_PATH = "/home/wangshuche/MUON_theory/modded-nanogpt/BIO_dataset/data/qa_tail_m15.jsonl"
+    #M_FOR_POWERLAW = 15
+    #NUM_SAMPLES_FOR_DETAIL_EVAL = 5000
+
+
+for step in range(train_steps + 1):
+    last_step = (step == train_steps)
+
+    # --------- VALIDATION SECTION ---------
+    if step == 0 or last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0):
+        torch.cuda.synchronize()
+        if step > 0:
+            current_run_time = 1000 * (time.perf_counter() - t0)
+            training_time_ms += current_run_time
+
+        model_compiled.eval()
+        val_batch_size = world_size * args.val_seq_len
+        if args.val_tokens % val_batch_size != 0:
+            print0(f"PRINT: Warning: val_tokens ({args.val_tokens}) not perfectly divisible by val_batch_size ({val_batch_size}). Some tokens might be missed.", console=True)
+
+        val_num_steps = args.val_tokens // val_batch_size
+        val_loader = distributed_data_generator(args.val_files, val_batch_size, rank, world_size)
+        val_loss_sum = torch.zeros(1, device=device)
+        actual_val_steps = 0
+
+        with torch.no_grad():
+            for val_i in range(val_num_steps):
+                try:
+                    inputs, targets = next(val_loader)
+                    loss_val = model_compiled(inputs, targets, get_window_size_blocks(step))
+                    val_loss_sum += loss_val
+                    actual_val_steps += 1
+                except StopIteration:
+                    print0(f"PRINT: Validation data loader for '{args.val_files}' exhausted early at val_step {val_i+1}/{val_num_steps}.", console=True)
+                    break
+
+        if actual_val_steps > 0:
+            val_loss_avg = val_loss_sum / actual_val_steps
+        else:
+            val_loss_avg = torch.tensor(float('nan'), device=device)
+            print0(f"PRINT: Warning: No validation steps were completed. val_loss is NaN.", console=True)
+
+        del val_loader
+        dist.all_reduce(val_loss_avg, op=dist.ReduceOp.AVG)
+
+        if train_step_count > 0:
+            avg_train_loss = train_loss_sum / train_step_count
+            dist.all_reduce(avg_train_loss, op=dist.ReduceOp.AVG)
+            avg_train_loss = avg_train_loss.item()
+        else:
+            avg_train_loss = float('nan')
+
+        avg_step_time = training_time_ms / max(step, 1) if step > 0 else 0
+
+
+
+        avg_train_loss = float(avg_train_loss)
+        if step == 0:
+            print0(f"PRINT: step:{step}/{train_steps} val_loss:{val_loss_avg.item():.4f}  train_time:{training_time_ms:.0f}ms", console=True)
+        else:
+            print0(f"PRINT: step:{step}/{train_steps} train_loss:{avg_train_loss:.4f} val_loss:{val_loss_avg.item():.4f} train_time:{training_time_ms:.0f}ms step_avg:{avg_step_time:.2f}ms", console=True)
+
+            if master_process and step > 0:
+                selection_counts, class_groups_list = generate_powerlaw_selection_counts(M_FOR_POWERLAW)
+                class_to_group_map = {cid: gid for cid, gid in zip(selection_counts.keys(), class_groups_list)}
+            
+                model_for_inference.load_state_dict(model.state_dict())
+
+                
+                eval_results = run_detailed_evaluation(
+                    model=model_for_inference,
+                    tokenizer=tokenizer_for_eval,
+                    qa_data_path=QA_JSONL_PATH,
+                    device=device,
+                    m_val=M_FOR_POWERLAW,
+                    class_to_group_map=class_to_group_map,
+                    #num_samples=NUM_SAMPLES_FOR_DETAIL_EVAL
+                    fixed_indices=fixed_idx
+                )
+
+                #
+
+
+                print0("--- Detailed Evaluation Results (This Step) ---", console=True)
+                print0(f"  Total Loss: {eval_results['total_loss']:.4f}", console=True)
+                print0(f"  Total FTA (Unweighted): {eval_results['total_acc_unweighted']:.4f}", console=True)
+                print0(f"  Total FTA (Weighted):   {eval_results['total_acc_weighted']:.4f}", console=True)
+                for group_id, loss in sorted(eval_results['per_class_loss'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} Loss: {loss:.4f}", console=True)
+                for group_id, acc in sorted(eval_results['per_class_acc'].items(), key=lambda item: int(item[0])):
+                    print0(f"  Group {group_id} FTA: {acc:.4f}", console=True)
+                
+                
+                current_step_str = str(step)
+                history['total_loss'][current_step_str] = eval_results['total_loss']
+                history['total_acc'][current_step_str] = eval_results['total_acc_unweighted']  # Use simple average method
+                for group_id, loss in eval_results['per_class_loss'].items():
+                    history['per_class_loss'][group_id][current_step_str] = loss
+                for group_id, acc in eval_results['per_class_acc'].items():
+                    history['per_class_acc'][group_id][current_step_str] = acc
+                
+               
+                plot_curves(history['per_class_loss'], run_dir_path / "per_class_loss_curves.png", "Per-Class Loss", "Loss")
+                plot_curves(history['per_class_acc'], run_dir_path / "per_class_acc_curves.png", "Per-Class FTA", "Accuracy", y_lim=[0, 1])
+                plot_curves(history['total_loss'], run_dir_path / "total_loss_curve.png", "Total Detailed Loss", "Loss")
+                plot_curves(history['total_acc'], run_dir_path / "total_acc_curve.png", "Total Detailed FTA", "Accuracy", y_lim=[0, 1])
+
+            if world_size > 1:
+                dist.barrier()
+
+
+        if master_process and args.save_checkpoint and step > 0:
+            if run_dir_path_str:
+                
+                checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+                checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+                
+                
+                checkpoint_path = checkpoint_parent_dir / f"ckpt_epoch_{step}.pt"
+                
+                log_checkpoint = dict(
+                    step=step,
+                    code=code,
+                    model=model_compiled.state_dict(),
+                    optimizers=[opt.state_dict() for opt in optimizers]
+                )
+                
+                torch.save(log_checkpoint, str(checkpoint_path))
+                print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+            else:
+                print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+
+        train_loss_sum = torch.zeros(1, device=device)
+        train_step_count = torch.zeros(1, device=device)
+        model_compiled.train()
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+
+    #if last_step:
+    #    if master_process and args.save_checkpoint:
+    #        if run_dir_path_str:
+    #            checkpoint_parent_dir = Path(run_dir_path_str) / "checkpoints"
+    #            checkpoint_parent_dir.mkdir(parents=True, exist_ok=True)
+    #            checkpoint_path = checkpoint_parent_dir / f"state_step{step:06d}.pt"
+    #            log_checkpoint = dict(
+    #                step=step,
+    #                code=code,
+    #                model=model_compiled.state_dict(),
+    #                optimizers=[opt.state_dict() for opt in optimizers]
+    #            )
+    #            torch.save(log_checkpoint, str(checkpoint_path))
+    #            print0(f"PRINT: Saved checkpoint to {checkpoint_path}", console=True)
+    #        else:
+    #            print0("PRINT: Warning - run_dir_path_str not set, cannot save checkpoint.", console=True)
+    #    break
+
+    # --------- TRAINING SECTION ---------
+    try:
+        inputs, targets = next(train_loader)
+    except StopIteration:
+        
+        print0(f"PRINT: Training data loader for '{args.train_files}' exhausted. Ending training early at step {step}.", console=True)
+        break
+
+    loss_train = model_compiled(inputs, targets, get_window_size_blocks(step))
+    loss_train.backward()
+    train_loss_sum += loss_train.detach()/ args.train_seq_len
+    train_step_count += 1
+
+    for param in model_compiled.parameters():
+        if param.grad is not None:
+            dist.all_reduce(param.grad, op=dist.ReduceOp.AVG)
+
+    # Add gradient clipping for SGD mode to prevent gradient explosion
+    if exp_args.optimizer_mode == 9:
+        torch.nn.utils.clip_grad_norm_(model_compiled.parameters(), max_norm=1.0)
+
+    current_lr_val = get_lr(step)
+    for opt in optimizers:
+        for group in opt.param_groups:
+            group["lr"] = group["initial_lr"] * current_lr_val
+
+    if optimizer2 is not None:
+        for group in optimizer2.param_groups:
+            frac = min(step / 300, 1)
+            group["momentum"] = (1 - frac) * 0.85 + frac * 0.95
+
+    for opt in optimizers:
+        opt.step()
+
+    model_compiled.zero_grad(set_to_none=True)
+
+    if step > 0 and (step % 20 == 0 or step == train_steps - 1):
+        current_segment_time_ms = 1000 * (time.perf_counter() - t0)
+        approx_total_training_time_ms = training_time_ms + current_segment_time_ms
+        total_tokens_in_batch = args.train_seq_len * world_size
+        train_loss_per_token = loss_train.item() / total_tokens_in_batch if total_tokens_in_batch > 0 else loss_train.item()
+        print0(f"step:{step+1}/{train_steps} train_time:{approx_total_training_time_ms:.0f}ms step_avg:{approx_total_training_time_ms/max(1, step + 1):.2f}ms", console=True)
+
+print0(f"PRINT: --- Training Finished: {time.ctime()} ---", console=True)
+print0(f"PRINT: Peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+       f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB", console=True)
+
+if dist.is_initialized():
+    dist.destroy_process_group()
+[2025-09-04 13:01:26] [Rank 0] PRINT: Constructing model...
+[2025-09-04 13:01:26] [Rank 0] PRINT: Constructing model...
+[2025-09-04 13:01:27] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 13:01:27] [Rank 0] PRINT: Broadcasting model parameters...
+[2025-09-04 13:01:27] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 13:01:27] [Rank 0] PRINT: Model constructed and broadcasted.
+[2025-09-04 13:01:27] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 13:01:27] [Rank 0] PRINT: Testing model forward function:
+[2025-09-04 13:01:31] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 13:01:31] [Rank 0] PRINT: Model test - Result type: <class 'torch.Tensor'>
+[2025-09-04 13:01:31] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 13:01:31] [Rank 0] PRINT: Model test - Single result shape: torch.Size([1, 128, 50304])
+[2025-09-04 13:01:31] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 13:01:31] [Rank 0] PRINT: Saved original model reference for inference.
+[2025-09-04 13:01:31] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 13:01:31] [Rank 0] PRINT: Testing model with target_seq=None...
+[2025-09-04 13:01:31] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 13:01:31] [Rank 0] PRINT: Model returns: <class 'torch.Tensor'>
+[2025-09-04 13:01:31] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 13:01:31] [Rank 0] PRINT: Collecting parameters for optimizers...
+[2025-09-04 13:01:31] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 13:01:31] [Rank 0] PRINT: Configuring optimizers for EXPERIMENT_MODE = 10
+[2025-09-04 13:01:31] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 13:01:31] [Rank 0] PRINT: Mode 10: Muon on O Attn, MLP (Adam LR: 0.002).
+[2025-09-04 13:01:31] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 13:01:31] [Rank 0] PRINT: Optimizers configured. Total optimizers: 2
+[2025-09-04 13:01:31] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 13:01:31] [Rank 0] PRINT: Muon optimizer is active with 35 parameters.
+[2025-09-04 13:01:31] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 13:01:31] [Rank 0] PRINT: Compiling model with TorchInductor...
+[2025-09-04 13:01:35] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 13:01:35] [Rank 0] PRINT: Model compilation complete.
+[2025-09-04 13:01:35] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 13:01:35] [Rank 0] PRINT: Starting warmup...
+[2025-09-04 13:02:15] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 13:02:15] [Rank 0] PRINT: Warmup complete.
+[2025-09-04 13:02:15] [Rank 0] PRINT: Starting training...
+[2025-09-04 13:02:15] [Rank 0] PRINT: Starting training...
+[2025-09-04 13:02:22] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/fixed_eval_indices.json
+[2025-09-04 13:02:22] [Rank 0] PRINT: Built fixed eval set. Saved to logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/fixed_eval_indices.json
+[2025-09-04 13:02:22] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:02:22] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:02:26] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 13:02:26] [Rank 0] PRINT: step:0/10000 val_loss:10.8258  train_time:0ms
+[2025-09-04 13:03:01] [Rank 0] step:21/10000 train_time:35369ms step_avg:1684.22ms
+[2025-09-04 13:03:01] [Rank 0] step:21/10000 train_time:35369ms step_avg:1684.22ms
+[2025-09-04 13:03:02] [Rank 0] step:41/10000 train_time:36110ms step_avg:880.74ms
+[2025-09-04 13:03:02] [Rank 0] step:41/10000 train_time:36110ms step_avg:880.74ms
+[2025-09-04 13:03:03] [Rank 0] step:61/10000 train_time:36850ms step_avg:604.10ms
+[2025-09-04 13:03:03] [Rank 0] step:61/10000 train_time:36850ms step_avg:604.10ms
+[2025-09-04 13:03:03] [Rank 0] step:81/10000 train_time:37592ms step_avg:464.10ms
+[2025-09-04 13:03:03] [Rank 0] step:81/10000 train_time:37592ms step_avg:464.10ms
+[2025-09-04 13:03:04] [Rank 0] step:101/10000 train_time:38329ms step_avg:379.50ms
+[2025-09-04 13:03:04] [Rank 0] step:101/10000 train_time:38329ms step_avg:379.50ms
+[2025-09-04 13:03:05] [Rank 0] step:121/10000 train_time:39068ms step_avg:322.88ms
+[2025-09-04 13:03:05] [Rank 0] step:121/10000 train_time:39068ms step_avg:322.88ms
+[2025-09-04 13:03:06] [Rank 0] step:141/10000 train_time:39808ms step_avg:282.33ms
+[2025-09-04 13:03:06] [Rank 0] step:141/10000 train_time:39808ms step_avg:282.33ms
+[2025-09-04 13:03:06] [Rank 0] step:161/10000 train_time:40548ms step_avg:251.85ms
+[2025-09-04 13:03:06] [Rank 0] step:161/10000 train_time:40548ms step_avg:251.85ms
+[2025-09-04 13:03:07] [Rank 0] step:181/10000 train_time:41288ms step_avg:228.11ms
+[2025-09-04 13:03:07] [Rank 0] step:181/10000 train_time:41288ms step_avg:228.11ms
+[2025-09-04 13:03:08] [Rank 0] step:201/10000 train_time:42030ms step_avg:209.10ms
+[2025-09-04 13:03:08] [Rank 0] step:201/10000 train_time:42030ms step_avg:209.10ms
+[2025-09-04 13:03:09] [Rank 0] step:221/10000 train_time:42766ms step_avg:193.51ms
+[2025-09-04 13:03:09] [Rank 0] step:221/10000 train_time:42766ms step_avg:193.51ms
+[2025-09-04 13:03:09] [Rank 0] step:241/10000 train_time:43505ms step_avg:180.52ms
+[2025-09-04 13:03:09] [Rank 0] step:241/10000 train_time:43505ms step_avg:180.52ms
+[2025-09-04 13:03:10] [Rank 0] step:261/10000 train_time:44244ms step_avg:169.52ms
+[2025-09-04 13:03:10] [Rank 0] step:261/10000 train_time:44244ms step_avg:169.52ms
+[2025-09-04 13:03:11] [Rank 0] step:281/10000 train_time:44984ms step_avg:160.09ms
+[2025-09-04 13:03:11] [Rank 0] step:281/10000 train_time:44984ms step_avg:160.09ms
+[2025-09-04 13:03:12] [Rank 0] step:301/10000 train_time:45724ms step_avg:151.91ms
+[2025-09-04 13:03:12] [Rank 0] step:301/10000 train_time:45724ms step_avg:151.91ms
+[2025-09-04 13:03:12] [Rank 0] step:321/10000 train_time:46465ms step_avg:144.75ms
+[2025-09-04 13:03:12] [Rank 0] step:321/10000 train_time:46465ms step_avg:144.75ms
+[2025-09-04 13:03:13] [Rank 0] step:341/10000 train_time:47205ms step_avg:138.43ms
+[2025-09-04 13:03:13] [Rank 0] step:341/10000 train_time:47205ms step_avg:138.43ms
+[2025-09-04 13:03:14] [Rank 0] step:361/10000 train_time:47945ms step_avg:132.81ms
+[2025-09-04 13:03:14] [Rank 0] step:361/10000 train_time:47945ms step_avg:132.81ms
+[2025-09-04 13:03:15] [Rank 0] step:381/10000 train_time:48685ms step_avg:127.78ms
+[2025-09-04 13:03:15] [Rank 0] step:381/10000 train_time:48685ms step_avg:127.78ms
+[2025-09-04 13:03:15] [Rank 0] step:401/10000 train_time:49425ms step_avg:123.25ms
+[2025-09-04 13:03:15] [Rank 0] step:401/10000 train_time:49425ms step_avg:123.25ms
+[2025-09-04 13:03:16] [Rank 0] step:421/10000 train_time:50165ms step_avg:119.16ms
+[2025-09-04 13:03:16] [Rank 0] step:421/10000 train_time:50165ms step_avg:119.16ms
+[2025-09-04 13:03:17] [Rank 0] step:441/10000 train_time:50906ms step_avg:115.43ms
+[2025-09-04 13:03:17] [Rank 0] step:441/10000 train_time:50906ms step_avg:115.43ms
+[2025-09-04 13:03:18] [Rank 0] step:461/10000 train_time:51646ms step_avg:112.03ms
+[2025-09-04 13:03:18] [Rank 0] step:461/10000 train_time:51646ms step_avg:112.03ms
+[2025-09-04 13:03:18] [Rank 0] step:481/10000 train_time:52386ms step_avg:108.91ms
+[2025-09-04 13:03:18] [Rank 0] step:481/10000 train_time:52386ms step_avg:108.91ms
+[2025-09-04 13:03:19] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:03:19] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:03:19] [Rank 0] PRINT: step:500/10000 train_loss:3.1290 val_loss:1.1168 train_time:53130ms step_avg:106.26ms
+[2025-09-04 13:03:19] [Rank 0] PRINT: step:500/10000 train_loss:3.1290 val_loss:1.1168 train_time:53130ms step_avg:106.26ms
+[2025-09-04 13:03:20] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:03:20] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:03:20] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:03:20] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:04:55] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:04:55] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:04:55] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:04:55] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:04:55] [Rank 0]   Total Loss: 3.8600
+[2025-09-04 13:04:55] [Rank 0]   Total Loss: 3.8600
+[2025-09-04 13:04:55] [Rank 0]   Total FTA (Unweighted): 0.4869
+[2025-09-04 13:04:55] [Rank 0]   Total FTA (Unweighted): 0.4869
+[2025-09-04 13:04:55] [Rank 0]   Total FTA (Weighted):   0.4869
+[2025-09-04 13:04:55] [Rank 0]   Total FTA (Weighted):   0.4869
+[2025-09-04 13:04:55] [Rank 0]   Group 0 Loss: 3.3686
+[2025-09-04 13:04:55] [Rank 0]   Group 0 Loss: 3.3686
+[2025-09-04 13:04:55] [Rank 0]   Group 1 Loss: 3.1863
+[2025-09-04 13:04:55] [Rank 0]   Group 1 Loss: 3.1863
+[2025-09-04 13:04:55] [Rank 0]   Group 2 Loss: 3.0906
+[2025-09-04 13:04:55] [Rank 0]   Group 2 Loss: 3.0906
+[2025-09-04 13:04:55] [Rank 0]   Group 3 Loss: 3.4026
+[2025-09-04 13:04:55] [Rank 0]   Group 3 Loss: 3.4026
+[2025-09-04 13:04:55] [Rank 0]   Group 4 Loss: 3.4618
+[2025-09-04 13:04:55] [Rank 0]   Group 4 Loss: 3.4618
+[2025-09-04 13:04:55] [Rank 0]   Group 5 Loss: 3.5453
+[2025-09-04 13:04:55] [Rank 0]   Group 5 Loss: 3.5453
+[2025-09-04 13:04:55] [Rank 0]   Group 6 Loss: 3.5676
+[2025-09-04 13:04:55] [Rank 0]   Group 6 Loss: 3.5676
+[2025-09-04 13:04:55] [Rank 0]   Group 7 Loss: 3.7022
+[2025-09-04 13:04:55] [Rank 0]   Group 7 Loss: 3.7022
+[2025-09-04 13:04:55] [Rank 0]   Group 8 Loss: 3.9785
+[2025-09-04 13:04:55] [Rank 0]   Group 8 Loss: 3.9785
+[2025-09-04 13:04:55] [Rank 0]   Group 9 Loss: 4.0485
+[2025-09-04 13:04:55] [Rank 0]   Group 9 Loss: 4.0485
+[2025-09-04 13:04:55] [Rank 0]   Group 10 Loss: 4.2553
+[2025-09-04 13:04:55] [Rank 0]   Group 10 Loss: 4.2553
+[2025-09-04 13:04:56] [Rank 0]   Group 11 Loss: 4.3150
+[2025-09-04 13:04:56] [Rank 0]   Group 11 Loss: 4.3150
+[2025-09-04 13:04:56] [Rank 0]   Group 12 Loss: 4.3835
+[2025-09-04 13:04:56] [Rank 0]   Group 12 Loss: 4.3835
+[2025-09-04 13:04:56] [Rank 0]   Group 13 Loss: 4.4964
+[2025-09-04 13:04:56] [Rank 0]   Group 13 Loss: 4.4964
+[2025-09-04 13:04:56] [Rank 0]   Group 14 Loss: 4.4759
+[2025-09-04 13:04:56] [Rank 0]   Group 14 Loss: 4.4759
+[2025-09-04 13:04:56] [Rank 0]   Group 15 Loss: 4.4818
+[2025-09-04 13:04:56] [Rank 0]   Group 15 Loss: 4.4818
+[2025-09-04 13:04:56] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:04:56] [Rank 0]   Group 4 FTA: 0.9400
+[2025-09-04 13:04:56] [Rank 0]   Group 4 FTA: 0.9400
+[2025-09-04 13:04:56] [Rank 0]   Group 5 FTA: 0.6100
+[2025-09-04 13:04:56] [Rank 0]   Group 5 FTA: 0.6100
+[2025-09-04 13:04:56] [Rank 0]   Group 6 FTA: 0.4900
+[2025-09-04 13:04:56] [Rank 0]   Group 6 FTA: 0.4900
+[2025-09-04 13:04:56] [Rank 0]   Group 7 FTA: 0.4300
+[2025-09-04 13:04:56] [Rank 0]   Group 7 FTA: 0.4300
+[2025-09-04 13:04:56] [Rank 0]   Group 8 FTA: 0.3900
+[2025-09-04 13:04:56] [Rank 0]   Group 8 FTA: 0.3900
+[2025-09-04 13:04:56] [Rank 0]   Group 9 FTA: 0.2200
+[2025-09-04 13:04:56] [Rank 0]   Group 9 FTA: 0.2200
+[2025-09-04 13:04:56] [Rank 0]   Group 10 FTA: 0.1300
+[2025-09-04 13:04:56] [Rank 0]   Group 10 FTA: 0.1300
+[2025-09-04 13:04:56] [Rank 0]   Group 11 FTA: 0.1000
+[2025-09-04 13:04:56] [Rank 0]   Group 11 FTA: 0.1000
+[2025-09-04 13:04:56] [Rank 0]   Group 12 FTA: 0.0900
+[2025-09-04 13:04:56] [Rank 0]   Group 12 FTA: 0.0900
+[2025-09-04 13:04:56] [Rank 0]   Group 13 FTA: 0.1600
+[2025-09-04 13:04:56] [Rank 0]   Group 13 FTA: 0.1600
+[2025-09-04 13:04:56] [Rank 0]   Group 14 FTA: 0.1200
+[2025-09-04 13:04:56] [Rank 0]   Group 14 FTA: 0.1200
+[2025-09-04 13:04:56] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 13:04:56] [Rank 0]   Group 15 FTA: 0.1100
+[2025-09-04 13:04:56] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:04:56] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:04:56] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:04:56] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:04:57] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:04:57] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:04:57] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:04:57] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:04:57] [Rank 0] step:501/10000 train_time:53144ms step_avg:106.08ms
+[2025-09-04 13:04:57] [Rank 0] step:501/10000 train_time:53144ms step_avg:106.08ms
+[2025-09-04 13:04:58] [Rank 0] step:521/10000 train_time:53902ms step_avg:103.46ms
+[2025-09-04 13:04:58] [Rank 0] step:521/10000 train_time:53902ms step_avg:103.46ms
+[2025-09-04 13:04:58] [Rank 0] step:541/10000 train_time:54642ms step_avg:101.00ms
+[2025-09-04 13:04:58] [Rank 0] step:541/10000 train_time:54642ms step_avg:101.00ms
+[2025-09-04 13:04:59] [Rank 0] step:561/10000 train_time:55382ms step_avg:98.72ms
+[2025-09-04 13:04:59] [Rank 0] step:561/10000 train_time:55382ms step_avg:98.72ms
+[2025-09-04 13:05:00] [Rank 0] step:581/10000 train_time:56122ms step_avg:96.60ms
+[2025-09-04 13:05:00] [Rank 0] step:581/10000 train_time:56122ms step_avg:96.60ms
+[2025-09-04 13:05:01] [Rank 0] step:601/10000 train_time:56862ms step_avg:94.61ms
+[2025-09-04 13:05:01] [Rank 0] step:601/10000 train_time:56862ms step_avg:94.61ms
+[2025-09-04 13:05:01] [Rank 0] step:621/10000 train_time:57647ms step_avg:92.83ms
+[2025-09-04 13:05:01] [Rank 0] step:621/10000 train_time:57647ms step_avg:92.83ms
+[2025-09-04 13:05:02] [Rank 0] step:641/10000 train_time:58422ms step_avg:91.14ms
+[2025-09-04 13:05:02] [Rank 0] step:641/10000 train_time:58422ms step_avg:91.14ms
+[2025-09-04 13:05:03] [Rank 0] step:661/10000 train_time:59163ms step_avg:89.51ms
+[2025-09-04 13:05:03] [Rank 0] step:661/10000 train_time:59163ms step_avg:89.51ms
+[2025-09-04 13:05:04] [Rank 0] step:681/10000 train_time:59903ms step_avg:87.96ms
+[2025-09-04 13:05:04] [Rank 0] step:681/10000 train_time:59903ms step_avg:87.96ms
+[2025-09-04 13:05:04] [Rank 0] step:701/10000 train_time:60644ms step_avg:86.51ms
+[2025-09-04 13:05:04] [Rank 0] step:701/10000 train_time:60644ms step_avg:86.51ms
+[2025-09-04 13:05:05] [Rank 0] step:721/10000 train_time:61384ms step_avg:85.14ms
+[2025-09-04 13:05:05] [Rank 0] step:721/10000 train_time:61384ms step_avg:85.14ms
+[2025-09-04 13:05:06] [Rank 0] step:741/10000 train_time:62125ms step_avg:83.84ms
+[2025-09-04 13:05:06] [Rank 0] step:741/10000 train_time:62125ms step_avg:83.84ms
+[2025-09-04 13:05:07] [Rank 0] step:761/10000 train_time:62869ms step_avg:82.61ms
+[2025-09-04 13:05:07] [Rank 0] step:761/10000 train_time:62869ms step_avg:82.61ms
+[2025-09-04 13:05:07] [Rank 0] step:781/10000 train_time:63613ms step_avg:81.45ms
+[2025-09-04 13:05:07] [Rank 0] step:781/10000 train_time:63613ms step_avg:81.45ms
+[2025-09-04 13:05:08] [Rank 0] step:801/10000 train_time:64357ms step_avg:80.35ms
+[2025-09-04 13:05:08] [Rank 0] step:801/10000 train_time:64357ms step_avg:80.35ms
+[2025-09-04 13:05:09] [Rank 0] step:821/10000 train_time:65371ms step_avg:79.62ms
+[2025-09-04 13:05:09] [Rank 0] step:821/10000 train_time:65371ms step_avg:79.62ms
+[2025-09-04 13:05:10] [Rank 0] step:841/10000 train_time:66115ms step_avg:78.61ms
+[2025-09-04 13:05:10] [Rank 0] step:841/10000 train_time:66115ms step_avg:78.61ms
+[2025-09-04 13:05:11] [Rank 0] step:861/10000 train_time:66859ms step_avg:77.65ms
+[2025-09-04 13:05:11] [Rank 0] step:861/10000 train_time:66859ms step_avg:77.65ms
+[2025-09-04 13:05:11] [Rank 0] step:881/10000 train_time:67604ms step_avg:76.74ms
+[2025-09-04 13:05:11] [Rank 0] step:881/10000 train_time:67604ms step_avg:76.74ms
+[2025-09-04 13:05:12] [Rank 0] step:901/10000 train_time:68348ms step_avg:75.86ms
+[2025-09-04 13:05:12] [Rank 0] step:901/10000 train_time:68348ms step_avg:75.86ms
+[2025-09-04 13:05:13] [Rank 0] step:921/10000 train_time:69092ms step_avg:75.02ms
+[2025-09-04 13:05:13] [Rank 0] step:921/10000 train_time:69092ms step_avg:75.02ms
+[2025-09-04 13:05:14] [Rank 0] step:941/10000 train_time:69836ms step_avg:74.21ms
+[2025-09-04 13:05:14] [Rank 0] step:941/10000 train_time:69836ms step_avg:74.21ms
+[2025-09-04 13:05:14] [Rank 0] step:961/10000 train_time:70579ms step_avg:73.44ms
+[2025-09-04 13:05:14] [Rank 0] step:961/10000 train_time:70579ms step_avg:73.44ms
+[2025-09-04 13:05:15] [Rank 0] step:981/10000 train_time:71324ms step_avg:72.71ms
+[2025-09-04 13:05:15] [Rank 0] step:981/10000 train_time:71324ms step_avg:72.71ms
+[2025-09-04 13:05:16] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:05:16] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:05:16] [Rank 0] PRINT: step:1000/10000 train_loss:0.9704 val_loss:0.8663 train_time:72073ms step_avg:72.07ms
+[2025-09-04 13:05:16] [Rank 0] PRINT: step:1000/10000 train_loss:0.9704 val_loss:0.8663 train_time:72073ms step_avg:72.07ms
+[2025-09-04 13:05:16] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:05:16] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:05:17] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:05:17] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:06:52] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:06:52] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:06:52] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:06:52] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:06:52] [Rank 0]   Total Loss: 4.2720
+[2025-09-04 13:06:52] [Rank 0]   Total Loss: 4.2720
+[2025-09-04 13:06:52] [Rank 0]   Total FTA (Unweighted): 0.6725
+[2025-09-04 13:06:52] [Rank 0]   Total FTA (Unweighted): 0.6725
+[2025-09-04 13:06:52] [Rank 0]   Total FTA (Weighted):   0.6725
+[2025-09-04 13:06:52] [Rank 0]   Total FTA (Weighted):   0.6725
+[2025-09-04 13:06:52] [Rank 0]   Group 0 Loss: 4.2524
+[2025-09-04 13:06:52] [Rank 0]   Group 0 Loss: 4.2524
+[2025-09-04 13:06:52] [Rank 0]   Group 1 Loss: 3.7205
+[2025-09-04 13:06:52] [Rank 0]   Group 1 Loss: 3.7205
+[2025-09-04 13:06:52] [Rank 0]   Group 2 Loss: 3.7147
+[2025-09-04 13:06:52] [Rank 0]   Group 2 Loss: 3.7147
+[2025-09-04 13:06:52] [Rank 0]   Group 3 Loss: 4.0468
+[2025-09-04 13:06:52] [Rank 0]   Group 3 Loss: 4.0468
+[2025-09-04 13:06:52] [Rank 0]   Group 4 Loss: 4.0170
+[2025-09-04 13:06:52] [Rank 0]   Group 4 Loss: 4.0170
+[2025-09-04 13:06:52] [Rank 0]   Group 5 Loss: 4.0478
+[2025-09-04 13:06:52] [Rank 0]   Group 5 Loss: 4.0478
+[2025-09-04 13:06:52] [Rank 0]   Group 6 Loss: 3.9632
+[2025-09-04 13:06:52] [Rank 0]   Group 6 Loss: 3.9632
+[2025-09-04 13:06:52] [Rank 0]   Group 7 Loss: 4.0703
+[2025-09-04 13:06:52] [Rank 0]   Group 7 Loss: 4.0703
+[2025-09-04 13:06:52] [Rank 0]   Group 8 Loss: 4.2291
+[2025-09-04 13:06:52] [Rank 0]   Group 8 Loss: 4.2291
+[2025-09-04 13:06:52] [Rank 0]   Group 9 Loss: 4.2122
+[2025-09-04 13:06:52] [Rank 0]   Group 9 Loss: 4.2122
+[2025-09-04 13:06:52] [Rank 0]   Group 10 Loss: 4.4310
+[2025-09-04 13:06:52] [Rank 0]   Group 10 Loss: 4.4310
+[2025-09-04 13:06:52] [Rank 0]   Group 11 Loss: 4.5276
+[2025-09-04 13:06:52] [Rank 0]   Group 11 Loss: 4.5276
+[2025-09-04 13:06:52] [Rank 0]   Group 12 Loss: 4.5947
+[2025-09-04 13:06:52] [Rank 0]   Group 12 Loss: 4.5947
+[2025-09-04 13:06:52] [Rank 0]   Group 13 Loss: 4.7768
+[2025-09-04 13:06:52] [Rank 0]   Group 13 Loss: 4.7768
+[2025-09-04 13:06:52] [Rank 0]   Group 14 Loss: 4.8209
+[2025-09-04 13:06:52] [Rank 0]   Group 14 Loss: 4.8209
+[2025-09-04 13:06:52] [Rank 0]   Group 15 Loss: 4.9270
+[2025-09-04 13:06:52] [Rank 0]   Group 15 Loss: 4.9270
+[2025-09-04 13:06:52] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:06:52] [Rank 0]   Group 7 FTA: 0.9100
+[2025-09-04 13:06:52] [Rank 0]   Group 7 FTA: 0.9100
+[2025-09-04 13:06:52] [Rank 0]   Group 8 FTA: 0.7600
+[2025-09-04 13:06:52] [Rank 0]   Group 8 FTA: 0.7600
+[2025-09-04 13:06:52] [Rank 0]   Group 9 FTA: 0.6000
+[2025-09-04 13:06:52] [Rank 0]   Group 9 FTA: 0.6000
+[2025-09-04 13:06:52] [Rank 0]   Group 10 FTA: 0.6100
+[2025-09-04 13:06:52] [Rank 0]   Group 10 FTA: 0.6100
+[2025-09-04 13:06:52] [Rank 0]   Group 11 FTA: 0.3200
+[2025-09-04 13:06:52] [Rank 0]   Group 11 FTA: 0.3200
+[2025-09-04 13:06:52] [Rank 0]   Group 12 FTA: 0.1600
+[2025-09-04 13:06:52] [Rank 0]   Group 12 FTA: 0.1600
+[2025-09-04 13:06:52] [Rank 0]   Group 13 FTA: 0.1600
+[2025-09-04 13:06:52] [Rank 0]   Group 13 FTA: 0.1600
+[2025-09-04 13:06:52] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 13:06:52] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 13:06:52] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 13:06:52] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 13:06:53] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:06:53] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:06:53] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:06:53] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:06:54] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:06:54] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:06:54] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:06:54] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:06:54] [Rank 0] step:1001/10000 train_time:72088ms step_avg:72.02ms
+[2025-09-04 13:06:54] [Rank 0] step:1001/10000 train_time:72088ms step_avg:72.02ms
+[2025-09-04 13:06:55] [Rank 0] step:1021/10000 train_time:72830ms step_avg:71.33ms
+[2025-09-04 13:06:55] [Rank 0] step:1021/10000 train_time:72830ms step_avg:71.33ms
+[2025-09-04 13:06:55] [Rank 0] step:1041/10000 train_time:73573ms step_avg:70.68ms
+[2025-09-04 13:06:55] [Rank 0] step:1041/10000 train_time:73573ms step_avg:70.68ms
+[2025-09-04 13:06:56] [Rank 0] step:1061/10000 train_time:74317ms step_avg:70.04ms
+[2025-09-04 13:06:56] [Rank 0] step:1061/10000 train_time:74317ms step_avg:70.04ms
+[2025-09-04 13:06:57] [Rank 0] step:1081/10000 train_time:75061ms step_avg:69.44ms
+[2025-09-04 13:06:57] [Rank 0] step:1081/10000 train_time:75061ms step_avg:69.44ms
+[2025-09-04 13:06:58] [Rank 0] step:1101/10000 train_time:75805ms step_avg:68.85ms
+[2025-09-04 13:06:58] [Rank 0] step:1101/10000 train_time:75805ms step_avg:68.85ms
+[2025-09-04 13:06:58] [Rank 0] step:1121/10000 train_time:76549ms step_avg:68.29ms
+[2025-09-04 13:06:58] [Rank 0] step:1121/10000 train_time:76549ms step_avg:68.29ms
+[2025-09-04 13:06:59] [Rank 0] step:1141/10000 train_time:77292ms step_avg:67.74ms
+[2025-09-04 13:06:59] [Rank 0] step:1141/10000 train_time:77292ms step_avg:67.74ms
+[2025-09-04 13:07:00] [Rank 0] step:1161/10000 train_time:78035ms step_avg:67.21ms
+[2025-09-04 13:07:00] [Rank 0] step:1161/10000 train_time:78035ms step_avg:67.21ms
+[2025-09-04 13:07:01] [Rank 0] step:1181/10000 train_time:78779ms step_avg:66.71ms
+[2025-09-04 13:07:01] [Rank 0] step:1181/10000 train_time:78779ms step_avg:66.71ms
+[2025-09-04 13:07:01] [Rank 0] step:1201/10000 train_time:79523ms step_avg:66.21ms
+[2025-09-04 13:07:01] [Rank 0] step:1201/10000 train_time:79523ms step_avg:66.21ms
+[2025-09-04 13:07:02] [Rank 0] step:1221/10000 train_time:80266ms step_avg:65.74ms
+[2025-09-04 13:07:02] [Rank 0] step:1221/10000 train_time:80266ms step_avg:65.74ms
+[2025-09-04 13:07:03] [Rank 0] step:1241/10000 train_time:81011ms step_avg:65.28ms
+[2025-09-04 13:07:03] [Rank 0] step:1241/10000 train_time:81011ms step_avg:65.28ms
+[2025-09-04 13:07:03] [Rank 0] step:1261/10000 train_time:81755ms step_avg:64.83ms
+[2025-09-04 13:07:03] [Rank 0] step:1261/10000 train_time:81755ms step_avg:64.83ms
+[2025-09-04 13:07:04] [Rank 0] step:1281/10000 train_time:82499ms step_avg:64.40ms
+[2025-09-04 13:07:04] [Rank 0] step:1281/10000 train_time:82499ms step_avg:64.40ms
+[2025-09-04 13:07:05] [Rank 0] step:1301/10000 train_time:83243ms step_avg:63.98ms
+[2025-09-04 13:07:05] [Rank 0] step:1301/10000 train_time:83243ms step_avg:63.98ms
+[2025-09-04 13:07:06] [Rank 0] step:1321/10000 train_time:83987ms step_avg:63.58ms
+[2025-09-04 13:07:06] [Rank 0] step:1321/10000 train_time:83987ms step_avg:63.58ms
+[2025-09-04 13:07:06] [Rank 0] step:1341/10000 train_time:84731ms step_avg:63.19ms
+[2025-09-04 13:07:06] [Rank 0] step:1341/10000 train_time:84731ms step_avg:63.19ms
+[2025-09-04 13:07:07] [Rank 0] step:1361/10000 train_time:85475ms step_avg:62.80ms
+[2025-09-04 13:07:07] [Rank 0] step:1361/10000 train_time:85475ms step_avg:62.80ms
+[2025-09-04 13:07:08] [Rank 0] step:1381/10000 train_time:86218ms step_avg:62.43ms
+[2025-09-04 13:07:08] [Rank 0] step:1381/10000 train_time:86218ms step_avg:62.43ms
+[2025-09-04 13:07:09] [Rank 0] step:1401/10000 train_time:86962ms step_avg:62.07ms
+[2025-09-04 13:07:09] [Rank 0] step:1401/10000 train_time:86962ms step_avg:62.07ms
+[2025-09-04 13:07:09] [Rank 0] step:1421/10000 train_time:87705ms step_avg:61.72ms
+[2025-09-04 13:07:09] [Rank 0] step:1421/10000 train_time:87705ms step_avg:61.72ms
+[2025-09-04 13:07:10] [Rank 0] step:1441/10000 train_time:88449ms step_avg:61.38ms
+[2025-09-04 13:07:10] [Rank 0] step:1441/10000 train_time:88449ms step_avg:61.38ms
+[2025-09-04 13:07:11] [Rank 0] step:1461/10000 train_time:89193ms step_avg:61.05ms
+[2025-09-04 13:07:11] [Rank 0] step:1461/10000 train_time:89193ms step_avg:61.05ms
+[2025-09-04 13:07:12] [Rank 0] step:1481/10000 train_time:89938ms step_avg:60.73ms
+[2025-09-04 13:07:12] [Rank 0] step:1481/10000 train_time:89938ms step_avg:60.73ms
+[2025-09-04 13:07:12] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:07:12] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:07:13] [Rank 0] PRINT: step:1500/10000 train_loss:0.8300 val_loss:0.7839 train_time:90687ms step_avg:60.46ms
+[2025-09-04 13:07:13] [Rank 0] PRINT: step:1500/10000 train_loss:0.8300 val_loss:0.7839 train_time:90687ms step_avg:60.46ms
+[2025-09-04 13:07:13] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:07:13] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:07:13] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:07:13] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:08:49] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:08:49] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:08:49] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:08:49] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:08:49] [Rank 0]   Total Loss: 4.5139
+[2025-09-04 13:08:49] [Rank 0]   Total Loss: 4.5139
+[2025-09-04 13:08:49] [Rank 0]   Total FTA (Unweighted): 0.7406
+[2025-09-04 13:08:49] [Rank 0]   Total FTA (Unweighted): 0.7406
+[2025-09-04 13:08:49] [Rank 0]   Total FTA (Weighted):   0.7406
+[2025-09-04 13:08:49] [Rank 0]   Total FTA (Weighted):   0.7406
+[2025-09-04 13:08:49] [Rank 0]   Group 0 Loss: 4.5804
+[2025-09-04 13:08:49] [Rank 0]   Group 0 Loss: 4.5804
+[2025-09-04 13:08:49] [Rank 0]   Group 1 Loss: 4.0282
+[2025-09-04 13:08:49] [Rank 0]   Group 1 Loss: 4.0282
+[2025-09-04 13:08:49] [Rank 0]   Group 2 Loss: 3.9346
+[2025-09-04 13:08:49] [Rank 0]   Group 2 Loss: 3.9346
+[2025-09-04 13:08:49] [Rank 0]   Group 3 Loss: 4.4066
+[2025-09-04 13:08:49] [Rank 0]   Group 3 Loss: 4.4066
+[2025-09-04 13:08:49] [Rank 0]   Group 4 Loss: 4.2926
+[2025-09-04 13:08:49] [Rank 0]   Group 4 Loss: 4.2926
+[2025-09-04 13:08:49] [Rank 0]   Group 5 Loss: 4.3111
+[2025-09-04 13:08:49] [Rank 0]   Group 5 Loss: 4.3111
+[2025-09-04 13:08:49] [Rank 0]   Group 6 Loss: 4.2879
+[2025-09-04 13:08:49] [Rank 0]   Group 6 Loss: 4.2879
+[2025-09-04 13:08:49] [Rank 0]   Group 7 Loss: 4.3412
+[2025-09-04 13:08:49] [Rank 0]   Group 7 Loss: 4.3412
+[2025-09-04 13:08:49] [Rank 0]   Group 8 Loss: 4.4674
+[2025-09-04 13:08:49] [Rank 0]   Group 8 Loss: 4.4674
+[2025-09-04 13:08:49] [Rank 0]   Group 9 Loss: 4.4198
+[2025-09-04 13:08:49] [Rank 0]   Group 9 Loss: 4.4198
+[2025-09-04 13:08:49] [Rank 0]   Group 10 Loss: 4.6180
+[2025-09-04 13:08:49] [Rank 0]   Group 10 Loss: 4.6180
+[2025-09-04 13:08:49] [Rank 0]   Group 11 Loss: 4.7076
+[2025-09-04 13:08:49] [Rank 0]   Group 11 Loss: 4.7076
+[2025-09-04 13:08:49] [Rank 0]   Group 12 Loss: 4.7277
+[2025-09-04 13:08:49] [Rank 0]   Group 12 Loss: 4.7277
+[2025-09-04 13:08:49] [Rank 0]   Group 13 Loss: 4.9180
+[2025-09-04 13:08:49] [Rank 0]   Group 13 Loss: 4.9180
+[2025-09-04 13:08:49] [Rank 0]   Group 14 Loss: 5.0309
+[2025-09-04 13:08:49] [Rank 0]   Group 14 Loss: 5.0309
+[2025-09-04 13:08:49] [Rank 0]   Group 15 Loss: 5.1509
+[2025-09-04 13:08:49] [Rank 0]   Group 15 Loss: 5.1509
+[2025-09-04 13:08:49] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:08:49] [Rank 0]   Group 8 FTA: 0.9400
+[2025-09-04 13:08:49] [Rank 0]   Group 8 FTA: 0.9400
+[2025-09-04 13:08:49] [Rank 0]   Group 9 FTA: 0.8000
+[2025-09-04 13:08:49] [Rank 0]   Group 9 FTA: 0.8000
+[2025-09-04 13:08:49] [Rank 0]   Group 10 FTA: 0.8200
+[2025-09-04 13:08:49] [Rank 0]   Group 10 FTA: 0.8200
+[2025-09-04 13:08:49] [Rank 0]   Group 11 FTA: 0.5900
+[2025-09-04 13:08:49] [Rank 0]   Group 11 FTA: 0.5900
+[2025-09-04 13:08:49] [Rank 0]   Group 12 FTA: 0.2800
+[2025-09-04 13:08:49] [Rank 0]   Group 12 FTA: 0.2800
+[2025-09-04 13:08:49] [Rank 0]   Group 13 FTA: 0.1800
+[2025-09-04 13:08:49] [Rank 0]   Group 13 FTA: 0.1800
+[2025-09-04 13:08:49] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 13:08:49] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 13:08:49] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 13:08:49] [Rank 0]   Group 15 FTA: 0.0800
+[2025-09-04 13:08:50] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:08:50] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:08:50] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:08:50] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:08:50] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:08:50] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:08:51] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:08:51] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:08:51] [Rank 0] step:1501/10000 train_time:90702ms step_avg:60.43ms
+[2025-09-04 13:08:51] [Rank 0] step:1501/10000 train_time:90702ms step_avg:60.43ms
+[2025-09-04 13:08:52] [Rank 0] step:1521/10000 train_time:91613ms step_avg:60.23ms
+[2025-09-04 13:08:52] [Rank 0] step:1521/10000 train_time:91613ms step_avg:60.23ms
+[2025-09-04 13:08:52] [Rank 0] step:1541/10000 train_time:92497ms step_avg:60.02ms
+[2025-09-04 13:08:52] [Rank 0] step:1541/10000 train_time:92497ms step_avg:60.02ms
+[2025-09-04 13:08:53] [Rank 0] step:1561/10000 train_time:93241ms step_avg:59.73ms
+[2025-09-04 13:08:53] [Rank 0] step:1561/10000 train_time:93241ms step_avg:59.73ms
+[2025-09-04 13:08:54] [Rank 0] step:1581/10000 train_time:93984ms step_avg:59.45ms
+[2025-09-04 13:08:54] [Rank 0] step:1581/10000 train_time:93984ms step_avg:59.45ms
+[2025-09-04 13:08:55] [Rank 0] step:1601/10000 train_time:94728ms step_avg:59.17ms
+[2025-09-04 13:08:55] [Rank 0] step:1601/10000 train_time:94728ms step_avg:59.17ms
+[2025-09-04 13:08:55] [Rank 0] step:1621/10000 train_time:95473ms step_avg:58.90ms
+[2025-09-04 13:08:55] [Rank 0] step:1621/10000 train_time:95473ms step_avg:58.90ms
+[2025-09-04 13:08:56] [Rank 0] step:1641/10000 train_time:96492ms step_avg:58.80ms
+[2025-09-04 13:08:56] [Rank 0] step:1641/10000 train_time:96492ms step_avg:58.80ms
+[2025-09-04 13:08:57] [Rank 0] step:1661/10000 train_time:97236ms step_avg:58.54ms
+[2025-09-04 13:08:57] [Rank 0] step:1661/10000 train_time:97236ms step_avg:58.54ms
+[2025-09-04 13:08:58] [Rank 0] step:1681/10000 train_time:97980ms step_avg:58.29ms
+[2025-09-04 13:08:58] [Rank 0] step:1681/10000 train_time:97980ms step_avg:58.29ms
+[2025-09-04 13:08:59] [Rank 0] step:1701/10000 train_time:98723ms step_avg:58.04ms
+[2025-09-04 13:08:59] [Rank 0] step:1701/10000 train_time:98723ms step_avg:58.04ms
+[2025-09-04 13:08:59] [Rank 0] step:1721/10000 train_time:99467ms step_avg:57.80ms
+[2025-09-04 13:08:59] [Rank 0] step:1721/10000 train_time:99467ms step_avg:57.80ms
+[2025-09-04 13:09:00] [Rank 0] step:1741/10000 train_time:100211ms step_avg:57.56ms
+[2025-09-04 13:09:00] [Rank 0] step:1741/10000 train_time:100211ms step_avg:57.56ms
+[2025-09-04 13:09:01] [Rank 0] step:1761/10000 train_time:100955ms step_avg:57.33ms
+[2025-09-04 13:09:01] [Rank 0] step:1761/10000 train_time:100955ms step_avg:57.33ms
+[2025-09-04 13:09:02] [Rank 0] step:1781/10000 train_time:101699ms step_avg:57.10ms
+[2025-09-04 13:09:02] [Rank 0] step:1781/10000 train_time:101699ms step_avg:57.10ms
+[2025-09-04 13:09:02] [Rank 0] step:1801/10000 train_time:102443ms step_avg:56.88ms
+[2025-09-04 13:09:02] [Rank 0] step:1801/10000 train_time:102443ms step_avg:56.88ms
+[2025-09-04 13:09:03] [Rank 0] step:1821/10000 train_time:103188ms step_avg:56.67ms
+[2025-09-04 13:09:03] [Rank 0] step:1821/10000 train_time:103188ms step_avg:56.67ms
+[2025-09-04 13:09:04] [Rank 0] step:1841/10000 train_time:103932ms step_avg:56.45ms
+[2025-09-04 13:09:04] [Rank 0] step:1841/10000 train_time:103932ms step_avg:56.45ms
+[2025-09-04 13:09:05] [Rank 0] step:1861/10000 train_time:104676ms step_avg:56.25ms
+[2025-09-04 13:09:05] [Rank 0] step:1861/10000 train_time:104676ms step_avg:56.25ms
+[2025-09-04 13:09:05] [Rank 0] step:1881/10000 train_time:105420ms step_avg:56.04ms
+[2025-09-04 13:09:05] [Rank 0] step:1881/10000 train_time:105420ms step_avg:56.04ms
+[2025-09-04 13:09:06] [Rank 0] step:1901/10000 train_time:106165ms step_avg:55.85ms
+[2025-09-04 13:09:06] [Rank 0] step:1901/10000 train_time:106165ms step_avg:55.85ms
+[2025-09-04 13:09:07] [Rank 0] step:1921/10000 train_time:106909ms step_avg:55.65ms
+[2025-09-04 13:09:07] [Rank 0] step:1921/10000 train_time:106909ms step_avg:55.65ms
+[2025-09-04 13:09:08] [Rank 0] step:1941/10000 train_time:107653ms step_avg:55.46ms
+[2025-09-04 13:09:08] [Rank 0] step:1941/10000 train_time:107653ms step_avg:55.46ms
+[2025-09-04 13:09:08] [Rank 0] step:1961/10000 train_time:108397ms step_avg:55.28ms
+[2025-09-04 13:09:08] [Rank 0] step:1961/10000 train_time:108397ms step_avg:55.28ms
+[2025-09-04 13:09:09] [Rank 0] step:1981/10000 train_time:109141ms step_avg:55.09ms
+[2025-09-04 13:09:09] [Rank 0] step:1981/10000 train_time:109141ms step_avg:55.09ms
+[2025-09-04 13:09:10] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:09:10] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:09:10] [Rank 0] PRINT: step:2000/10000 train_loss:0.7710 val_loss:0.7381 train_time:109891ms step_avg:54.95ms
+[2025-09-04 13:09:10] [Rank 0] PRINT: step:2000/10000 train_loss:0.7710 val_loss:0.7381 train_time:109891ms step_avg:54.95ms
+[2025-09-04 13:09:10] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:09:10] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:09:11] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:09:11] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:10:46] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:10:46] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:10:46] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:10:46] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:10:46] [Rank 0]   Total Loss: 4.5733
+[2025-09-04 13:10:46] [Rank 0]   Total Loss: 4.5733
+[2025-09-04 13:10:46] [Rank 0]   Total FTA (Unweighted): 0.7875
+[2025-09-04 13:10:46] [Rank 0]   Total FTA (Unweighted): 0.7875
+[2025-09-04 13:10:46] [Rank 0]   Total FTA (Weighted):   0.7875
+[2025-09-04 13:10:46] [Rank 0]   Total FTA (Weighted):   0.7875
+[2025-09-04 13:10:46] [Rank 0]   Group 0 Loss: 4.6571
+[2025-09-04 13:10:46] [Rank 0]   Group 0 Loss: 4.6571
+[2025-09-04 13:10:46] [Rank 0]   Group 1 Loss: 4.1598
+[2025-09-04 13:10:46] [Rank 0]   Group 1 Loss: 4.1598
+[2025-09-04 13:10:46] [Rank 0]   Group 2 Loss: 4.0351
+[2025-09-04 13:10:46] [Rank 0]   Group 2 Loss: 4.0351
+[2025-09-04 13:10:46] [Rank 0]   Group 3 Loss: 4.4490
+[2025-09-04 13:10:46] [Rank 0]   Group 3 Loss: 4.4490
+[2025-09-04 13:10:46] [Rank 0]   Group 4 Loss: 4.4026
+[2025-09-04 13:10:46] [Rank 0]   Group 4 Loss: 4.4026
+[2025-09-04 13:10:46] [Rank 0]   Group 5 Loss: 4.3866
+[2025-09-04 13:10:46] [Rank 0]   Group 5 Loss: 4.3866
+[2025-09-04 13:10:46] [Rank 0]   Group 6 Loss: 4.3898
+[2025-09-04 13:10:46] [Rank 0]   Group 6 Loss: 4.3898
+[2025-09-04 13:10:46] [Rank 0]   Group 7 Loss: 4.4265
+[2025-09-04 13:10:46] [Rank 0]   Group 7 Loss: 4.4265
+[2025-09-04 13:10:46] [Rank 0]   Group 8 Loss: 4.5591
+[2025-09-04 13:10:46] [Rank 0]   Group 8 Loss: 4.5591
+[2025-09-04 13:10:46] [Rank 0]   Group 9 Loss: 4.4698
+[2025-09-04 13:10:46] [Rank 0]   Group 9 Loss: 4.4698
+[2025-09-04 13:10:46] [Rank 0]   Group 10 Loss: 4.7009
+[2025-09-04 13:10:46] [Rank 0]   Group 10 Loss: 4.7009
+[2025-09-04 13:10:46] [Rank 0]   Group 11 Loss: 4.7424
+[2025-09-04 13:10:46] [Rank 0]   Group 11 Loss: 4.7424
+[2025-09-04 13:10:46] [Rank 0]   Group 12 Loss: 4.7711
+[2025-09-04 13:10:46] [Rank 0]   Group 12 Loss: 4.7711
+[2025-09-04 13:10:46] [Rank 0]   Group 13 Loss: 4.9293
+[2025-09-04 13:10:46] [Rank 0]   Group 13 Loss: 4.9293
+[2025-09-04 13:10:46] [Rank 0]   Group 14 Loss: 4.9851
+[2025-09-04 13:10:46] [Rank 0]   Group 14 Loss: 4.9851
+[2025-09-04 13:10:46] [Rank 0]   Group 15 Loss: 5.1086
+[2025-09-04 13:10:46] [Rank 0]   Group 15 Loss: 5.1086
+[2025-09-04 13:10:46] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:10:46] [Rank 0]   Group 9 FTA: 0.9200
+[2025-09-04 13:10:46] [Rank 0]   Group 9 FTA: 0.9200
+[2025-09-04 13:10:46] [Rank 0]   Group 10 FTA: 0.9200
+[2025-09-04 13:10:46] [Rank 0]   Group 10 FTA: 0.9200
+[2025-09-04 13:10:46] [Rank 0]   Group 11 FTA: 0.8200
+[2025-09-04 13:10:46] [Rank 0]   Group 11 FTA: 0.8200
+[2025-09-04 13:10:46] [Rank 0]   Group 12 FTA: 0.4800
+[2025-09-04 13:10:46] [Rank 0]   Group 12 FTA: 0.4800
+[2025-09-04 13:10:46] [Rank 0]   Group 13 FTA: 0.2300
+[2025-09-04 13:10:46] [Rank 0]   Group 13 FTA: 0.2300
+[2025-09-04 13:10:46] [Rank 0]   Group 14 FTA: 0.1300
+[2025-09-04 13:10:46] [Rank 0]   Group 14 FTA: 0.1300
+[2025-09-04 13:10:46] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 13:10:46] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 13:10:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:10:47] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:10:47] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:10:47] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:10:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:10:48] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:10:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:10:48] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:10:48] [Rank 0] step:2001/10000 train_time:109906ms step_avg:54.93ms
+[2025-09-04 13:10:48] [Rank 0] step:2001/10000 train_time:109906ms step_avg:54.93ms
+[2025-09-04 13:10:49] [Rank 0] step:2021/10000 train_time:110918ms step_avg:54.88ms
+[2025-09-04 13:10:49] [Rank 0] step:2021/10000 train_time:110918ms step_avg:54.88ms
+[2025-09-04 13:10:50] [Rank 0] step:2041/10000 train_time:111662ms step_avg:54.71ms
+[2025-09-04 13:10:50] [Rank 0] step:2041/10000 train_time:111662ms step_avg:54.71ms
+[2025-09-04 13:10:50] [Rank 0] step:2061/10000 train_time:112407ms step_avg:54.54ms
+[2025-09-04 13:10:50] [Rank 0] step:2061/10000 train_time:112407ms step_avg:54.54ms
+[2025-09-04 13:10:51] [Rank 0] step:2081/10000 train_time:113151ms step_avg:54.37ms
+[2025-09-04 13:10:51] [Rank 0] step:2081/10000 train_time:113151ms step_avg:54.37ms
+[2025-09-04 13:10:52] [Rank 0] step:2101/10000 train_time:113894ms step_avg:54.21ms
+[2025-09-04 13:10:52] [Rank 0] step:2101/10000 train_time:113894ms step_avg:54.21ms
+[2025-09-04 13:10:53] [Rank 0] step:2121/10000 train_time:114637ms step_avg:54.05ms
+[2025-09-04 13:10:53] [Rank 0] step:2121/10000 train_time:114637ms step_avg:54.05ms
+[2025-09-04 13:10:53] [Rank 0] step:2141/10000 train_time:115380ms step_avg:53.89ms
+[2025-09-04 13:10:53] [Rank 0] step:2141/10000 train_time:115380ms step_avg:53.89ms
+[2025-09-04 13:10:54] [Rank 0] step:2161/10000 train_time:116124ms step_avg:53.74ms
+[2025-09-04 13:10:54] [Rank 0] step:2161/10000 train_time:116124ms step_avg:53.74ms
+[2025-09-04 13:10:55] [Rank 0] step:2181/10000 train_time:116868ms step_avg:53.58ms
+[2025-09-04 13:10:55] [Rank 0] step:2181/10000 train_time:116868ms step_avg:53.58ms
+[2025-09-04 13:10:56] [Rank 0] step:2201/10000 train_time:117756ms step_avg:53.50ms
+[2025-09-04 13:10:56] [Rank 0] step:2201/10000 train_time:117756ms step_avg:53.50ms
+[2025-09-04 13:10:57] [Rank 0] step:2221/10000 train_time:118574ms step_avg:53.39ms
+[2025-09-04 13:10:57] [Rank 0] step:2221/10000 train_time:118574ms step_avg:53.39ms
+[2025-09-04 13:10:57] [Rank 0] step:2241/10000 train_time:119328ms step_avg:53.25ms
+[2025-09-04 13:10:57] [Rank 0] step:2241/10000 train_time:119328ms step_avg:53.25ms
+[2025-09-04 13:10:58] [Rank 0] step:2261/10000 train_time:120383ms step_avg:53.24ms
+[2025-09-04 13:10:58] [Rank 0] step:2261/10000 train_time:120383ms step_avg:53.24ms
+[2025-09-04 13:10:59] [Rank 0] step:2281/10000 train_time:121138ms step_avg:53.11ms
+[2025-09-04 13:10:59] [Rank 0] step:2281/10000 train_time:121138ms step_avg:53.11ms
+[2025-09-04 13:11:00] [Rank 0] step:2301/10000 train_time:121892ms step_avg:52.97ms
+[2025-09-04 13:11:00] [Rank 0] step:2301/10000 train_time:121892ms step_avg:52.97ms
+[2025-09-04 13:11:01] [Rank 0] step:2321/10000 train_time:122646ms step_avg:52.84ms
+[2025-09-04 13:11:01] [Rank 0] step:2321/10000 train_time:122646ms step_avg:52.84ms
+[2025-09-04 13:11:01] [Rank 0] step:2341/10000 train_time:123400ms step_avg:52.71ms
+[2025-09-04 13:11:01] [Rank 0] step:2341/10000 train_time:123400ms step_avg:52.71ms
+[2025-09-04 13:11:02] [Rank 0] step:2361/10000 train_time:124156ms step_avg:52.59ms
+[2025-09-04 13:11:02] [Rank 0] step:2361/10000 train_time:124156ms step_avg:52.59ms
+[2025-09-04 13:11:03] [Rank 0] step:2381/10000 train_time:124910ms step_avg:52.46ms
+[2025-09-04 13:11:03] [Rank 0] step:2381/10000 train_time:124910ms step_avg:52.46ms
+[2025-09-04 13:11:04] [Rank 0] step:2401/10000 train_time:125665ms step_avg:52.34ms
+[2025-09-04 13:11:04] [Rank 0] step:2401/10000 train_time:125665ms step_avg:52.34ms
+[2025-09-04 13:11:04] [Rank 0] step:2421/10000 train_time:126419ms step_avg:52.22ms
+[2025-09-04 13:11:04] [Rank 0] step:2421/10000 train_time:126419ms step_avg:52.22ms
+[2025-09-04 13:11:05] [Rank 0] step:2441/10000 train_time:127173ms step_avg:52.10ms
+[2025-09-04 13:11:05] [Rank 0] step:2441/10000 train_time:127173ms step_avg:52.10ms
+[2025-09-04 13:11:06] [Rank 0] step:2461/10000 train_time:127928ms step_avg:51.98ms
+[2025-09-04 13:11:06] [Rank 0] step:2461/10000 train_time:127928ms step_avg:51.98ms
+[2025-09-04 13:11:07] [Rank 0] step:2481/10000 train_time:128682ms step_avg:51.87ms
+[2025-09-04 13:11:07] [Rank 0] step:2481/10000 train_time:128682ms step_avg:51.87ms
+[2025-09-04 13:11:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:11:07] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:11:08] [Rank 0] PRINT: step:2500/10000 train_loss:0.7340 val_loss:0.7063 train_time:129442ms step_avg:51.78ms
+[2025-09-04 13:11:08] [Rank 0] PRINT: step:2500/10000 train_loss:0.7340 val_loss:0.7063 train_time:129442ms step_avg:51.78ms
+[2025-09-04 13:11:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:11:08] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:11:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:11:08] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:12:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:12:44] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:12:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:12:44] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:12:44] [Rank 0]   Total Loss: 4.7736
+[2025-09-04 13:12:44] [Rank 0]   Total Loss: 4.7736
+[2025-09-04 13:12:44] [Rank 0]   Total FTA (Unweighted): 0.8163
+[2025-09-04 13:12:44] [Rank 0]   Total FTA (Unweighted): 0.8163
+[2025-09-04 13:12:44] [Rank 0]   Total FTA (Weighted):   0.8163
+[2025-09-04 13:12:44] [Rank 0]   Total FTA (Weighted):   0.8163
+[2025-09-04 13:12:44] [Rank 0]   Group 0 Loss: 4.6483
+[2025-09-04 13:12:44] [Rank 0]   Group 0 Loss: 4.6483
+[2025-09-04 13:12:44] [Rank 0]   Group 1 Loss: 4.3660
+[2025-09-04 13:12:44] [Rank 0]   Group 1 Loss: 4.3660
+[2025-09-04 13:12:44] [Rank 0]   Group 2 Loss: 4.2180
+[2025-09-04 13:12:44] [Rank 0]   Group 2 Loss: 4.2180
+[2025-09-04 13:12:44] [Rank 0]   Group 3 Loss: 4.7445
+[2025-09-04 13:12:44] [Rank 0]   Group 3 Loss: 4.7445
+[2025-09-04 13:12:44] [Rank 0]   Group 4 Loss: 4.6468
+[2025-09-04 13:12:44] [Rank 0]   Group 4 Loss: 4.6468
+[2025-09-04 13:12:44] [Rank 0]   Group 5 Loss: 4.6739
+[2025-09-04 13:12:44] [Rank 0]   Group 5 Loss: 4.6739
+[2025-09-04 13:12:44] [Rank 0]   Group 6 Loss: 4.6052
+[2025-09-04 13:12:44] [Rank 0]   Group 6 Loss: 4.6052
+[2025-09-04 13:12:44] [Rank 0]   Group 7 Loss: 4.6651
+[2025-09-04 13:12:44] [Rank 0]   Group 7 Loss: 4.6651
+[2025-09-04 13:12:44] [Rank 0]   Group 8 Loss: 4.8063
+[2025-09-04 13:12:44] [Rank 0]   Group 8 Loss: 4.8063
+[2025-09-04 13:12:44] [Rank 0]   Group 9 Loss: 4.7376
+[2025-09-04 13:12:44] [Rank 0]   Group 9 Loss: 4.7376
+[2025-09-04 13:12:44] [Rank 0]   Group 10 Loss: 4.9349
+[2025-09-04 13:12:44] [Rank 0]   Group 10 Loss: 4.9349
+[2025-09-04 13:12:44] [Rank 0]   Group 11 Loss: 4.9636
+[2025-09-04 13:12:44] [Rank 0]   Group 11 Loss: 4.9636
+[2025-09-04 13:12:44] [Rank 0]   Group 12 Loss: 4.9264
+[2025-09-04 13:12:44] [Rank 0]   Group 12 Loss: 4.9264
+[2025-09-04 13:12:44] [Rank 0]   Group 13 Loss: 5.0881
+[2025-09-04 13:12:44] [Rank 0]   Group 13 Loss: 5.0881
+[2025-09-04 13:12:44] [Rank 0]   Group 14 Loss: 5.1299
+[2025-09-04 13:12:44] [Rank 0]   Group 14 Loss: 5.1299
+[2025-09-04 13:12:44] [Rank 0]   Group 15 Loss: 5.2235
+[2025-09-04 13:12:44] [Rank 0]   Group 15 Loss: 5.2235
+[2025-09-04 13:12:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:12:44] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 13:12:44] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 13:12:44] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 13:12:44] [Rank 0]   Group 10 FTA: 0.9600
+[2025-09-04 13:12:44] [Rank 0]   Group 11 FTA: 0.8700
+[2025-09-04 13:12:44] [Rank 0]   Group 11 FTA: 0.8700
+[2025-09-04 13:12:44] [Rank 0]   Group 12 FTA: 0.7100
+[2025-09-04 13:12:44] [Rank 0]   Group 12 FTA: 0.7100
+[2025-09-04 13:12:44] [Rank 0]   Group 13 FTA: 0.3300
+[2025-09-04 13:12:44] [Rank 0]   Group 13 FTA: 0.3300
+[2025-09-04 13:12:44] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 13:12:44] [Rank 0]   Group 14 FTA: 0.1600
+[2025-09-04 13:12:44] [Rank 0]   Group 15 FTA: 0.0600
+[2025-09-04 13:12:44] [Rank 0]   Group 15 FTA: 0.0600
+[2025-09-04 13:12:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:12:44] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:12:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:12:45] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:12:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:12:45] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:12:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:12:45] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:12:45] [Rank 0] step:2501/10000 train_time:129457ms step_avg:51.76ms
+[2025-09-04 13:12:45] [Rank 0] step:2501/10000 train_time:129457ms step_avg:51.76ms
+[2025-09-04 13:12:46] [Rank 0] step:2521/10000 train_time:130225ms step_avg:51.66ms
+[2025-09-04 13:12:46] [Rank 0] step:2521/10000 train_time:130225ms step_avg:51.66ms
+[2025-09-04 13:12:47] [Rank 0] step:2541/10000 train_time:130980ms step_avg:51.55ms
+[2025-09-04 13:12:47] [Rank 0] step:2541/10000 train_time:130980ms step_avg:51.55ms
+[2025-09-04 13:12:48] [Rank 0] step:2561/10000 train_time:131734ms step_avg:51.44ms
+[2025-09-04 13:12:48] [Rank 0] step:2561/10000 train_time:131734ms step_avg:51.44ms
+[2025-09-04 13:12:48] [Rank 0] step:2581/10000 train_time:132488ms step_avg:51.33ms
+[2025-09-04 13:12:48] [Rank 0] step:2581/10000 train_time:132488ms step_avg:51.33ms
+[2025-09-04 13:12:49] [Rank 0] step:2601/10000 train_time:133243ms step_avg:51.23ms
+[2025-09-04 13:12:49] [Rank 0] step:2601/10000 train_time:133243ms step_avg:51.23ms
+[2025-09-04 13:12:50] [Rank 0] step:2621/10000 train_time:133997ms step_avg:51.12ms
+[2025-09-04 13:12:50] [Rank 0] step:2621/10000 train_time:133997ms step_avg:51.12ms
+[2025-09-04 13:12:51] [Rank 0] step:2641/10000 train_time:134751ms step_avg:51.02ms
+[2025-09-04 13:12:51] [Rank 0] step:2641/10000 train_time:134751ms step_avg:51.02ms
+[2025-09-04 13:12:51] [Rank 0] step:2661/10000 train_time:135505ms step_avg:50.92ms
+[2025-09-04 13:12:51] [Rank 0] step:2661/10000 train_time:135505ms step_avg:50.92ms
+[2025-09-04 13:12:52] [Rank 0] step:2681/10000 train_time:136260ms step_avg:50.82ms
+[2025-09-04 13:12:52] [Rank 0] step:2681/10000 train_time:136260ms step_avg:50.82ms
+[2025-09-04 13:12:53] [Rank 0] step:2701/10000 train_time:137014ms step_avg:50.73ms
+[2025-09-04 13:12:53] [Rank 0] step:2701/10000 train_time:137014ms step_avg:50.73ms
+[2025-09-04 13:12:54] [Rank 0] step:2721/10000 train_time:137768ms step_avg:50.63ms
+[2025-09-04 13:12:54] [Rank 0] step:2721/10000 train_time:137768ms step_avg:50.63ms
+[2025-09-04 13:12:54] [Rank 0] step:2741/10000 train_time:138523ms step_avg:50.54ms
+[2025-09-04 13:12:54] [Rank 0] step:2741/10000 train_time:138523ms step_avg:50.54ms
+[2025-09-04 13:12:55] [Rank 0] step:2761/10000 train_time:139277ms step_avg:50.44ms
+[2025-09-04 13:12:55] [Rank 0] step:2761/10000 train_time:139277ms step_avg:50.44ms
+[2025-09-04 13:12:56] [Rank 0] step:2781/10000 train_time:140031ms step_avg:50.35ms
+[2025-09-04 13:12:56] [Rank 0] step:2781/10000 train_time:140031ms step_avg:50.35ms
+[2025-09-04 13:12:57] [Rank 0] step:2801/10000 train_time:140786ms step_avg:50.26ms
+[2025-09-04 13:12:57] [Rank 0] step:2801/10000 train_time:140786ms step_avg:50.26ms
+[2025-09-04 13:12:58] [Rank 0] step:2821/10000 train_time:141813ms step_avg:50.27ms
+[2025-09-04 13:12:58] [Rank 0] step:2821/10000 train_time:141813ms step_avg:50.27ms
+[2025-09-04 13:12:58] [Rank 0] step:2841/10000 train_time:142568ms step_avg:50.18ms
+[2025-09-04 13:12:58] [Rank 0] step:2841/10000 train_time:142568ms step_avg:50.18ms
+[2025-09-04 13:12:59] [Rank 0] step:2861/10000 train_time:143323ms step_avg:50.10ms
+[2025-09-04 13:12:59] [Rank 0] step:2861/10000 train_time:143323ms step_avg:50.10ms
+[2025-09-04 13:13:00] [Rank 0] step:2881/10000 train_time:144078ms step_avg:50.01ms
+[2025-09-04 13:13:00] [Rank 0] step:2881/10000 train_time:144078ms step_avg:50.01ms
+[2025-09-04 13:13:01] [Rank 0] step:2901/10000 train_time:144832ms step_avg:49.92ms
+[2025-09-04 13:13:01] [Rank 0] step:2901/10000 train_time:144832ms step_avg:49.92ms
+[2025-09-04 13:13:01] [Rank 0] step:2921/10000 train_time:145587ms step_avg:49.84ms
+[2025-09-04 13:13:01] [Rank 0] step:2921/10000 train_time:145587ms step_avg:49.84ms
+[2025-09-04 13:13:02] [Rank 0] step:2941/10000 train_time:146342ms step_avg:49.76ms
+[2025-09-04 13:13:02] [Rank 0] step:2941/10000 train_time:146342ms step_avg:49.76ms
+[2025-09-04 13:13:03] [Rank 0] step:2961/10000 train_time:147378ms step_avg:49.77ms
+[2025-09-04 13:13:03] [Rank 0] step:2961/10000 train_time:147378ms step_avg:49.77ms
+[2025-09-04 13:13:04] [Rank 0] step:2981/10000 train_time:148135ms step_avg:49.69ms
+[2025-09-04 13:13:04] [Rank 0] step:2981/10000 train_time:148135ms step_avg:49.69ms
+[2025-09-04 13:13:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:13:05] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:13:05] [Rank 0] PRINT: step:3000/10000 train_loss:0.7079 val_loss:0.6871 train_time:148895ms step_avg:49.63ms
+[2025-09-04 13:13:05] [Rank 0] PRINT: step:3000/10000 train_loss:0.7079 val_loss:0.6871 train_time:148895ms step_avg:49.63ms
+[2025-09-04 13:13:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:13:05] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:13:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:13:05] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:14:41] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:14:41] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:14:41] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:14:41] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:14:41] [Rank 0]   Total Loss: 4.8105
+[2025-09-04 13:14:41] [Rank 0]   Total Loss: 4.8105
+[2025-09-04 13:14:41] [Rank 0]   Total FTA (Unweighted): 0.8438
+[2025-09-04 13:14:41] [Rank 0]   Total FTA (Unweighted): 0.8438
+[2025-09-04 13:14:41] [Rank 0]   Total FTA (Weighted):   0.8438
+[2025-09-04 13:14:41] [Rank 0]   Total FTA (Weighted):   0.8438
+[2025-09-04 13:14:41] [Rank 0]   Group 0 Loss: 4.7618
+[2025-09-04 13:14:41] [Rank 0]   Group 0 Loss: 4.7618
+[2025-09-04 13:14:41] [Rank 0]   Group 1 Loss: 4.4495
+[2025-09-04 13:14:41] [Rank 0]   Group 1 Loss: 4.4495
+[2025-09-04 13:14:41] [Rank 0]   Group 2 Loss: 4.2602
+[2025-09-04 13:14:41] [Rank 0]   Group 2 Loss: 4.2602
+[2025-09-04 13:14:41] [Rank 0]   Group 3 Loss: 4.7355
+[2025-09-04 13:14:41] [Rank 0]   Group 3 Loss: 4.7355
+[2025-09-04 13:14:41] [Rank 0]   Group 4 Loss: 4.7128
+[2025-09-04 13:14:41] [Rank 0]   Group 4 Loss: 4.7128
+[2025-09-04 13:14:41] [Rank 0]   Group 5 Loss: 4.7181
+[2025-09-04 13:14:41] [Rank 0]   Group 5 Loss: 4.7181
+[2025-09-04 13:14:41] [Rank 0]   Group 6 Loss: 4.6347
+[2025-09-04 13:14:41] [Rank 0]   Group 6 Loss: 4.6347
+[2025-09-04 13:14:41] [Rank 0]   Group 7 Loss: 4.7090
+[2025-09-04 13:14:41] [Rank 0]   Group 7 Loss: 4.7090
+[2025-09-04 13:14:41] [Rank 0]   Group 8 Loss: 4.8998
+[2025-09-04 13:14:41] [Rank 0]   Group 8 Loss: 4.8998
+[2025-09-04 13:14:41] [Rank 0]   Group 9 Loss: 4.8205
+[2025-09-04 13:14:41] [Rank 0]   Group 9 Loss: 4.8205
+[2025-09-04 13:14:41] [Rank 0]   Group 10 Loss: 4.9472
+[2025-09-04 13:14:41] [Rank 0]   Group 10 Loss: 4.9472
+[2025-09-04 13:14:41] [Rank 0]   Group 11 Loss: 5.0259
+[2025-09-04 13:14:41] [Rank 0]   Group 11 Loss: 5.0259
+[2025-09-04 13:14:41] [Rank 0]   Group 12 Loss: 4.9572
+[2025-09-04 13:14:41] [Rank 0]   Group 12 Loss: 4.9572
+[2025-09-04 13:14:41] [Rank 0]   Group 13 Loss: 5.0801
+[2025-09-04 13:14:41] [Rank 0]   Group 13 Loss: 5.0801
+[2025-09-04 13:14:41] [Rank 0]   Group 14 Loss: 5.0966
+[2025-09-04 13:14:41] [Rank 0]   Group 14 Loss: 5.0966
+[2025-09-04 13:14:41] [Rank 0]   Group 15 Loss: 5.1589
+[2025-09-04 13:14:41] [Rank 0]   Group 15 Loss: 5.1589
+[2025-09-04 13:14:41] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:14:41] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 13:14:41] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 13:14:41] [Rank 0]   Group 11 FTA: 0.9700
+[2025-09-04 13:14:41] [Rank 0]   Group 11 FTA: 0.9700
+[2025-09-04 13:14:41] [Rank 0]   Group 12 FTA: 0.8500
+[2025-09-04 13:14:41] [Rank 0]   Group 12 FTA: 0.8500
+[2025-09-04 13:14:41] [Rank 0]   Group 13 FTA: 0.4100
+[2025-09-04 13:14:41] [Rank 0]   Group 13 FTA: 0.4100
+[2025-09-04 13:14:41] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 13:14:41] [Rank 0]   Group 14 FTA: 0.1500
+[2025-09-04 13:14:41] [Rank 0]   Group 15 FTA: 0.1400
+[2025-09-04 13:14:41] [Rank 0]   Group 15 FTA: 0.1400
+[2025-09-04 13:14:42] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:14:42] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:14:42] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:14:42] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:14:42] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:14:42] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:14:43] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:14:43] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:14:43] [Rank 0] step:3001/10000 train_time:148909ms step_avg:49.62ms
+[2025-09-04 13:14:43] [Rank 0] step:3001/10000 train_time:148909ms step_avg:49.62ms
+[2025-09-04 13:14:43] [Rank 0] step:3021/10000 train_time:149681ms step_avg:49.55ms
+[2025-09-04 13:14:43] [Rank 0] step:3021/10000 train_time:149681ms step_avg:49.55ms
+[2025-09-04 13:14:44] [Rank 0] step:3041/10000 train_time:150435ms step_avg:49.47ms
+[2025-09-04 13:14:44] [Rank 0] step:3041/10000 train_time:150435ms step_avg:49.47ms
+[2025-09-04 13:14:45] [Rank 0] step:3061/10000 train_time:151190ms step_avg:49.39ms
+[2025-09-04 13:14:45] [Rank 0] step:3061/10000 train_time:151190ms step_avg:49.39ms
+[2025-09-04 13:14:46] [Rank 0] step:3081/10000 train_time:151944ms step_avg:49.32ms
+[2025-09-04 13:14:46] [Rank 0] step:3081/10000 train_time:151944ms step_avg:49.32ms
+[2025-09-04 13:14:46] [Rank 0] step:3101/10000 train_time:152699ms step_avg:49.24ms
+[2025-09-04 13:14:46] [Rank 0] step:3101/10000 train_time:152699ms step_avg:49.24ms
+[2025-09-04 13:14:47] [Rank 0] step:3121/10000 train_time:153452ms step_avg:49.17ms
+[2025-09-04 13:14:47] [Rank 0] step:3121/10000 train_time:153452ms step_avg:49.17ms
+[2025-09-04 13:14:48] [Rank 0] step:3141/10000 train_time:154207ms step_avg:49.09ms
+[2025-09-04 13:14:48] [Rank 0] step:3141/10000 train_time:154207ms step_avg:49.09ms
+[2025-09-04 13:14:49] [Rank 0] step:3161/10000 train_time:154961ms step_avg:49.02ms
+[2025-09-04 13:14:49] [Rank 0] step:3161/10000 train_time:154961ms step_avg:49.02ms
+[2025-09-04 13:14:49] [Rank 0] step:3181/10000 train_time:155715ms step_avg:48.95ms
+[2025-09-04 13:14:49] [Rank 0] step:3181/10000 train_time:155715ms step_avg:48.95ms
+[2025-09-04 13:14:50] [Rank 0] step:3201/10000 train_time:156468ms step_avg:48.88ms
+[2025-09-04 13:14:50] [Rank 0] step:3201/10000 train_time:156468ms step_avg:48.88ms
+[2025-09-04 13:14:51] [Rank 0] step:3221/10000 train_time:157222ms step_avg:48.81ms
+[2025-09-04 13:14:51] [Rank 0] step:3221/10000 train_time:157222ms step_avg:48.81ms
+[2025-09-04 13:14:52] [Rank 0] step:3241/10000 train_time:157976ms step_avg:48.74ms
+[2025-09-04 13:14:52] [Rank 0] step:3241/10000 train_time:157976ms step_avg:48.74ms
+[2025-09-04 13:14:52] [Rank 0] step:3261/10000 train_time:158730ms step_avg:48.68ms
+[2025-09-04 13:14:52] [Rank 0] step:3261/10000 train_time:158730ms step_avg:48.68ms
+[2025-09-04 13:14:53] [Rank 0] step:3281/10000 train_time:159484ms step_avg:48.61ms
+[2025-09-04 13:14:53] [Rank 0] step:3281/10000 train_time:159484ms step_avg:48.61ms
+[2025-09-04 13:14:54] [Rank 0] step:3301/10000 train_time:160239ms step_avg:48.54ms
+[2025-09-04 13:14:54] [Rank 0] step:3301/10000 train_time:160239ms step_avg:48.54ms
+[2025-09-04 13:14:55] [Rank 0] step:3321/10000 train_time:160993ms step_avg:48.48ms
+[2025-09-04 13:14:55] [Rank 0] step:3321/10000 train_time:160993ms step_avg:48.48ms
+[2025-09-04 13:14:55] [Rank 0] step:3341/10000 train_time:161748ms step_avg:48.41ms
+[2025-09-04 13:14:55] [Rank 0] step:3341/10000 train_time:161748ms step_avg:48.41ms
+[2025-09-04 13:14:56] [Rank 0] step:3361/10000 train_time:162506ms step_avg:48.35ms
+[2025-09-04 13:14:56] [Rank 0] step:3361/10000 train_time:162506ms step_avg:48.35ms
+[2025-09-04 13:14:57] [Rank 0] step:3381/10000 train_time:163267ms step_avg:48.29ms
+[2025-09-04 13:14:57] [Rank 0] step:3381/10000 train_time:163267ms step_avg:48.29ms
+[2025-09-04 13:14:58] [Rank 0] step:3401/10000 train_time:164023ms step_avg:48.23ms
+[2025-09-04 13:14:58] [Rank 0] step:3401/10000 train_time:164023ms step_avg:48.23ms
+[2025-09-04 13:14:59] [Rank 0] step:3421/10000 train_time:164777ms step_avg:48.17ms
+[2025-09-04 13:14:59] [Rank 0] step:3421/10000 train_time:164777ms step_avg:48.17ms
+[2025-09-04 13:14:59] [Rank 0] step:3441/10000 train_time:165532ms step_avg:48.11ms
+[2025-09-04 13:14:59] [Rank 0] step:3441/10000 train_time:165532ms step_avg:48.11ms
+[2025-09-04 13:15:00] [Rank 0] step:3461/10000 train_time:166288ms step_avg:48.05ms
+[2025-09-04 13:15:00] [Rank 0] step:3461/10000 train_time:166288ms step_avg:48.05ms
+[2025-09-04 13:15:01] [Rank 0] step:3481/10000 train_time:167043ms step_avg:47.99ms
+[2025-09-04 13:15:01] [Rank 0] step:3481/10000 train_time:167043ms step_avg:47.99ms
+[2025-09-04 13:15:02] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:15:02] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:15:02] [Rank 0] PRINT: step:3500/10000 train_loss:0.6914 val_loss:0.6736 train_time:167837ms step_avg:47.95ms
+[2025-09-04 13:15:02] [Rank 0] PRINT: step:3500/10000 train_loss:0.6914 val_loss:0.6736 train_time:167837ms step_avg:47.95ms
+[2025-09-04 13:15:02] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:15:02] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:15:02] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:15:02] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:16:38] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:16:38] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:16:38] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:16:38] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:16:38] [Rank 0]   Total Loss: 4.8170
+[2025-09-04 13:16:38] [Rank 0]   Total Loss: 4.8170
+[2025-09-04 13:16:38] [Rank 0]   Total FTA (Unweighted): 0.8619
+[2025-09-04 13:16:38] [Rank 0]   Total FTA (Unweighted): 0.8619
+[2025-09-04 13:16:38] [Rank 0]   Total FTA (Weighted):   0.8619
+[2025-09-04 13:16:38] [Rank 0]   Total FTA (Weighted):   0.8619
+[2025-09-04 13:16:38] [Rank 0]   Group 0 Loss: 4.7278
+[2025-09-04 13:16:38] [Rank 0]   Group 0 Loss: 4.7278
+[2025-09-04 13:16:38] [Rank 0]   Group 1 Loss: 4.4221
+[2025-09-04 13:16:38] [Rank 0]   Group 1 Loss: 4.4221
+[2025-09-04 13:16:38] [Rank 0]   Group 2 Loss: 4.2938
+[2025-09-04 13:16:38] [Rank 0]   Group 2 Loss: 4.2938
+[2025-09-04 13:16:38] [Rank 0]   Group 3 Loss: 4.7405
+[2025-09-04 13:16:38] [Rank 0]   Group 3 Loss: 4.7405
+[2025-09-04 13:16:38] [Rank 0]   Group 4 Loss: 4.7193
+[2025-09-04 13:16:38] [Rank 0]   Group 4 Loss: 4.7193
+[2025-09-04 13:16:38] [Rank 0]   Group 5 Loss: 4.7409
+[2025-09-04 13:16:38] [Rank 0]   Group 5 Loss: 4.7409
+[2025-09-04 13:16:38] [Rank 0]   Group 6 Loss: 4.6282
+[2025-09-04 13:16:38] [Rank 0]   Group 6 Loss: 4.6282
+[2025-09-04 13:16:38] [Rank 0]   Group 7 Loss: 4.7223
+[2025-09-04 13:16:38] [Rank 0]   Group 7 Loss: 4.7223
+[2025-09-04 13:16:38] [Rank 0]   Group 8 Loss: 4.9004
+[2025-09-04 13:16:38] [Rank 0]   Group 8 Loss: 4.9004
+[2025-09-04 13:16:38] [Rank 0]   Group 9 Loss: 4.8293
+[2025-09-04 13:16:38] [Rank 0]   Group 9 Loss: 4.8293
+[2025-09-04 13:16:38] [Rank 0]   Group 10 Loss: 4.9943
+[2025-09-04 13:16:38] [Rank 0]   Group 10 Loss: 4.9943
+[2025-09-04 13:16:38] [Rank 0]   Group 11 Loss: 5.0632
+[2025-09-04 13:16:38] [Rank 0]   Group 11 Loss: 5.0632
+[2025-09-04 13:16:38] [Rank 0]   Group 12 Loss: 4.9531
+[2025-09-04 13:16:38] [Rank 0]   Group 12 Loss: 4.9531
+[2025-09-04 13:16:38] [Rank 0]   Group 13 Loss: 5.0649
+[2025-09-04 13:16:38] [Rank 0]   Group 13 Loss: 5.0649
+[2025-09-04 13:16:38] [Rank 0]   Group 14 Loss: 5.1321
+[2025-09-04 13:16:38] [Rank 0]   Group 14 Loss: 5.1321
+[2025-09-04 13:16:38] [Rank 0]   Group 15 Loss: 5.1405
+[2025-09-04 13:16:38] [Rank 0]   Group 15 Loss: 5.1405
+[2025-09-04 13:16:38] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:16:38] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 13:16:38] [Rank 0]   Group 9 FTA: 0.9700
+[2025-09-04 13:16:38] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 13:16:38] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 13:16:38] [Rank 0]   Group 11 FTA: 0.9400
+[2025-09-04 13:16:38] [Rank 0]   Group 11 FTA: 0.9400
+[2025-09-04 13:16:38] [Rank 0]   Group 12 FTA: 0.9600
+[2025-09-04 13:16:38] [Rank 0]   Group 12 FTA: 0.9600
+[2025-09-04 13:16:38] [Rank 0]   Group 13 FTA: 0.6100
+[2025-09-04 13:16:38] [Rank 0]   Group 13 FTA: 0.6100
+[2025-09-04 13:16:38] [Rank 0]   Group 14 FTA: 0.2300
+[2025-09-04 13:16:38] [Rank 0]   Group 14 FTA: 0.2300
+[2025-09-04 13:16:38] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 13:16:38] [Rank 0]   Group 15 FTA: 0.1000
+[2025-09-04 13:16:38] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:16:38] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:16:38] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:16:38] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:16:39] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:16:39] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:16:39] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:16:39] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:16:39] [Rank 0] step:3501/10000 train_time:167853ms step_avg:47.94ms
+[2025-09-04 13:16:39] [Rank 0] step:3501/10000 train_time:167853ms step_avg:47.94ms
+[2025-09-04 13:16:40] [Rank 0] step:3521/10000 train_time:168619ms step_avg:47.89ms
+[2025-09-04 13:16:40] [Rank 0] step:3521/10000 train_time:168619ms step_avg:47.89ms
+[2025-09-04 13:16:41] [Rank 0] step:3541/10000 train_time:169477ms step_avg:47.86ms
+[2025-09-04 13:16:41] [Rank 0] step:3541/10000 train_time:169477ms step_avg:47.86ms
+[2025-09-04 13:16:41] [Rank 0] step:3561/10000 train_time:170231ms step_avg:47.80ms
+[2025-09-04 13:16:41] [Rank 0] step:3561/10000 train_time:170231ms step_avg:47.80ms
+[2025-09-04 13:16:42] [Rank 0] step:3581/10000 train_time:170986ms step_avg:47.75ms
+[2025-09-04 13:16:42] [Rank 0] step:3581/10000 train_time:170986ms step_avg:47.75ms
+[2025-09-04 13:16:43] [Rank 0] step:3601/10000 train_time:171742ms step_avg:47.69ms
+[2025-09-04 13:16:43] [Rank 0] step:3601/10000 train_time:171742ms step_avg:47.69ms
+[2025-09-04 13:16:44] [Rank 0] step:3621/10000 train_time:172495ms step_avg:47.64ms
+[2025-09-04 13:16:44] [Rank 0] step:3621/10000 train_time:172495ms step_avg:47.64ms
+[2025-09-04 13:16:45] [Rank 0] step:3641/10000 train_time:173522ms step_avg:47.66ms
+[2025-09-04 13:16:45] [Rank 0] step:3641/10000 train_time:173522ms step_avg:47.66ms
+[2025-09-04 13:16:45] [Rank 0] step:3661/10000 train_time:174276ms step_avg:47.60ms
+[2025-09-04 13:16:45] [Rank 0] step:3661/10000 train_time:174276ms step_avg:47.60ms
+[2025-09-04 13:16:46] [Rank 0] step:3681/10000 train_time:175030ms step_avg:47.55ms
+[2025-09-04 13:16:46] [Rank 0] step:3681/10000 train_time:175030ms step_avg:47.55ms
+[2025-09-04 13:16:47] [Rank 0] step:3701/10000 train_time:175784ms step_avg:47.50ms
+[2025-09-04 13:16:47] [Rank 0] step:3701/10000 train_time:175784ms step_avg:47.50ms
+[2025-09-04 13:16:48] [Rank 0] step:3721/10000 train_time:176538ms step_avg:47.44ms
+[2025-09-04 13:16:48] [Rank 0] step:3721/10000 train_time:176538ms step_avg:47.44ms
+[2025-09-04 13:16:49] [Rank 0] step:3741/10000 train_time:177293ms step_avg:47.39ms
+[2025-09-04 13:16:49] [Rank 0] step:3741/10000 train_time:177293ms step_avg:47.39ms
+[2025-09-04 13:16:49] [Rank 0] step:3761/10000 train_time:178047ms step_avg:47.34ms
+[2025-09-04 13:16:49] [Rank 0] step:3761/10000 train_time:178047ms step_avg:47.34ms
+[2025-09-04 13:16:50] [Rank 0] step:3781/10000 train_time:178802ms step_avg:47.29ms
+[2025-09-04 13:16:50] [Rank 0] step:3781/10000 train_time:178802ms step_avg:47.29ms
+[2025-09-04 13:16:51] [Rank 0] step:3801/10000 train_time:179556ms step_avg:47.24ms
+[2025-09-04 13:16:51] [Rank 0] step:3801/10000 train_time:179556ms step_avg:47.24ms
+[2025-09-04 13:16:52] [Rank 0] step:3821/10000 train_time:180310ms step_avg:47.19ms
+[2025-09-04 13:16:52] [Rank 0] step:3821/10000 train_time:180310ms step_avg:47.19ms
+[2025-09-04 13:16:52] [Rank 0] step:3841/10000 train_time:181064ms step_avg:47.14ms
+[2025-09-04 13:16:52] [Rank 0] step:3841/10000 train_time:181064ms step_avg:47.14ms
+[2025-09-04 13:16:53] [Rank 0] step:3861/10000 train_time:181818ms step_avg:47.09ms
+[2025-09-04 13:16:53] [Rank 0] step:3861/10000 train_time:181818ms step_avg:47.09ms
+[2025-09-04 13:16:54] [Rank 0] step:3881/10000 train_time:182573ms step_avg:47.04ms
+[2025-09-04 13:16:54] [Rank 0] step:3881/10000 train_time:182573ms step_avg:47.04ms
+[2025-09-04 13:16:55] [Rank 0] step:3901/10000 train_time:183327ms step_avg:46.99ms
+[2025-09-04 13:16:55] [Rank 0] step:3901/10000 train_time:183327ms step_avg:46.99ms
+[2025-09-04 13:16:55] [Rank 0] step:3921/10000 train_time:184081ms step_avg:46.95ms
+[2025-09-04 13:16:55] [Rank 0] step:3921/10000 train_time:184081ms step_avg:46.95ms
+[2025-09-04 13:16:56] [Rank 0] step:3941/10000 train_time:184837ms step_avg:46.90ms
+[2025-09-04 13:16:56] [Rank 0] step:3941/10000 train_time:184837ms step_avg:46.90ms
+[2025-09-04 13:16:57] [Rank 0] step:3961/10000 train_time:185590ms step_avg:46.85ms
+[2025-09-04 13:16:57] [Rank 0] step:3961/10000 train_time:185590ms step_avg:46.85ms
+[2025-09-04 13:16:58] [Rank 0] step:3981/10000 train_time:186345ms step_avg:46.81ms
+[2025-09-04 13:16:58] [Rank 0] step:3981/10000 train_time:186345ms step_avg:46.81ms
+[2025-09-04 13:16:58] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:16:58] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:16:59] [Rank 0] PRINT: step:4000/10000 train_loss:0.6790 val_loss:0.6616 train_time:187105ms step_avg:46.78ms
+[2025-09-04 13:16:59] [Rank 0] PRINT: step:4000/10000 train_loss:0.6790 val_loss:0.6616 train_time:187105ms step_avg:46.78ms
+[2025-09-04 13:16:59] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:16:59] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:16:59] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:16:59] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:18:35] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:18:35] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:18:35] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:18:35] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:18:35] [Rank 0]   Total Loss: 4.8728
+[2025-09-04 13:18:35] [Rank 0]   Total Loss: 4.8728
+[2025-09-04 13:18:35] [Rank 0]   Total FTA (Unweighted): 0.8881
+[2025-09-04 13:18:35] [Rank 0]   Total FTA (Unweighted): 0.8881
+[2025-09-04 13:18:35] [Rank 0]   Total FTA (Weighted):   0.8881
+[2025-09-04 13:18:35] [Rank 0]   Total FTA (Weighted):   0.8881
+[2025-09-04 13:18:35] [Rank 0]   Group 0 Loss: 4.7628
+[2025-09-04 13:18:35] [Rank 0]   Group 0 Loss: 4.7628
+[2025-09-04 13:18:35] [Rank 0]   Group 1 Loss: 4.4508
+[2025-09-04 13:18:35] [Rank 0]   Group 1 Loss: 4.4508
+[2025-09-04 13:18:35] [Rank 0]   Group 2 Loss: 4.3769
+[2025-09-04 13:18:35] [Rank 0]   Group 2 Loss: 4.3769
+[2025-09-04 13:18:35] [Rank 0]   Group 3 Loss: 4.8168
+[2025-09-04 13:18:35] [Rank 0]   Group 3 Loss: 4.8168
+[2025-09-04 13:18:35] [Rank 0]   Group 4 Loss: 4.7183
+[2025-09-04 13:18:35] [Rank 0]   Group 4 Loss: 4.7183
+[2025-09-04 13:18:35] [Rank 0]   Group 5 Loss: 4.8074
+[2025-09-04 13:18:35] [Rank 0]   Group 5 Loss: 4.8074
+[2025-09-04 13:18:35] [Rank 0]   Group 6 Loss: 4.7328
+[2025-09-04 13:18:35] [Rank 0]   Group 6 Loss: 4.7328
+[2025-09-04 13:18:35] [Rank 0]   Group 7 Loss: 4.8059
+[2025-09-04 13:18:35] [Rank 0]   Group 7 Loss: 4.8059
+[2025-09-04 13:18:35] [Rank 0]   Group 8 Loss: 4.9941
+[2025-09-04 13:18:35] [Rank 0]   Group 8 Loss: 4.9941
+[2025-09-04 13:18:35] [Rank 0]   Group 9 Loss: 4.8916
+[2025-09-04 13:18:35] [Rank 0]   Group 9 Loss: 4.8916
+[2025-09-04 13:18:35] [Rank 0]   Group 10 Loss: 5.0798
+[2025-09-04 13:18:35] [Rank 0]   Group 10 Loss: 5.0798
+[2025-09-04 13:18:35] [Rank 0]   Group 11 Loss: 5.1072
+[2025-09-04 13:18:35] [Rank 0]   Group 11 Loss: 5.1072
+[2025-09-04 13:18:35] [Rank 0]   Group 12 Loss: 4.9871
+[2025-09-04 13:18:35] [Rank 0]   Group 12 Loss: 4.9871
+[2025-09-04 13:18:35] [Rank 0]   Group 13 Loss: 5.1086
+[2025-09-04 13:18:35] [Rank 0]   Group 13 Loss: 5.1086
+[2025-09-04 13:18:35] [Rank 0]   Group 14 Loss: 5.1421
+[2025-09-04 13:18:35] [Rank 0]   Group 14 Loss: 5.1421
+[2025-09-04 13:18:35] [Rank 0]   Group 15 Loss: 5.1824
+[2025-09-04 13:18:35] [Rank 0]   Group 15 Loss: 5.1824
+[2025-09-04 13:18:35] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:18:35] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 13:18:35] [Rank 0]   Group 10 FTA: 0.9800
+[2025-09-04 13:18:35] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 13:18:35] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 13:18:35] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 13:18:35] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 13:18:35] [Rank 0]   Group 13 FTA: 0.7500
+[2025-09-04 13:18:35] [Rank 0]   Group 13 FTA: 0.7500
+[2025-09-04 13:18:35] [Rank 0]   Group 14 FTA: 0.3400
+[2025-09-04 13:18:35] [Rank 0]   Group 14 FTA: 0.3400
+[2025-09-04 13:18:35] [Rank 0]   Group 15 FTA: 0.1800
+[2025-09-04 13:18:35] [Rank 0]   Group 15 FTA: 0.1800
+[2025-09-04 13:18:35] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:18:35] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:18:36] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:18:36] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:18:36] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:18:36] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:18:36] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:18:36] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:18:36] [Rank 0] step:4001/10000 train_time:187120ms step_avg:46.77ms
+[2025-09-04 13:18:36] [Rank 0] step:4001/10000 train_time:187120ms step_avg:46.77ms
+[2025-09-04 13:18:37] [Rank 0] step:4021/10000 train_time:188163ms step_avg:46.79ms
+[2025-09-04 13:18:37] [Rank 0] step:4021/10000 train_time:188163ms step_avg:46.79ms
+[2025-09-04 13:18:38] [Rank 0] step:4041/10000 train_time:188917ms step_avg:46.75ms
+[2025-09-04 13:18:38] [Rank 0] step:4041/10000 train_time:188917ms step_avg:46.75ms
+[2025-09-04 13:18:39] [Rank 0] step:4061/10000 train_time:189671ms step_avg:46.71ms
+[2025-09-04 13:18:39] [Rank 0] step:4061/10000 train_time:189671ms step_avg:46.71ms
+[2025-09-04 13:18:40] [Rank 0] step:4081/10000 train_time:190425ms step_avg:46.66ms
+[2025-09-04 13:18:40] [Rank 0] step:4081/10000 train_time:190425ms step_avg:46.66ms
+[2025-09-04 13:18:40] [Rank 0] step:4101/10000 train_time:191178ms step_avg:46.62ms
+[2025-09-04 13:18:40] [Rank 0] step:4101/10000 train_time:191178ms step_avg:46.62ms
+[2025-09-04 13:18:41] [Rank 0] step:4121/10000 train_time:191932ms step_avg:46.57ms
+[2025-09-04 13:18:41] [Rank 0] step:4121/10000 train_time:191932ms step_avg:46.57ms
+[2025-09-04 13:18:42] [Rank 0] step:4141/10000 train_time:192686ms step_avg:46.53ms
+[2025-09-04 13:18:42] [Rank 0] step:4141/10000 train_time:192686ms step_avg:46.53ms
+[2025-09-04 13:18:43] [Rank 0] step:4161/10000 train_time:193440ms step_avg:46.49ms
+[2025-09-04 13:18:43] [Rank 0] step:4161/10000 train_time:193440ms step_avg:46.49ms
+[2025-09-04 13:18:43] [Rank 0] step:4181/10000 train_time:194193ms step_avg:46.45ms
+[2025-09-04 13:18:43] [Rank 0] step:4181/10000 train_time:194193ms step_avg:46.45ms
+[2025-09-04 13:18:44] [Rank 0] step:4201/10000 train_time:194948ms step_avg:46.41ms
+[2025-09-04 13:18:44] [Rank 0] step:4201/10000 train_time:194948ms step_avg:46.41ms
+[2025-09-04 13:18:45] [Rank 0] step:4221/10000 train_time:195701ms step_avg:46.36ms
+[2025-09-04 13:18:45] [Rank 0] step:4221/10000 train_time:195701ms step_avg:46.36ms
+[2025-09-04 13:18:46] [Rank 0] step:4241/10000 train_time:196455ms step_avg:46.32ms
+[2025-09-04 13:18:46] [Rank 0] step:4241/10000 train_time:196455ms step_avg:46.32ms
+[2025-09-04 13:18:46] [Rank 0] step:4261/10000 train_time:197209ms step_avg:46.28ms
+[2025-09-04 13:18:46] [Rank 0] step:4261/10000 train_time:197209ms step_avg:46.28ms
+[2025-09-04 13:18:47] [Rank 0] step:4281/10000 train_time:197963ms step_avg:46.24ms
+[2025-09-04 13:18:47] [Rank 0] step:4281/10000 train_time:197963ms step_avg:46.24ms
+[2025-09-04 13:18:48] [Rank 0] step:4301/10000 train_time:198717ms step_avg:46.20ms
+[2025-09-04 13:18:48] [Rank 0] step:4301/10000 train_time:198717ms step_avg:46.20ms
+[2025-09-04 13:18:49] [Rank 0] step:4321/10000 train_time:199471ms step_avg:46.16ms
+[2025-09-04 13:18:49] [Rank 0] step:4321/10000 train_time:199471ms step_avg:46.16ms
+[2025-09-04 13:18:49] [Rank 0] step:4341/10000 train_time:200224ms step_avg:46.12ms
+[2025-09-04 13:18:49] [Rank 0] step:4341/10000 train_time:200224ms step_avg:46.12ms
+[2025-09-04 13:18:50] [Rank 0] step:4361/10000 train_time:200979ms step_avg:46.09ms
+[2025-09-04 13:18:50] [Rank 0] step:4361/10000 train_time:200979ms step_avg:46.09ms
+[2025-09-04 13:18:51] [Rank 0] step:4381/10000 train_time:201733ms step_avg:46.05ms
+[2025-09-04 13:18:51] [Rank 0] step:4381/10000 train_time:201733ms step_avg:46.05ms
+[2025-09-04 13:18:52] [Rank 0] step:4401/10000 train_time:202487ms step_avg:46.01ms
+[2025-09-04 13:18:52] [Rank 0] step:4401/10000 train_time:202487ms step_avg:46.01ms
+[2025-09-04 13:18:52] [Rank 0] step:4421/10000 train_time:203241ms step_avg:45.97ms
+[2025-09-04 13:18:52] [Rank 0] step:4421/10000 train_time:203241ms step_avg:45.97ms
+[2025-09-04 13:18:53] [Rank 0] step:4441/10000 train_time:203995ms step_avg:45.93ms
+[2025-09-04 13:18:53] [Rank 0] step:4441/10000 train_time:203995ms step_avg:45.93ms
+[2025-09-04 13:18:54] [Rank 0] step:4461/10000 train_time:204750ms step_avg:45.90ms
+[2025-09-04 13:18:54] [Rank 0] step:4461/10000 train_time:204750ms step_avg:45.90ms
+[2025-09-04 13:18:55] [Rank 0] step:4481/10000 train_time:205504ms step_avg:45.86ms
+[2025-09-04 13:18:55] [Rank 0] step:4481/10000 train_time:205504ms step_avg:45.86ms
+[2025-09-04 13:18:55] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:18:55] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:18:56] [Rank 0] PRINT: step:4500/10000 train_loss:0.6687 val_loss:0.6525 train_time:206263ms step_avg:45.84ms
+[2025-09-04 13:18:56] [Rank 0] PRINT: step:4500/10000 train_loss:0.6687 val_loss:0.6525 train_time:206263ms step_avg:45.84ms
+[2025-09-04 13:18:56] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:18:56] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:18:56] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:18:56] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:20:31] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:20:31] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:20:31] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:20:31] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:20:31] [Rank 0]   Total Loss: 4.8689
+[2025-09-04 13:20:31] [Rank 0]   Total Loss: 4.8689
+[2025-09-04 13:20:31] [Rank 0]   Total FTA (Unweighted): 0.8987
+[2025-09-04 13:20:31] [Rank 0]   Total FTA (Unweighted): 0.8987
+[2025-09-04 13:20:31] [Rank 0]   Total FTA (Weighted):   0.8988
+[2025-09-04 13:20:31] [Rank 0]   Total FTA (Weighted):   0.8988
+[2025-09-04 13:20:31] [Rank 0]   Group 0 Loss: 4.8566
+[2025-09-04 13:20:31] [Rank 0]   Group 0 Loss: 4.8566
+[2025-09-04 13:20:31] [Rank 0]   Group 1 Loss: 4.3847
+[2025-09-04 13:20:31] [Rank 0]   Group 1 Loss: 4.3847
+[2025-09-04 13:20:31] [Rank 0]   Group 2 Loss: 4.3071
+[2025-09-04 13:20:31] [Rank 0]   Group 2 Loss: 4.3071
+[2025-09-04 13:20:31] [Rank 0]   Group 3 Loss: 4.7965
+[2025-09-04 13:20:31] [Rank 0]   Group 3 Loss: 4.7965
+[2025-09-04 13:20:31] [Rank 0]   Group 4 Loss: 4.7442
+[2025-09-04 13:20:31] [Rank 0]   Group 4 Loss: 4.7442
+[2025-09-04 13:20:31] [Rank 0]   Group 5 Loss: 4.8383
+[2025-09-04 13:20:31] [Rank 0]   Group 5 Loss: 4.8383
+[2025-09-04 13:20:31] [Rank 0]   Group 6 Loss: 4.7355
+[2025-09-04 13:20:31] [Rank 0]   Group 6 Loss: 4.7355
+[2025-09-04 13:20:31] [Rank 0]   Group 7 Loss: 4.8048
+[2025-09-04 13:20:31] [Rank 0]   Group 7 Loss: 4.8048
+[2025-09-04 13:20:31] [Rank 0]   Group 8 Loss: 4.9766
+[2025-09-04 13:20:31] [Rank 0]   Group 8 Loss: 4.9766
+[2025-09-04 13:20:31] [Rank 0]   Group 9 Loss: 4.9051
+[2025-09-04 13:20:31] [Rank 0]   Group 9 Loss: 4.9051
+[2025-09-04 13:20:31] [Rank 0]   Group 10 Loss: 5.0610
+[2025-09-04 13:20:31] [Rank 0]   Group 10 Loss: 5.0610
+[2025-09-04 13:20:31] [Rank 0]   Group 11 Loss: 5.1217
+[2025-09-04 13:20:31] [Rank 0]   Group 11 Loss: 5.1217
+[2025-09-04 13:20:31] [Rank 0]   Group 12 Loss: 5.0062
+[2025-09-04 13:20:31] [Rank 0]   Group 12 Loss: 5.0062
+[2025-09-04 13:20:31] [Rank 0]   Group 13 Loss: 5.1072
+[2025-09-04 13:20:31] [Rank 0]   Group 13 Loss: 5.1072
+[2025-09-04 13:20:31] [Rank 0]   Group 14 Loss: 5.1100
+[2025-09-04 13:20:31] [Rank 0]   Group 14 Loss: 5.1100
+[2025-09-04 13:20:31] [Rank 0]   Group 15 Loss: 5.1475
+[2025-09-04 13:20:31] [Rank 0]   Group 15 Loss: 5.1475
+[2025-09-04 13:20:31] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 13:20:31] [Rank 0]   Group 10 FTA: 0.9900
+[2025-09-04 13:20:31] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:20:31] [Rank 0]   Group 12 FTA: 0.9700
+[2025-09-04 13:20:31] [Rank 0]   Group 12 FTA: 0.9700
+[2025-09-04 13:20:31] [Rank 0]   Group 13 FTA: 0.8600
+[2025-09-04 13:20:31] [Rank 0]   Group 13 FTA: 0.8600
+[2025-09-04 13:20:31] [Rank 0]   Group 14 FTA: 0.3700
+[2025-09-04 13:20:31] [Rank 0]   Group 14 FTA: 0.3700
+[2025-09-04 13:20:31] [Rank 0]   Group 15 FTA: 0.1900
+[2025-09-04 13:20:31] [Rank 0]   Group 15 FTA: 0.1900
+[2025-09-04 13:20:32] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:20:32] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:20:32] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:20:32] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:20:32] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:20:32] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:20:33] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:20:33] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:20:33] [Rank 0] step:4501/10000 train_time:206278ms step_avg:45.83ms
+[2025-09-04 13:20:33] [Rank 0] step:4501/10000 train_time:206278ms step_avg:45.83ms
+[2025-09-04 13:20:33] [Rank 0] step:4521/10000 train_time:207029ms step_avg:45.79ms
+[2025-09-04 13:20:33] [Rank 0] step:4521/10000 train_time:207029ms step_avg:45.79ms
+[2025-09-04 13:20:34] [Rank 0] step:4541/10000 train_time:207784ms step_avg:45.76ms
+[2025-09-04 13:20:34] [Rank 0] step:4541/10000 train_time:207784ms step_avg:45.76ms
+[2025-09-04 13:20:35] [Rank 0] step:4561/10000 train_time:208538ms step_avg:45.72ms
+[2025-09-04 13:20:35] [Rank 0] step:4561/10000 train_time:208538ms step_avg:45.72ms
+[2025-09-04 13:20:36] [Rank 0] step:4581/10000 train_time:209293ms step_avg:45.69ms
+[2025-09-04 13:20:36] [Rank 0] step:4581/10000 train_time:209293ms step_avg:45.69ms
+[2025-09-04 13:20:36] [Rank 0] step:4601/10000 train_time:210047ms step_avg:45.65ms
+[2025-09-04 13:20:36] [Rank 0] step:4601/10000 train_time:210047ms step_avg:45.65ms
+[2025-09-04 13:20:37] [Rank 0] step:4621/10000 train_time:210801ms step_avg:45.62ms
+[2025-09-04 13:20:37] [Rank 0] step:4621/10000 train_time:210801ms step_avg:45.62ms
+[2025-09-04 13:20:38] [Rank 0] step:4641/10000 train_time:211556ms step_avg:45.58ms
+[2025-09-04 13:20:38] [Rank 0] step:4641/10000 train_time:211556ms step_avg:45.58ms
+[2025-09-04 13:20:39] [Rank 0] step:4661/10000 train_time:212311ms step_avg:45.55ms
+[2025-09-04 13:20:39] [Rank 0] step:4661/10000 train_time:212311ms step_avg:45.55ms
+[2025-09-04 13:20:39] [Rank 0] step:4681/10000 train_time:213065ms step_avg:45.52ms
+[2025-09-04 13:20:39] [Rank 0] step:4681/10000 train_time:213065ms step_avg:45.52ms
+[2025-09-04 13:20:40] [Rank 0] step:4701/10000 train_time:213819ms step_avg:45.48ms
+[2025-09-04 13:20:40] [Rank 0] step:4701/10000 train_time:213819ms step_avg:45.48ms
+[2025-09-04 13:20:41] [Rank 0] step:4721/10000 train_time:214574ms step_avg:45.45ms
+[2025-09-04 13:20:41] [Rank 0] step:4721/10000 train_time:214574ms step_avg:45.45ms
+[2025-09-04 13:20:42] [Rank 0] step:4741/10000 train_time:215329ms step_avg:45.42ms
+[2025-09-04 13:20:42] [Rank 0] step:4741/10000 train_time:215329ms step_avg:45.42ms
+[2025-09-04 13:20:42] [Rank 0] step:4761/10000 train_time:216083ms step_avg:45.39ms
+[2025-09-04 13:20:42] [Rank 0] step:4761/10000 train_time:216083ms step_avg:45.39ms
+[2025-09-04 13:20:43] [Rank 0] step:4781/10000 train_time:216838ms step_avg:45.35ms
+[2025-09-04 13:20:43] [Rank 0] step:4781/10000 train_time:216838ms step_avg:45.35ms
+[2025-09-04 13:20:44] [Rank 0] step:4801/10000 train_time:217593ms step_avg:45.32ms
+[2025-09-04 13:20:44] [Rank 0] step:4801/10000 train_time:217593ms step_avg:45.32ms
+[2025-09-04 13:20:45] [Rank 0] step:4821/10000 train_time:218348ms step_avg:45.29ms
+[2025-09-04 13:20:45] [Rank 0] step:4821/10000 train_time:218348ms step_avg:45.29ms
+[2025-09-04 13:20:46] [Rank 0] step:4841/10000 train_time:219844ms step_avg:45.41ms
+[2025-09-04 13:20:46] [Rank 0] step:4841/10000 train_time:219844ms step_avg:45.41ms
+[2025-09-04 13:20:47] [Rank 0] step:4861/10000 train_time:220598ms step_avg:45.38ms
+[2025-09-04 13:20:47] [Rank 0] step:4861/10000 train_time:220598ms step_avg:45.38ms
+[2025-09-04 13:20:48] [Rank 0] step:4881/10000 train_time:221352ms step_avg:45.35ms
+[2025-09-04 13:20:48] [Rank 0] step:4881/10000 train_time:221352ms step_avg:45.35ms
+[2025-09-04 13:20:49] [Rank 0] step:4901/10000 train_time:222107ms step_avg:45.32ms
+[2025-09-04 13:20:49] [Rank 0] step:4901/10000 train_time:222107ms step_avg:45.32ms
+[2025-09-04 13:20:49] [Rank 0] step:4921/10000 train_time:222860ms step_avg:45.29ms
+[2025-09-04 13:20:49] [Rank 0] step:4921/10000 train_time:222860ms step_avg:45.29ms
+[2025-09-04 13:20:50] [Rank 0] step:4941/10000 train_time:223615ms step_avg:45.26ms
+[2025-09-04 13:20:50] [Rank 0] step:4941/10000 train_time:223615ms step_avg:45.26ms
+[2025-09-04 13:20:51] [Rank 0] step:4961/10000 train_time:224369ms step_avg:45.23ms
+[2025-09-04 13:20:51] [Rank 0] step:4961/10000 train_time:224369ms step_avg:45.23ms
+[2025-09-04 13:20:52] [Rank 0] step:4981/10000 train_time:225124ms step_avg:45.20ms
+[2025-09-04 13:20:52] [Rank 0] step:4981/10000 train_time:225124ms step_avg:45.20ms
+[2025-09-04 13:20:52] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:20:52] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:20:53] [Rank 0] PRINT: step:5000/10000 train_loss:0.6599 val_loss:0.6439 train_time:225884ms step_avg:45.18ms
+[2025-09-04 13:20:53] [Rank 0] PRINT: step:5000/10000 train_loss:0.6599 val_loss:0.6439 train_time:225884ms step_avg:45.18ms
+[2025-09-04 13:20:53] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:20:53] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:20:53] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:20:53] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:22:29] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:22:29] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:22:29] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:22:29] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:22:29] [Rank 0]   Total Loss: 4.9149
+[2025-09-04 13:22:29] [Rank 0]   Total Loss: 4.9149
+[2025-09-04 13:22:29] [Rank 0]   Total FTA (Unweighted): 0.9163
+[2025-09-04 13:22:29] [Rank 0]   Total FTA (Unweighted): 0.9163
+[2025-09-04 13:22:29] [Rank 0]   Total FTA (Weighted):   0.9163
+[2025-09-04 13:22:29] [Rank 0]   Total FTA (Weighted):   0.9163
+[2025-09-04 13:22:29] [Rank 0]   Group 0 Loss: 4.8495
+[2025-09-04 13:22:29] [Rank 0]   Group 0 Loss: 4.8495
+[2025-09-04 13:22:29] [Rank 0]   Group 1 Loss: 4.4483
+[2025-09-04 13:22:29] [Rank 0]   Group 1 Loss: 4.4483
+[2025-09-04 13:22:29] [Rank 0]   Group 2 Loss: 4.3986
+[2025-09-04 13:22:29] [Rank 0]   Group 2 Loss: 4.3986
+[2025-09-04 13:22:29] [Rank 0]   Group 3 Loss: 4.8394
+[2025-09-04 13:22:29] [Rank 0]   Group 3 Loss: 4.8394
+[2025-09-04 13:22:29] [Rank 0]   Group 4 Loss: 4.8372
+[2025-09-04 13:22:29] [Rank 0]   Group 4 Loss: 4.8372
+[2025-09-04 13:22:29] [Rank 0]   Group 5 Loss: 4.8636
+[2025-09-04 13:22:29] [Rank 0]   Group 5 Loss: 4.8636
+[2025-09-04 13:22:29] [Rank 0]   Group 6 Loss: 4.7664
+[2025-09-04 13:22:29] [Rank 0]   Group 6 Loss: 4.7664
+[2025-09-04 13:22:29] [Rank 0]   Group 7 Loss: 4.8446
+[2025-09-04 13:22:29] [Rank 0]   Group 7 Loss: 4.8446
+[2025-09-04 13:22:29] [Rank 0]   Group 8 Loss: 5.0355
+[2025-09-04 13:22:29] [Rank 0]   Group 8 Loss: 5.0355
+[2025-09-04 13:22:29] [Rank 0]   Group 9 Loss: 4.9462
+[2025-09-04 13:22:29] [Rank 0]   Group 9 Loss: 4.9462
+[2025-09-04 13:22:29] [Rank 0]   Group 10 Loss: 5.1183
+[2025-09-04 13:22:29] [Rank 0]   Group 10 Loss: 5.1183
+[2025-09-04 13:22:29] [Rank 0]   Group 11 Loss: 5.1660
+[2025-09-04 13:22:29] [Rank 0]   Group 11 Loss: 5.1660
+[2025-09-04 13:22:29] [Rank 0]   Group 12 Loss: 5.0694
+[2025-09-04 13:22:29] [Rank 0]   Group 12 Loss: 5.0694
+[2025-09-04 13:22:29] [Rank 0]   Group 13 Loss: 5.1672
+[2025-09-04 13:22:29] [Rank 0]   Group 13 Loss: 5.1672
+[2025-09-04 13:22:29] [Rank 0]   Group 14 Loss: 5.1478
+[2025-09-04 13:22:29] [Rank 0]   Group 14 Loss: 5.1478
+[2025-09-04 13:22:29] [Rank 0]   Group 15 Loss: 5.1404
+[2025-09-04 13:22:29] [Rank 0]   Group 15 Loss: 5.1404
+[2025-09-04 13:22:29] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:22:29] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 13:22:29] [Rank 0]   Group 12 FTA: 0.9800
+[2025-09-04 13:22:29] [Rank 0]   Group 13 FTA: 0.9100
+[2025-09-04 13:22:29] [Rank 0]   Group 13 FTA: 0.9100
+[2025-09-04 13:22:29] [Rank 0]   Group 14 FTA: 0.5300
+[2025-09-04 13:22:29] [Rank 0]   Group 14 FTA: 0.5300
+[2025-09-04 13:22:29] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 13:22:29] [Rank 0]   Group 15 FTA: 0.2400
+[2025-09-04 13:22:29] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:22:29] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:22:30] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:22:30] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:22:30] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:22:30] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:22:30] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:22:30] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:22:30] [Rank 0] step:5001/10000 train_time:225898ms step_avg:45.17ms
+[2025-09-04 13:22:30] [Rank 0] step:5001/10000 train_time:225898ms step_avg:45.17ms
+[2025-09-04 13:22:31] [Rank 0] step:5021/10000 train_time:226680ms step_avg:45.15ms
+[2025-09-04 13:22:31] [Rank 0] step:5021/10000 train_time:226680ms step_avg:45.15ms
+[2025-09-04 13:22:32] [Rank 0] step:5041/10000 train_time:227617ms step_avg:45.15ms
+[2025-09-04 13:22:32] [Rank 0] step:5041/10000 train_time:227617ms step_avg:45.15ms
+[2025-09-04 13:22:33] [Rank 0] step:5061/10000 train_time:228445ms step_avg:45.14ms
+[2025-09-04 13:22:33] [Rank 0] step:5061/10000 train_time:228445ms step_avg:45.14ms
+[2025-09-04 13:22:34] [Rank 0] step:5081/10000 train_time:229198ms step_avg:45.11ms
+[2025-09-04 13:22:34] [Rank 0] step:5081/10000 train_time:229198ms step_avg:45.11ms
+[2025-09-04 13:22:35] [Rank 0] step:5101/10000 train_time:230108ms step_avg:45.11ms
+[2025-09-04 13:22:35] [Rank 0] step:5101/10000 train_time:230108ms step_avg:45.11ms
+[2025-09-04 13:22:35] [Rank 0] step:5121/10000 train_time:230978ms step_avg:45.10ms
+[2025-09-04 13:22:35] [Rank 0] step:5121/10000 train_time:230978ms step_avg:45.10ms
+[2025-09-04 13:22:36] [Rank 0] step:5141/10000 train_time:231731ms step_avg:45.08ms
+[2025-09-04 13:22:36] [Rank 0] step:5141/10000 train_time:231731ms step_avg:45.08ms
+[2025-09-04 13:22:37] [Rank 0] step:5161/10000 train_time:232485ms step_avg:45.05ms
+[2025-09-04 13:22:37] [Rank 0] step:5161/10000 train_time:232485ms step_avg:45.05ms
+[2025-09-04 13:22:38] [Rank 0] step:5181/10000 train_time:233239ms step_avg:45.02ms
+[2025-09-04 13:22:38] [Rank 0] step:5181/10000 train_time:233239ms step_avg:45.02ms
+[2025-09-04 13:22:38] [Rank 0] step:5201/10000 train_time:233992ms step_avg:44.99ms
+[2025-09-04 13:22:38] [Rank 0] step:5201/10000 train_time:233992ms step_avg:44.99ms
+[2025-09-04 13:22:39] [Rank 0] step:5221/10000 train_time:234749ms step_avg:44.96ms
+[2025-09-04 13:22:39] [Rank 0] step:5221/10000 train_time:234749ms step_avg:44.96ms
+[2025-09-04 13:22:40] [Rank 0] step:5241/10000 train_time:235503ms step_avg:44.93ms
+[2025-09-04 13:22:40] [Rank 0] step:5241/10000 train_time:235503ms step_avg:44.93ms
+[2025-09-04 13:22:41] [Rank 0] step:5261/10000 train_time:236257ms step_avg:44.91ms
+[2025-09-04 13:22:41] [Rank 0] step:5261/10000 train_time:236257ms step_avg:44.91ms
+[2025-09-04 13:22:41] [Rank 0] step:5281/10000 train_time:237010ms step_avg:44.88ms
+[2025-09-04 13:22:41] [Rank 0] step:5281/10000 train_time:237010ms step_avg:44.88ms
+[2025-09-04 13:22:42] [Rank 0] step:5301/10000 train_time:237764ms step_avg:44.85ms
+[2025-09-04 13:22:42] [Rank 0] step:5301/10000 train_time:237764ms step_avg:44.85ms
+[2025-09-04 13:22:43] [Rank 0] step:5321/10000 train_time:238518ms step_avg:44.83ms
+[2025-09-04 13:22:43] [Rank 0] step:5321/10000 train_time:238518ms step_avg:44.83ms
+[2025-09-04 13:22:44] [Rank 0] step:5341/10000 train_time:239273ms step_avg:44.80ms
+[2025-09-04 13:22:44] [Rank 0] step:5341/10000 train_time:239273ms step_avg:44.80ms
+[2025-09-04 13:22:44] [Rank 0] step:5361/10000 train_time:240027ms step_avg:44.77ms
+[2025-09-04 13:22:44] [Rank 0] step:5361/10000 train_time:240027ms step_avg:44.77ms
+[2025-09-04 13:22:45] [Rank 0] step:5381/10000 train_time:240781ms step_avg:44.75ms
+[2025-09-04 13:22:45] [Rank 0] step:5381/10000 train_time:240781ms step_avg:44.75ms
+[2025-09-04 13:22:46] [Rank 0] step:5401/10000 train_time:241536ms step_avg:44.72ms
+[2025-09-04 13:22:46] [Rank 0] step:5401/10000 train_time:241536ms step_avg:44.72ms
+[2025-09-04 13:22:47] [Rank 0] step:5421/10000 train_time:242290ms step_avg:44.69ms
+[2025-09-04 13:22:47] [Rank 0] step:5421/10000 train_time:242290ms step_avg:44.69ms
+[2025-09-04 13:22:47] [Rank 0] step:5441/10000 train_time:243044ms step_avg:44.67ms
+[2025-09-04 13:22:47] [Rank 0] step:5441/10000 train_time:243044ms step_avg:44.67ms
+[2025-09-04 13:22:48] [Rank 0] step:5461/10000 train_time:243799ms step_avg:44.64ms
+[2025-09-04 13:22:48] [Rank 0] step:5461/10000 train_time:243799ms step_avg:44.64ms
+[2025-09-04 13:22:49] [Rank 0] step:5481/10000 train_time:244554ms step_avg:44.62ms
+[2025-09-04 13:22:49] [Rank 0] step:5481/10000 train_time:244554ms step_avg:44.62ms
+[2025-09-04 13:22:50] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:22:50] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:22:50] [Rank 0] PRINT: step:5500/10000 train_loss:0.6522 val_loss:0.6371 train_time:245313ms step_avg:44.60ms
+[2025-09-04 13:22:50] [Rank 0] PRINT: step:5500/10000 train_loss:0.6522 val_loss:0.6371 train_time:245313ms step_avg:44.60ms
+[2025-09-04 13:22:50] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:22:50] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:22:50] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:22:50] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:24:26] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:24:26] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:24:26] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:24:26] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:24:26] [Rank 0]   Total Loss: 4.9839
+[2025-09-04 13:24:26] [Rank 0]   Total Loss: 4.9839
+[2025-09-04 13:24:26] [Rank 0]   Total FTA (Unweighted): 0.9225
+[2025-09-04 13:24:26] [Rank 0]   Total FTA (Unweighted): 0.9225
+[2025-09-04 13:24:26] [Rank 0]   Total FTA (Weighted):   0.9225
+[2025-09-04 13:24:26] [Rank 0]   Total FTA (Weighted):   0.9225
+[2025-09-04 13:24:26] [Rank 0]   Group 0 Loss: 4.9770
+[2025-09-04 13:24:26] [Rank 0]   Group 0 Loss: 4.9770
+[2025-09-04 13:24:26] [Rank 0]   Group 1 Loss: 4.4953
+[2025-09-04 13:24:26] [Rank 0]   Group 1 Loss: 4.4953
+[2025-09-04 13:24:26] [Rank 0]   Group 2 Loss: 4.4669
+[2025-09-04 13:24:26] [Rank 0]   Group 2 Loss: 4.4669
+[2025-09-04 13:24:26] [Rank 0]   Group 3 Loss: 4.9068
+[2025-09-04 13:24:26] [Rank 0]   Group 3 Loss: 4.9068
+[2025-09-04 13:24:26] [Rank 0]   Group 4 Loss: 4.9188
+[2025-09-04 13:24:26] [Rank 0]   Group 4 Loss: 4.9188
+[2025-09-04 13:24:26] [Rank 0]   Group 5 Loss: 4.9249
+[2025-09-04 13:24:26] [Rank 0]   Group 5 Loss: 4.9249
+[2025-09-04 13:24:26] [Rank 0]   Group 6 Loss: 4.8384
+[2025-09-04 13:24:26] [Rank 0]   Group 6 Loss: 4.8384
+[2025-09-04 13:24:26] [Rank 0]   Group 7 Loss: 4.9227
+[2025-09-04 13:24:26] [Rank 0]   Group 7 Loss: 4.9227
+[2025-09-04 13:24:26] [Rank 0]   Group 8 Loss: 5.0923
+[2025-09-04 13:24:26] [Rank 0]   Group 8 Loss: 5.0923
+[2025-09-04 13:24:26] [Rank 0]   Group 9 Loss: 4.9697
+[2025-09-04 13:24:26] [Rank 0]   Group 9 Loss: 4.9697
+[2025-09-04 13:24:26] [Rank 0]   Group 10 Loss: 5.2060
+[2025-09-04 13:24:26] [Rank 0]   Group 10 Loss: 5.2060
+[2025-09-04 13:24:26] [Rank 0]   Group 11 Loss: 5.2312
+[2025-09-04 13:24:26] [Rank 0]   Group 11 Loss: 5.2312
+[2025-09-04 13:24:26] [Rank 0]   Group 12 Loss: 5.1365
+[2025-09-04 13:24:26] [Rank 0]   Group 12 Loss: 5.1365
+[2025-09-04 13:24:26] [Rank 0]   Group 13 Loss: 5.2055
+[2025-09-04 13:24:26] [Rank 0]   Group 13 Loss: 5.2055
+[2025-09-04 13:24:26] [Rank 0]   Group 14 Loss: 5.2322
+[2025-09-04 13:24:26] [Rank 0]   Group 14 Loss: 5.2322
+[2025-09-04 13:24:26] [Rank 0]   Group 15 Loss: 5.2180
+[2025-09-04 13:24:26] [Rank 0]   Group 15 Loss: 5.2180
+[2025-09-04 13:24:26] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:24:26] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:24:27] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:24:27] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:24:27] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:24:27] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:24:27] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 13:24:27] [Rank 0]   Group 11 FTA: 0.9900
+[2025-09-04 13:24:27] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 13:24:27] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 13:24:27] [Rank 0]   Group 13 FTA: 0.9700
+[2025-09-04 13:24:27] [Rank 0]   Group 13 FTA: 0.9700
+[2025-09-04 13:24:27] [Rank 0]   Group 14 FTA: 0.5800
+[2025-09-04 13:24:27] [Rank 0]   Group 14 FTA: 0.5800
+[2025-09-04 13:24:27] [Rank 0]   Group 15 FTA: 0.2300
+[2025-09-04 13:24:27] [Rank 0]   Group 15 FTA: 0.2300
+[2025-09-04 13:24:27] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:24:27] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:24:27] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:24:27] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:24:28] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:24:28] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:24:28] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:24:28] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:24:28] [Rank 0] step:5501/10000 train_time:245328ms step_avg:44.60ms
+[2025-09-04 13:24:28] [Rank 0] step:5501/10000 train_time:245328ms step_avg:44.60ms
+[2025-09-04 13:24:29] [Rank 0] step:5521/10000 train_time:246092ms step_avg:44.57ms
+[2025-09-04 13:24:29] [Rank 0] step:5521/10000 train_time:246092ms step_avg:44.57ms
+[2025-09-04 13:24:29] [Rank 0] step:5541/10000 train_time:246847ms step_avg:44.55ms
+[2025-09-04 13:24:29] [Rank 0] step:5541/10000 train_time:246847ms step_avg:44.55ms
+[2025-09-04 13:24:30] [Rank 0] step:5561/10000 train_time:247601ms step_avg:44.52ms
+[2025-09-04 13:24:30] [Rank 0] step:5561/10000 train_time:247601ms step_avg:44.52ms
+[2025-09-04 13:24:31] [Rank 0] step:5581/10000 train_time:248356ms step_avg:44.50ms
+[2025-09-04 13:24:31] [Rank 0] step:5581/10000 train_time:248356ms step_avg:44.50ms
+[2025-09-04 13:24:32] [Rank 0] step:5601/10000 train_time:249110ms step_avg:44.48ms
+[2025-09-04 13:24:32] [Rank 0] step:5601/10000 train_time:249110ms step_avg:44.48ms
+[2025-09-04 13:24:32] [Rank 0] step:5621/10000 train_time:249865ms step_avg:44.45ms
+[2025-09-04 13:24:32] [Rank 0] step:5621/10000 train_time:249865ms step_avg:44.45ms
+[2025-09-04 13:24:33] [Rank 0] step:5641/10000 train_time:250885ms step_avg:44.48ms
+[2025-09-04 13:24:33] [Rank 0] step:5641/10000 train_time:250885ms step_avg:44.48ms
+[2025-09-04 13:24:34] [Rank 0] step:5661/10000 train_time:251639ms step_avg:44.45ms
+[2025-09-04 13:24:34] [Rank 0] step:5661/10000 train_time:251639ms step_avg:44.45ms
+[2025-09-04 13:24:35] [Rank 0] step:5681/10000 train_time:252393ms step_avg:44.43ms
+[2025-09-04 13:24:35] [Rank 0] step:5681/10000 train_time:252393ms step_avg:44.43ms
+[2025-09-04 13:24:36] [Rank 0] step:5701/10000 train_time:253148ms step_avg:44.40ms
+[2025-09-04 13:24:36] [Rank 0] step:5701/10000 train_time:253148ms step_avg:44.40ms
+[2025-09-04 13:24:37] [Rank 0] step:5721/10000 train_time:253903ms step_avg:44.38ms
+[2025-09-04 13:24:37] [Rank 0] step:5721/10000 train_time:253903ms step_avg:44.38ms
+[2025-09-04 13:24:37] [Rank 0] step:5741/10000 train_time:254658ms step_avg:44.36ms
+[2025-09-04 13:24:37] [Rank 0] step:5741/10000 train_time:254658ms step_avg:44.36ms
+[2025-09-04 13:24:38] [Rank 0] step:5761/10000 train_time:255411ms step_avg:44.33ms
+[2025-09-04 13:24:38] [Rank 0] step:5761/10000 train_time:255411ms step_avg:44.33ms
+[2025-09-04 13:24:39] [Rank 0] step:5781/10000 train_time:256403ms step_avg:44.35ms
+[2025-09-04 13:24:39] [Rank 0] step:5781/10000 train_time:256403ms step_avg:44.35ms
+[2025-09-04 13:24:40] [Rank 0] step:5801/10000 train_time:257158ms step_avg:44.33ms
+[2025-09-04 13:24:40] [Rank 0] step:5801/10000 train_time:257158ms step_avg:44.33ms
+[2025-09-04 13:24:41] [Rank 0] step:5821/10000 train_time:257912ms step_avg:44.31ms
+[2025-09-04 13:24:41] [Rank 0] step:5821/10000 train_time:257912ms step_avg:44.31ms
+[2025-09-04 13:24:42] [Rank 0] step:5841/10000 train_time:258901ms step_avg:44.32ms
+[2025-09-04 13:24:42] [Rank 0] step:5841/10000 train_time:258901ms step_avg:44.32ms
+[2025-09-04 13:24:42] [Rank 0] step:5861/10000 train_time:259656ms step_avg:44.30ms
+[2025-09-04 13:24:42] [Rank 0] step:5861/10000 train_time:259656ms step_avg:44.30ms
+[2025-09-04 13:24:43] [Rank 0] step:5881/10000 train_time:260410ms step_avg:44.28ms
+[2025-09-04 13:24:43] [Rank 0] step:5881/10000 train_time:260410ms step_avg:44.28ms
+[2025-09-04 13:24:44] [Rank 0] step:5901/10000 train_time:261165ms step_avg:44.26ms
+[2025-09-04 13:24:44] [Rank 0] step:5901/10000 train_time:261165ms step_avg:44.26ms
+[2025-09-04 13:24:45] [Rank 0] step:5921/10000 train_time:261919ms step_avg:44.24ms
+[2025-09-04 13:24:45] [Rank 0] step:5921/10000 train_time:261919ms step_avg:44.24ms
+[2025-09-04 13:24:45] [Rank 0] step:5941/10000 train_time:262674ms step_avg:44.21ms
+[2025-09-04 13:24:45] [Rank 0] step:5941/10000 train_time:262674ms step_avg:44.21ms
+[2025-09-04 13:24:46] [Rank 0] step:5961/10000 train_time:263428ms step_avg:44.19ms
+[2025-09-04 13:24:46] [Rank 0] step:5961/10000 train_time:263428ms step_avg:44.19ms
+[2025-09-04 13:24:47] [Rank 0] step:5981/10000 train_time:264182ms step_avg:44.17ms
+[2025-09-04 13:24:47] [Rank 0] step:5981/10000 train_time:264182ms step_avg:44.17ms
+[2025-09-04 13:24:48] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:24:48] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:24:48] [Rank 0] PRINT: step:6000/10000 train_loss:0.6450 val_loss:0.6309 train_time:264942ms step_avg:44.16ms
+[2025-09-04 13:24:48] [Rank 0] PRINT: step:6000/10000 train_loss:0.6450 val_loss:0.6309 train_time:264942ms step_avg:44.16ms
+[2025-09-04 13:24:48] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:24:48] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:24:48] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:24:48] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:26:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:26:24] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:26:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:26:24] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:26:24] [Rank 0]   Total Loss: 4.9662
+[2025-09-04 13:26:24] [Rank 0]   Total Loss: 4.9662
+[2025-09-04 13:26:24] [Rank 0]   Total FTA (Unweighted): 0.9406
+[2025-09-04 13:26:24] [Rank 0]   Total FTA (Unweighted): 0.9406
+[2025-09-04 13:26:24] [Rank 0]   Total FTA (Weighted):   0.9406
+[2025-09-04 13:26:24] [Rank 0]   Total FTA (Weighted):   0.9406
+[2025-09-04 13:26:24] [Rank 0]   Group 0 Loss: 4.9480
+[2025-09-04 13:26:24] [Rank 0]   Group 0 Loss: 4.9480
+[2025-09-04 13:26:24] [Rank 0]   Group 1 Loss: 4.5031
+[2025-09-04 13:26:24] [Rank 0]   Group 1 Loss: 4.5031
+[2025-09-04 13:26:24] [Rank 0]   Group 2 Loss: 4.4111
+[2025-09-04 13:26:24] [Rank 0]   Group 2 Loss: 4.4111
+[2025-09-04 13:26:24] [Rank 0]   Group 3 Loss: 4.8624
+[2025-09-04 13:26:24] [Rank 0]   Group 3 Loss: 4.8624
+[2025-09-04 13:26:24] [Rank 0]   Group 4 Loss: 4.9012
+[2025-09-04 13:26:24] [Rank 0]   Group 4 Loss: 4.9012
+[2025-09-04 13:26:24] [Rank 0]   Group 5 Loss: 4.9246
+[2025-09-04 13:26:24] [Rank 0]   Group 5 Loss: 4.9246
+[2025-09-04 13:26:24] [Rank 0]   Group 6 Loss: 4.8216
+[2025-09-04 13:26:24] [Rank 0]   Group 6 Loss: 4.8216
+[2025-09-04 13:26:24] [Rank 0]   Group 7 Loss: 4.9097
+[2025-09-04 13:26:24] [Rank 0]   Group 7 Loss: 4.9097
+[2025-09-04 13:26:24] [Rank 0]   Group 8 Loss: 5.0859
+[2025-09-04 13:26:24] [Rank 0]   Group 8 Loss: 5.0859
+[2025-09-04 13:26:24] [Rank 0]   Group 9 Loss: 4.9614
+[2025-09-04 13:26:24] [Rank 0]   Group 9 Loss: 4.9614
+[2025-09-04 13:26:24] [Rank 0]   Group 10 Loss: 5.1996
+[2025-09-04 13:26:24] [Rank 0]   Group 10 Loss: 5.1996
+[2025-09-04 13:26:24] [Rank 0]   Group 11 Loss: 5.2391
+[2025-09-04 13:26:24] [Rank 0]   Group 11 Loss: 5.2391
+[2025-09-04 13:26:24] [Rank 0]   Group 12 Loss: 5.0903
+[2025-09-04 13:26:24] [Rank 0]   Group 12 Loss: 5.0903
+[2025-09-04 13:26:24] [Rank 0]   Group 13 Loss: 5.1941
+[2025-09-04 13:26:24] [Rank 0]   Group 13 Loss: 5.1941
+[2025-09-04 13:26:24] [Rank 0]   Group 14 Loss: 5.2183
+[2025-09-04 13:26:24] [Rank 0]   Group 14 Loss: 5.2183
+[2025-09-04 13:26:24] [Rank 0]   Group 15 Loss: 5.1891
+[2025-09-04 13:26:24] [Rank 0]   Group 15 Loss: 5.1891
+[2025-09-04 13:26:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 13:26:24] [Rank 0]   Group 11 FTA: 0.9800
+[2025-09-04 13:26:24] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:26:24] [Rank 0]   Group 13 FTA: 0.9600
+[2025-09-04 13:26:24] [Rank 0]   Group 13 FTA: 0.9600
+[2025-09-04 13:26:24] [Rank 0]   Group 14 FTA: 0.8000
+[2025-09-04 13:26:24] [Rank 0]   Group 14 FTA: 0.8000
+[2025-09-04 13:26:24] [Rank 0]   Group 15 FTA: 0.3100
+[2025-09-04 13:26:24] [Rank 0]   Group 15 FTA: 0.3100
+[2025-09-04 13:26:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:26:24] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:26:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:26:25] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:26:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:26:25] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:26:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:26:25] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:26:25] [Rank 0] step:6001/10000 train_time:264959ms step_avg:44.15ms
+[2025-09-04 13:26:25] [Rank 0] step:6001/10000 train_time:264959ms step_avg:44.15ms
+[2025-09-04 13:26:26] [Rank 0] step:6021/10000 train_time:265800ms step_avg:44.15ms
+[2025-09-04 13:26:26] [Rank 0] step:6021/10000 train_time:265800ms step_avg:44.15ms
+[2025-09-04 13:26:27] [Rank 0] step:6041/10000 train_time:266554ms step_avg:44.12ms
+[2025-09-04 13:26:27] [Rank 0] step:6041/10000 train_time:266554ms step_avg:44.12ms
+[2025-09-04 13:26:28] [Rank 0] step:6061/10000 train_time:267309ms step_avg:44.10ms
+[2025-09-04 13:26:28] [Rank 0] step:6061/10000 train_time:267309ms step_avg:44.10ms
+[2025-09-04 13:26:29] [Rank 0] step:6081/10000 train_time:268063ms step_avg:44.08ms
+[2025-09-04 13:26:29] [Rank 0] step:6081/10000 train_time:268063ms step_avg:44.08ms
+[2025-09-04 13:26:29] [Rank 0] step:6101/10000 train_time:268816ms step_avg:44.06ms
+[2025-09-04 13:26:29] [Rank 0] step:6101/10000 train_time:268816ms step_avg:44.06ms
+[2025-09-04 13:26:30] [Rank 0] step:6121/10000 train_time:269570ms step_avg:44.04ms
+[2025-09-04 13:26:30] [Rank 0] step:6121/10000 train_time:269570ms step_avg:44.04ms
+[2025-09-04 13:26:31] [Rank 0] step:6141/10000 train_time:270324ms step_avg:44.02ms
+[2025-09-04 13:26:31] [Rank 0] step:6141/10000 train_time:270324ms step_avg:44.02ms
+[2025-09-04 13:26:32] [Rank 0] step:6161/10000 train_time:271078ms step_avg:44.00ms
+[2025-09-04 13:26:32] [Rank 0] step:6161/10000 train_time:271078ms step_avg:44.00ms
+[2025-09-04 13:26:32] [Rank 0] step:6181/10000 train_time:271832ms step_avg:43.98ms
+[2025-09-04 13:26:32] [Rank 0] step:6181/10000 train_time:271832ms step_avg:43.98ms
+[2025-09-04 13:26:33] [Rank 0] step:6201/10000 train_time:272585ms step_avg:43.96ms
+[2025-09-04 13:26:33] [Rank 0] step:6201/10000 train_time:272585ms step_avg:43.96ms
+[2025-09-04 13:26:34] [Rank 0] step:6221/10000 train_time:273339ms step_avg:43.94ms
+[2025-09-04 13:26:34] [Rank 0] step:6221/10000 train_time:273339ms step_avg:43.94ms
+[2025-09-04 13:26:35] [Rank 0] step:6241/10000 train_time:274092ms step_avg:43.92ms
+[2025-09-04 13:26:35] [Rank 0] step:6241/10000 train_time:274092ms step_avg:43.92ms
+[2025-09-04 13:26:35] [Rank 0] step:6261/10000 train_time:274846ms step_avg:43.90ms
+[2025-09-04 13:26:35] [Rank 0] step:6261/10000 train_time:274846ms step_avg:43.90ms
+[2025-09-04 13:26:36] [Rank 0] step:6281/10000 train_time:275625ms step_avg:43.88ms
+[2025-09-04 13:26:36] [Rank 0] step:6281/10000 train_time:275625ms step_avg:43.88ms
+[2025-09-04 13:26:37] [Rank 0] step:6301/10000 train_time:276379ms step_avg:43.86ms
+[2025-09-04 13:26:37] [Rank 0] step:6301/10000 train_time:276379ms step_avg:43.86ms
+[2025-09-04 13:26:38] [Rank 0] step:6321/10000 train_time:277132ms step_avg:43.84ms
+[2025-09-04 13:26:38] [Rank 0] step:6321/10000 train_time:277132ms step_avg:43.84ms
+[2025-09-04 13:26:38] [Rank 0] step:6341/10000 train_time:277886ms step_avg:43.82ms
+[2025-09-04 13:26:38] [Rank 0] step:6341/10000 train_time:277886ms step_avg:43.82ms
+[2025-09-04 13:26:39] [Rank 0] step:6361/10000 train_time:278640ms step_avg:43.80ms
+[2025-09-04 13:26:39] [Rank 0] step:6361/10000 train_time:278640ms step_avg:43.80ms
+[2025-09-04 13:26:40] [Rank 0] step:6381/10000 train_time:279394ms step_avg:43.79ms
+[2025-09-04 13:26:40] [Rank 0] step:6381/10000 train_time:279394ms step_avg:43.79ms
+[2025-09-04 13:26:41] [Rank 0] step:6401/10000 train_time:280148ms step_avg:43.77ms
+[2025-09-04 13:26:41] [Rank 0] step:6401/10000 train_time:280148ms step_avg:43.77ms
+[2025-09-04 13:26:41] [Rank 0] step:6421/10000 train_time:280902ms step_avg:43.75ms
+[2025-09-04 13:26:41] [Rank 0] step:6421/10000 train_time:280902ms step_avg:43.75ms
+[2025-09-04 13:26:42] [Rank 0] step:6441/10000 train_time:281657ms step_avg:43.73ms
+[2025-09-04 13:26:42] [Rank 0] step:6441/10000 train_time:281657ms step_avg:43.73ms
+[2025-09-04 13:26:43] [Rank 0] step:6461/10000 train_time:282411ms step_avg:43.71ms
+[2025-09-04 13:26:43] [Rank 0] step:6461/10000 train_time:282411ms step_avg:43.71ms
+[2025-09-04 13:26:44] [Rank 0] step:6481/10000 train_time:283165ms step_avg:43.69ms
+[2025-09-04 13:26:44] [Rank 0] step:6481/10000 train_time:283165ms step_avg:43.69ms
+[2025-09-04 13:26:44] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:26:44] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:26:45] [Rank 0] PRINT: step:6500/10000 train_loss:0.6385 val_loss:0.6256 train_time:283924ms step_avg:43.68ms
+[2025-09-04 13:26:45] [Rank 0] PRINT: step:6500/10000 train_loss:0.6385 val_loss:0.6256 train_time:283924ms step_avg:43.68ms
+[2025-09-04 13:26:45] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:26:45] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:26:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:26:45] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:28:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:28:21] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:28:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:28:21] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:28:21] [Rank 0]   Total Loss: 4.9253
+[2025-09-04 13:28:21] [Rank 0]   Total Loss: 4.9253
+[2025-09-04 13:28:21] [Rank 0]   Total FTA (Unweighted): 0.9569
+[2025-09-04 13:28:21] [Rank 0]   Total FTA (Unweighted): 0.9569
+[2025-09-04 13:28:22] [Rank 0]   Total FTA (Weighted):   0.9569
+[2025-09-04 13:28:22] [Rank 0]   Total FTA (Weighted):   0.9569
+[2025-09-04 13:28:22] [Rank 0]   Group 0 Loss: 4.8455
+[2025-09-04 13:28:22] [Rank 0]   Group 0 Loss: 4.8455
+[2025-09-04 13:28:22] [Rank 0]   Group 1 Loss: 4.4206
+[2025-09-04 13:28:22] [Rank 0]   Group 1 Loss: 4.4206
+[2025-09-04 13:28:22] [Rank 0]   Group 2 Loss: 4.3969
+[2025-09-04 13:28:22] [Rank 0]   Group 2 Loss: 4.3969
+[2025-09-04 13:28:22] [Rank 0]   Group 3 Loss: 4.7932
+[2025-09-04 13:28:22] [Rank 0]   Group 3 Loss: 4.7932
+[2025-09-04 13:28:22] [Rank 0]   Group 4 Loss: 4.8734
+[2025-09-04 13:28:22] [Rank 0]   Group 4 Loss: 4.8734
+[2025-09-04 13:28:22] [Rank 0]   Group 5 Loss: 4.9114
+[2025-09-04 13:28:22] [Rank 0]   Group 5 Loss: 4.9114
+[2025-09-04 13:28:22] [Rank 0]   Group 6 Loss: 4.8110
+[2025-09-04 13:28:22] [Rank 0]   Group 6 Loss: 4.8110
+[2025-09-04 13:28:22] [Rank 0]   Group 7 Loss: 4.8879
+[2025-09-04 13:28:22] [Rank 0]   Group 7 Loss: 4.8879
+[2025-09-04 13:28:22] [Rank 0]   Group 8 Loss: 5.0222
+[2025-09-04 13:28:22] [Rank 0]   Group 8 Loss: 5.0222
+[2025-09-04 13:28:22] [Rank 0]   Group 9 Loss: 4.9427
+[2025-09-04 13:28:22] [Rank 0]   Group 9 Loss: 4.9427
+[2025-09-04 13:28:22] [Rank 0]   Group 10 Loss: 5.1542
+[2025-09-04 13:28:22] [Rank 0]   Group 10 Loss: 5.1542
+[2025-09-04 13:28:22] [Rank 0]   Group 11 Loss: 5.1850
+[2025-09-04 13:28:22] [Rank 0]   Group 11 Loss: 5.1850
+[2025-09-04 13:28:22] [Rank 0]   Group 12 Loss: 5.0777
+[2025-09-04 13:28:22] [Rank 0]   Group 12 Loss: 5.0777
+[2025-09-04 13:28:22] [Rank 0]   Group 13 Loss: 5.1427
+[2025-09-04 13:28:22] [Rank 0]   Group 13 Loss: 5.1427
+[2025-09-04 13:28:22] [Rank 0]   Group 14 Loss: 5.1789
+[2025-09-04 13:28:22] [Rank 0]   Group 14 Loss: 5.1789
+[2025-09-04 13:28:22] [Rank 0]   Group 15 Loss: 5.1619
+[2025-09-04 13:28:22] [Rank 0]   Group 15 Loss: 5.1619
+[2025-09-04 13:28:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 13:28:22] [Rank 0]   Group 8 FTA: 0.9900
+[2025-09-04 13:28:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:28:22] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:28:22] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:28:22] [Rank 0]   Group 14 FTA: 0.8500
+[2025-09-04 13:28:22] [Rank 0]   Group 14 FTA: 0.8500
+[2025-09-04 13:28:22] [Rank 0]   Group 15 FTA: 0.4800
+[2025-09-04 13:28:22] [Rank 0]   Group 15 FTA: 0.4800
+[2025-09-04 13:28:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:28:22] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:28:23] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:28:23] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:28:23] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:28:23] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:28:23] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:28:23] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:28:23] [Rank 0] step:6501/10000 train_time:283940ms step_avg:43.68ms
+[2025-09-04 13:28:23] [Rank 0] step:6501/10000 train_time:283940ms step_avg:43.68ms
+[2025-09-04 13:28:24] [Rank 0] step:6521/10000 train_time:284704ms step_avg:43.66ms
+[2025-09-04 13:28:24] [Rank 0] step:6521/10000 train_time:284704ms step_avg:43.66ms
+[2025-09-04 13:28:25] [Rank 0] step:6541/10000 train_time:285459ms step_avg:43.64ms
+[2025-09-04 13:28:25] [Rank 0] step:6541/10000 train_time:285459ms step_avg:43.64ms
+[2025-09-04 13:28:25] [Rank 0] step:6561/10000 train_time:286214ms step_avg:43.62ms
+[2025-09-04 13:28:25] [Rank 0] step:6561/10000 train_time:286214ms step_avg:43.62ms
+[2025-09-04 13:28:26] [Rank 0] step:6581/10000 train_time:286969ms step_avg:43.61ms
+[2025-09-04 13:28:26] [Rank 0] step:6581/10000 train_time:286969ms step_avg:43.61ms
+[2025-09-04 13:28:27] [Rank 0] step:6601/10000 train_time:287724ms step_avg:43.59ms
+[2025-09-04 13:28:27] [Rank 0] step:6601/10000 train_time:287724ms step_avg:43.59ms
+[2025-09-04 13:28:28] [Rank 0] step:6621/10000 train_time:288480ms step_avg:43.57ms
+[2025-09-04 13:28:28] [Rank 0] step:6621/10000 train_time:288480ms step_avg:43.57ms
+[2025-09-04 13:28:28] [Rank 0] step:6641/10000 train_time:289236ms step_avg:43.55ms
+[2025-09-04 13:28:28] [Rank 0] step:6641/10000 train_time:289236ms step_avg:43.55ms
+[2025-09-04 13:28:29] [Rank 0] step:6661/10000 train_time:289992ms step_avg:43.54ms
+[2025-09-04 13:28:29] [Rank 0] step:6661/10000 train_time:289992ms step_avg:43.54ms
+[2025-09-04 13:28:30] [Rank 0] step:6681/10000 train_time:290746ms step_avg:43.52ms
+[2025-09-04 13:28:30] [Rank 0] step:6681/10000 train_time:290746ms step_avg:43.52ms
+[2025-09-04 13:28:31] [Rank 0] step:6701/10000 train_time:291502ms step_avg:43.50ms
+[2025-09-04 13:28:31] [Rank 0] step:6701/10000 train_time:291502ms step_avg:43.50ms
+[2025-09-04 13:28:31] [Rank 0] step:6721/10000 train_time:292259ms step_avg:43.48ms
+[2025-09-04 13:28:31] [Rank 0] step:6721/10000 train_time:292259ms step_avg:43.48ms
+[2025-09-04 13:28:32] [Rank 0] step:6741/10000 train_time:293015ms step_avg:43.47ms
+[2025-09-04 13:28:32] [Rank 0] step:6741/10000 train_time:293015ms step_avg:43.47ms
+[2025-09-04 13:28:33] [Rank 0] step:6761/10000 train_time:293770ms step_avg:43.45ms
+[2025-09-04 13:28:33] [Rank 0] step:6761/10000 train_time:293770ms step_avg:43.45ms
+[2025-09-04 13:28:34] [Rank 0] step:6781/10000 train_time:294525ms step_avg:43.43ms
+[2025-09-04 13:28:34] [Rank 0] step:6781/10000 train_time:294525ms step_avg:43.43ms
+[2025-09-04 13:28:35] [Rank 0] step:6801/10000 train_time:295283ms step_avg:43.42ms
+[2025-09-04 13:28:35] [Rank 0] step:6801/10000 train_time:295283ms step_avg:43.42ms
+[2025-09-04 13:28:35] [Rank 0] step:6821/10000 train_time:296038ms step_avg:43.40ms
+[2025-09-04 13:28:35] [Rank 0] step:6821/10000 train_time:296038ms step_avg:43.40ms
+[2025-09-04 13:28:37] [Rank 0] step:6841/10000 train_time:297495ms step_avg:43.49ms
+[2025-09-04 13:28:37] [Rank 0] step:6841/10000 train_time:297495ms step_avg:43.49ms
+[2025-09-04 13:28:37] [Rank 0] step:6861/10000 train_time:298251ms step_avg:43.47ms
+[2025-09-04 13:28:37] [Rank 0] step:6861/10000 train_time:298251ms step_avg:43.47ms
+[2025-09-04 13:28:38] [Rank 0] step:6881/10000 train_time:299006ms step_avg:43.45ms
+[2025-09-04 13:28:38] [Rank 0] step:6881/10000 train_time:299006ms step_avg:43.45ms
+[2025-09-04 13:28:39] [Rank 0] step:6901/10000 train_time:299761ms step_avg:43.44ms
+[2025-09-04 13:28:39] [Rank 0] step:6901/10000 train_time:299761ms step_avg:43.44ms
+[2025-09-04 13:28:40] [Rank 0] step:6921/10000 train_time:300517ms step_avg:43.42ms
+[2025-09-04 13:28:40] [Rank 0] step:6921/10000 train_time:300517ms step_avg:43.42ms
+[2025-09-04 13:28:41] [Rank 0] step:6941/10000 train_time:301272ms step_avg:43.40ms
+[2025-09-04 13:28:41] [Rank 0] step:6941/10000 train_time:301272ms step_avg:43.40ms
+[2025-09-04 13:28:41] [Rank 0] step:6961/10000 train_time:302027ms step_avg:43.39ms
+[2025-09-04 13:28:41] [Rank 0] step:6961/10000 train_time:302027ms step_avg:43.39ms
+[2025-09-04 13:28:42] [Rank 0] step:6981/10000 train_time:302782ms step_avg:43.37ms
+[2025-09-04 13:28:42] [Rank 0] step:6981/10000 train_time:302782ms step_avg:43.37ms
+[2025-09-04 13:28:43] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:28:43] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:28:43] [Rank 0] PRINT: step:7000/10000 train_loss:0.6324 val_loss:0.6207 train_time:303542ms step_avg:43.36ms
+[2025-09-04 13:28:43] [Rank 0] PRINT: step:7000/10000 train_loss:0.6324 val_loss:0.6207 train_time:303542ms step_avg:43.36ms
+[2025-09-04 13:28:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:28:43] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:28:43] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:28:43] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:30:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:30:20] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:30:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:30:20] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:30:20] [Rank 0]   Total Loss: 4.9255
+[2025-09-04 13:30:20] [Rank 0]   Total Loss: 4.9255
+[2025-09-04 13:30:20] [Rank 0]   Total FTA (Unweighted): 0.9631
+[2025-09-04 13:30:20] [Rank 0]   Total FTA (Unweighted): 0.9631
+[2025-09-04 13:30:20] [Rank 0]   Total FTA (Weighted):   0.9631
+[2025-09-04 13:30:20] [Rank 0]   Total FTA (Weighted):   0.9631
+[2025-09-04 13:30:20] [Rank 0]   Group 0 Loss: 4.8546
+[2025-09-04 13:30:20] [Rank 0]   Group 0 Loss: 4.8546
+[2025-09-04 13:30:20] [Rank 0]   Group 1 Loss: 4.4488
+[2025-09-04 13:30:20] [Rank 0]   Group 1 Loss: 4.4488
+[2025-09-04 13:30:20] [Rank 0]   Group 2 Loss: 4.4055
+[2025-09-04 13:30:20] [Rank 0]   Group 2 Loss: 4.4055
+[2025-09-04 13:30:20] [Rank 0]   Group 3 Loss: 4.7944
+[2025-09-04 13:30:20] [Rank 0]   Group 3 Loss: 4.7944
+[2025-09-04 13:30:20] [Rank 0]   Group 4 Loss: 4.8514
+[2025-09-04 13:30:20] [Rank 0]   Group 4 Loss: 4.8514
+[2025-09-04 13:30:20] [Rank 0]   Group 5 Loss: 4.8917
+[2025-09-04 13:30:20] [Rank 0]   Group 5 Loss: 4.8917
+[2025-09-04 13:30:20] [Rank 0]   Group 6 Loss: 4.7864
+[2025-09-04 13:30:20] [Rank 0]   Group 6 Loss: 4.7864
+[2025-09-04 13:30:20] [Rank 0]   Group 7 Loss: 4.8721
+[2025-09-04 13:30:20] [Rank 0]   Group 7 Loss: 4.8721
+[2025-09-04 13:30:20] [Rank 0]   Group 8 Loss: 5.0168
+[2025-09-04 13:30:20] [Rank 0]   Group 8 Loss: 5.0168
+[2025-09-04 13:30:20] [Rank 0]   Group 9 Loss: 4.9448
+[2025-09-04 13:30:20] [Rank 0]   Group 9 Loss: 4.9448
+[2025-09-04 13:30:20] [Rank 0]   Group 10 Loss: 5.1372
+[2025-09-04 13:30:20] [Rank 0]   Group 10 Loss: 5.1372
+[2025-09-04 13:30:20] [Rank 0]   Group 11 Loss: 5.1956
+[2025-09-04 13:30:20] [Rank 0]   Group 11 Loss: 5.1956
+[2025-09-04 13:30:20] [Rank 0]   Group 12 Loss: 5.0763
+[2025-09-04 13:30:20] [Rank 0]   Group 12 Loss: 5.0763
+[2025-09-04 13:30:20] [Rank 0]   Group 13 Loss: 5.1779
+[2025-09-04 13:30:20] [Rank 0]   Group 13 Loss: 5.1779
+[2025-09-04 13:30:20] [Rank 0]   Group 14 Loss: 5.1950
+[2025-09-04 13:30:20] [Rank 0]   Group 14 Loss: 5.1950
+[2025-09-04 13:30:20] [Rank 0]   Group 15 Loss: 5.1588
+[2025-09-04 13:30:20] [Rank 0]   Group 15 Loss: 5.1588
+[2025-09-04 13:30:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:30:20] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:30:20] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:30:20] [Rank 0]   Group 14 FTA: 0.9000
+[2025-09-04 13:30:20] [Rank 0]   Group 14 FTA: 0.9000
+[2025-09-04 13:30:20] [Rank 0]   Group 15 FTA: 0.5200
+[2025-09-04 13:30:20] [Rank 0]   Group 15 FTA: 0.5200
+[2025-09-04 13:30:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:30:20] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:30:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:30:21] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:30:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:30:21] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:30:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:30:21] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:30:21] [Rank 0] step:7001/10000 train_time:303558ms step_avg:43.36ms
+[2025-09-04 13:30:21] [Rank 0] step:7001/10000 train_time:303558ms step_avg:43.36ms
+[2025-09-04 13:30:22] [Rank 0] step:7021/10000 train_time:304323ms step_avg:43.34ms
+[2025-09-04 13:30:22] [Rank 0] step:7021/10000 train_time:304323ms step_avg:43.34ms
+[2025-09-04 13:30:23] [Rank 0] step:7041/10000 train_time:305077ms step_avg:43.33ms
+[2025-09-04 13:30:23] [Rank 0] step:7041/10000 train_time:305077ms step_avg:43.33ms
+[2025-09-04 13:30:23] [Rank 0] step:7061/10000 train_time:305830ms step_avg:43.31ms
+[2025-09-04 13:30:23] [Rank 0] step:7061/10000 train_time:305830ms step_avg:43.31ms
+[2025-09-04 13:30:24] [Rank 0] step:7081/10000 train_time:306584ms step_avg:43.30ms
+[2025-09-04 13:30:24] [Rank 0] step:7081/10000 train_time:306584ms step_avg:43.30ms
+[2025-09-04 13:30:25] [Rank 0] step:7101/10000 train_time:307338ms step_avg:43.28ms
+[2025-09-04 13:30:25] [Rank 0] step:7101/10000 train_time:307338ms step_avg:43.28ms
+[2025-09-04 13:30:26] [Rank 0] step:7121/10000 train_time:308093ms step_avg:43.27ms
+[2025-09-04 13:30:26] [Rank 0] step:7121/10000 train_time:308093ms step_avg:43.27ms
+[2025-09-04 13:30:26] [Rank 0] step:7141/10000 train_time:308846ms step_avg:43.25ms
+[2025-09-04 13:30:26] [Rank 0] step:7141/10000 train_time:308846ms step_avg:43.25ms
+[2025-09-04 13:30:27] [Rank 0] step:7161/10000 train_time:309600ms step_avg:43.23ms
+[2025-09-04 13:30:27] [Rank 0] step:7161/10000 train_time:309600ms step_avg:43.23ms
+[2025-09-04 13:30:28] [Rank 0] step:7181/10000 train_time:310354ms step_avg:43.22ms
+[2025-09-04 13:30:28] [Rank 0] step:7181/10000 train_time:310354ms step_avg:43.22ms
+[2025-09-04 13:30:29] [Rank 0] step:7201/10000 train_time:311108ms step_avg:43.20ms
+[2025-09-04 13:30:29] [Rank 0] step:7201/10000 train_time:311108ms step_avg:43.20ms
+[2025-09-04 13:30:29] [Rank 0] step:7221/10000 train_time:311862ms step_avg:43.19ms
+[2025-09-04 13:30:29] [Rank 0] step:7221/10000 train_time:311862ms step_avg:43.19ms
+[2025-09-04 13:30:30] [Rank 0] step:7241/10000 train_time:312616ms step_avg:43.17ms
+[2025-09-04 13:30:30] [Rank 0] step:7241/10000 train_time:312616ms step_avg:43.17ms
+[2025-09-04 13:30:31] [Rank 0] step:7261/10000 train_time:313370ms step_avg:43.16ms
+[2025-09-04 13:30:31] [Rank 0] step:7261/10000 train_time:313370ms step_avg:43.16ms
+[2025-09-04 13:30:32] [Rank 0] step:7281/10000 train_time:314124ms step_avg:43.14ms
+[2025-09-04 13:30:32] [Rank 0] step:7281/10000 train_time:314124ms step_avg:43.14ms
+[2025-09-04 13:30:32] [Rank 0] step:7301/10000 train_time:314878ms step_avg:43.13ms
+[2025-09-04 13:30:32] [Rank 0] step:7301/10000 train_time:314878ms step_avg:43.13ms
+[2025-09-04 13:30:33] [Rank 0] step:7321/10000 train_time:315633ms step_avg:43.11ms
+[2025-09-04 13:30:33] [Rank 0] step:7321/10000 train_time:315633ms step_avg:43.11ms
+[2025-09-04 13:30:34] [Rank 0] step:7341/10000 train_time:316387ms step_avg:43.10ms
+[2025-09-04 13:30:34] [Rank 0] step:7341/10000 train_time:316387ms step_avg:43.10ms
+[2025-09-04 13:30:35] [Rank 0] step:7361/10000 train_time:317140ms step_avg:43.08ms
+[2025-09-04 13:30:35] [Rank 0] step:7361/10000 train_time:317140ms step_avg:43.08ms
+[2025-09-04 13:30:35] [Rank 0] step:7381/10000 train_time:317895ms step_avg:43.07ms
+[2025-09-04 13:30:35] [Rank 0] step:7381/10000 train_time:317895ms step_avg:43.07ms
+[2025-09-04 13:30:36] [Rank 0] step:7401/10000 train_time:318649ms step_avg:43.05ms
+[2025-09-04 13:30:36] [Rank 0] step:7401/10000 train_time:318649ms step_avg:43.05ms
+[2025-09-04 13:30:37] [Rank 0] step:7421/10000 train_time:319403ms step_avg:43.04ms
+[2025-09-04 13:30:37] [Rank 0] step:7421/10000 train_time:319403ms step_avg:43.04ms
+[2025-09-04 13:30:38] [Rank 0] step:7441/10000 train_time:320157ms step_avg:43.03ms
+[2025-09-04 13:30:38] [Rank 0] step:7441/10000 train_time:320157ms step_avg:43.03ms
+[2025-09-04 13:30:38] [Rank 0] step:7461/10000 train_time:320911ms step_avg:43.01ms
+[2025-09-04 13:30:38] [Rank 0] step:7461/10000 train_time:320911ms step_avg:43.01ms
+[2025-09-04 13:30:39] [Rank 0] step:7481/10000 train_time:321665ms step_avg:43.00ms
+[2025-09-04 13:30:39] [Rank 0] step:7481/10000 train_time:321665ms step_avg:43.00ms
+[2025-09-04 13:30:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:30:40] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:30:40] [Rank 0] PRINT: step:7500/10000 train_loss:0.6270 val_loss:0.6167 train_time:322424ms step_avg:42.99ms
+[2025-09-04 13:30:40] [Rank 0] PRINT: step:7500/10000 train_loss:0.6270 val_loss:0.6167 train_time:322424ms step_avg:42.99ms
+[2025-09-04 13:30:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:30:40] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:30:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:30:41] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:32:17] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:32:17] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:32:17] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:32:17] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:32:17] [Rank 0]   Total Loss: 4.9393
+[2025-09-04 13:32:17] [Rank 0]   Total Loss: 4.9393
+[2025-09-04 13:32:17] [Rank 0]   Total FTA (Unweighted): 0.9700
+[2025-09-04 13:32:17] [Rank 0]   Total FTA (Unweighted): 0.9700
+[2025-09-04 13:32:17] [Rank 0]   Total FTA (Weighted):   0.9700
+[2025-09-04 13:32:17] [Rank 0]   Total FTA (Weighted):   0.9700
+[2025-09-04 13:32:17] [Rank 0]   Group 0 Loss: 4.9580
+[2025-09-04 13:32:17] [Rank 0]   Group 0 Loss: 4.9580
+[2025-09-04 13:32:17] [Rank 0]   Group 1 Loss: 4.4753
+[2025-09-04 13:32:17] [Rank 0]   Group 1 Loss: 4.4753
+[2025-09-04 13:32:17] [Rank 0]   Group 2 Loss: 4.4189
+[2025-09-04 13:32:17] [Rank 0]   Group 2 Loss: 4.4189
+[2025-09-04 13:32:17] [Rank 0]   Group 3 Loss: 4.8311
+[2025-09-04 13:32:17] [Rank 0]   Group 3 Loss: 4.8311
+[2025-09-04 13:32:17] [Rank 0]   Group 4 Loss: 4.8531
+[2025-09-04 13:32:17] [Rank 0]   Group 4 Loss: 4.8531
+[2025-09-04 13:32:17] [Rank 0]   Group 5 Loss: 4.8797
+[2025-09-04 13:32:17] [Rank 0]   Group 5 Loss: 4.8797
+[2025-09-04 13:32:17] [Rank 0]   Group 6 Loss: 4.8071
+[2025-09-04 13:32:17] [Rank 0]   Group 6 Loss: 4.8071
+[2025-09-04 13:32:17] [Rank 0]   Group 7 Loss: 4.8948
+[2025-09-04 13:32:17] [Rank 0]   Group 7 Loss: 4.8948
+[2025-09-04 13:32:17] [Rank 0]   Group 8 Loss: 5.0138
+[2025-09-04 13:32:17] [Rank 0]   Group 8 Loss: 5.0138
+[2025-09-04 13:32:17] [Rank 0]   Group 9 Loss: 4.9597
+[2025-09-04 13:32:17] [Rank 0]   Group 9 Loss: 4.9597
+[2025-09-04 13:32:17] [Rank 0]   Group 10 Loss: 5.1407
+[2025-09-04 13:32:17] [Rank 0]   Group 10 Loss: 5.1407
+[2025-09-04 13:32:17] [Rank 0]   Group 11 Loss: 5.1961
+[2025-09-04 13:32:17] [Rank 0]   Group 11 Loss: 5.1961
+[2025-09-04 13:32:17] [Rank 0]   Group 12 Loss: 5.0999
+[2025-09-04 13:32:17] [Rank 0]   Group 12 Loss: 5.0999
+[2025-09-04 13:32:17] [Rank 0]   Group 13 Loss: 5.1739
+[2025-09-04 13:32:17] [Rank 0]   Group 13 Loss: 5.1739
+[2025-09-04 13:32:17] [Rank 0]   Group 14 Loss: 5.1712
+[2025-09-04 13:32:17] [Rank 0]   Group 14 Loss: 5.1712
+[2025-09-04 13:32:17] [Rank 0]   Group 15 Loss: 5.1553
+[2025-09-04 13:32:17] [Rank 0]   Group 15 Loss: 5.1553
+[2025-09-04 13:32:17] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 13:32:17] [Rank 0]   Group 12 FTA: 0.9900
+[2025-09-04 13:32:17] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:32:17] [Rank 0]   Group 14 FTA: 0.9100
+[2025-09-04 13:32:17] [Rank 0]   Group 14 FTA: 0.9100
+[2025-09-04 13:32:17] [Rank 0]   Group 15 FTA: 0.6200
+[2025-09-04 13:32:17] [Rank 0]   Group 15 FTA: 0.6200
+[2025-09-04 13:32:18] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:32:18] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:32:18] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:32:18] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:32:18] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:32:18] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:32:19] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:32:19] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:32:19] [Rank 0] step:7501/10000 train_time:322439ms step_avg:42.99ms
+[2025-09-04 13:32:19] [Rank 0] step:7501/10000 train_time:322439ms step_avg:42.99ms
+[2025-09-04 13:32:19] [Rank 0] step:7521/10000 train_time:323206ms step_avg:42.97ms
+[2025-09-04 13:32:19] [Rank 0] step:7521/10000 train_time:323206ms step_avg:42.97ms
+[2025-09-04 13:32:20] [Rank 0] step:7541/10000 train_time:323960ms step_avg:42.96ms
+[2025-09-04 13:32:20] [Rank 0] step:7541/10000 train_time:323960ms step_avg:42.96ms
+[2025-09-04 13:32:21] [Rank 0] step:7561/10000 train_time:324715ms step_avg:42.95ms
+[2025-09-04 13:32:21] [Rank 0] step:7561/10000 train_time:324715ms step_avg:42.95ms
+[2025-09-04 13:32:22] [Rank 0] step:7581/10000 train_time:325469ms step_avg:42.93ms
+[2025-09-04 13:32:22] [Rank 0] step:7581/10000 train_time:325469ms step_avg:42.93ms
+[2025-09-04 13:32:23] [Rank 0] step:7601/10000 train_time:326227ms step_avg:42.92ms
+[2025-09-04 13:32:23] [Rank 0] step:7601/10000 train_time:326227ms step_avg:42.92ms
+[2025-09-04 13:32:23] [Rank 0] step:7621/10000 train_time:326981ms step_avg:42.91ms
+[2025-09-04 13:32:23] [Rank 0] step:7621/10000 train_time:326981ms step_avg:42.91ms
+[2025-09-04 13:32:25] [Rank 0] step:7641/10000 train_time:328430ms step_avg:42.98ms
+[2025-09-04 13:32:25] [Rank 0] step:7641/10000 train_time:328430ms step_avg:42.98ms
+[2025-09-04 13:32:25] [Rank 0] step:7661/10000 train_time:329185ms step_avg:42.97ms
+[2025-09-04 13:32:25] [Rank 0] step:7661/10000 train_time:329185ms step_avg:42.97ms
+[2025-09-04 13:32:26] [Rank 0] step:7681/10000 train_time:329940ms step_avg:42.96ms
+[2025-09-04 13:32:26] [Rank 0] step:7681/10000 train_time:329940ms step_avg:42.96ms
+[2025-09-04 13:32:27] [Rank 0] step:7701/10000 train_time:330695ms step_avg:42.94ms
+[2025-09-04 13:32:27] [Rank 0] step:7701/10000 train_time:330695ms step_avg:42.94ms
+[2025-09-04 13:32:28] [Rank 0] step:7721/10000 train_time:331452ms step_avg:42.93ms
+[2025-09-04 13:32:28] [Rank 0] step:7721/10000 train_time:331452ms step_avg:42.93ms
+[2025-09-04 13:32:28] [Rank 0] step:7741/10000 train_time:332207ms step_avg:42.92ms
+[2025-09-04 13:32:28] [Rank 0] step:7741/10000 train_time:332207ms step_avg:42.92ms
+[2025-09-04 13:32:29] [Rank 0] step:7761/10000 train_time:332962ms step_avg:42.90ms
+[2025-09-04 13:32:29] [Rank 0] step:7761/10000 train_time:332962ms step_avg:42.90ms
+[2025-09-04 13:32:30] [Rank 0] step:7781/10000 train_time:333717ms step_avg:42.89ms
+[2025-09-04 13:32:30] [Rank 0] step:7781/10000 train_time:333717ms step_avg:42.89ms
+[2025-09-04 13:32:31] [Rank 0] step:7801/10000 train_time:334473ms step_avg:42.88ms
+[2025-09-04 13:32:31] [Rank 0] step:7801/10000 train_time:334473ms step_avg:42.88ms
+[2025-09-04 13:32:32] [Rank 0] step:7821/10000 train_time:335228ms step_avg:42.86ms
+[2025-09-04 13:32:32] [Rank 0] step:7821/10000 train_time:335228ms step_avg:42.86ms
+[2025-09-04 13:32:32] [Rank 0] step:7841/10000 train_time:335982ms step_avg:42.85ms
+[2025-09-04 13:32:32] [Rank 0] step:7841/10000 train_time:335982ms step_avg:42.85ms
+[2025-09-04 13:32:33] [Rank 0] step:7861/10000 train_time:336737ms step_avg:42.84ms
+[2025-09-04 13:32:33] [Rank 0] step:7861/10000 train_time:336737ms step_avg:42.84ms
+[2025-09-04 13:32:34] [Rank 0] step:7881/10000 train_time:337492ms step_avg:42.82ms
+[2025-09-04 13:32:34] [Rank 0] step:7881/10000 train_time:337492ms step_avg:42.82ms
+[2025-09-04 13:32:35] [Rank 0] step:7901/10000 train_time:338248ms step_avg:42.81ms
+[2025-09-04 13:32:35] [Rank 0] step:7901/10000 train_time:338248ms step_avg:42.81ms
+[2025-09-04 13:32:35] [Rank 0] step:7921/10000 train_time:339003ms step_avg:42.80ms
+[2025-09-04 13:32:35] [Rank 0] step:7921/10000 train_time:339003ms step_avg:42.80ms
+[2025-09-04 13:32:36] [Rank 0] step:7941/10000 train_time:339814ms step_avg:42.79ms
+[2025-09-04 13:32:36] [Rank 0] step:7941/10000 train_time:339814ms step_avg:42.79ms
+[2025-09-04 13:32:37] [Rank 0] step:7961/10000 train_time:340568ms step_avg:42.78ms
+[2025-09-04 13:32:37] [Rank 0] step:7961/10000 train_time:340568ms step_avg:42.78ms
+[2025-09-04 13:32:38] [Rank 0] step:7981/10000 train_time:341321ms step_avg:42.77ms
+[2025-09-04 13:32:38] [Rank 0] step:7981/10000 train_time:341321ms step_avg:42.77ms
+[2025-09-04 13:32:38] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:32:38] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:32:39] [Rank 0] PRINT: step:8000/10000 train_loss:0.6222 val_loss:0.6130 train_time:342081ms step_avg:42.76ms
+[2025-09-04 13:32:39] [Rank 0] PRINT: step:8000/10000 train_loss:0.6222 val_loss:0.6130 train_time:342081ms step_avg:42.76ms
+[2025-09-04 13:32:39] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:32:39] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:32:39] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:32:39] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:34:15] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:34:15] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:34:15] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:34:15] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:34:15] [Rank 0]   Total Loss: 4.9553
+[2025-09-04 13:34:15] [Rank 0]   Total Loss: 4.9553
+[2025-09-04 13:34:15] [Rank 0]   Total FTA (Unweighted): 0.9781
+[2025-09-04 13:34:15] [Rank 0]   Total FTA (Unweighted): 0.9781
+[2025-09-04 13:34:15] [Rank 0]   Total FTA (Weighted):   0.9781
+[2025-09-04 13:34:15] [Rank 0]   Total FTA (Weighted):   0.9781
+[2025-09-04 13:34:15] [Rank 0]   Group 0 Loss: 4.8652
+[2025-09-04 13:34:15] [Rank 0]   Group 0 Loss: 4.8652
+[2025-09-04 13:34:15] [Rank 0]   Group 1 Loss: 4.5301
+[2025-09-04 13:34:15] [Rank 0]   Group 1 Loss: 4.5301
+[2025-09-04 13:34:15] [Rank 0]   Group 2 Loss: 4.4300
+[2025-09-04 13:34:15] [Rank 0]   Group 2 Loss: 4.4300
+[2025-09-04 13:34:15] [Rank 0]   Group 3 Loss: 4.8388
+[2025-09-04 13:34:15] [Rank 0]   Group 3 Loss: 4.8388
+[2025-09-04 13:34:15] [Rank 0]   Group 4 Loss: 4.8712
+[2025-09-04 13:34:15] [Rank 0]   Group 4 Loss: 4.8712
+[2025-09-04 13:34:15] [Rank 0]   Group 5 Loss: 4.8997
+[2025-09-04 13:34:15] [Rank 0]   Group 5 Loss: 4.8997
+[2025-09-04 13:34:15] [Rank 0]   Group 6 Loss: 4.8212
+[2025-09-04 13:34:15] [Rank 0]   Group 6 Loss: 4.8212
+[2025-09-04 13:34:15] [Rank 0]   Group 7 Loss: 4.8988
+[2025-09-04 13:34:15] [Rank 0]   Group 7 Loss: 4.8988
+[2025-09-04 13:34:15] [Rank 0]   Group 8 Loss: 5.0473
+[2025-09-04 13:34:15] [Rank 0]   Group 8 Loss: 5.0473
+[2025-09-04 13:34:15] [Rank 0]   Group 9 Loss: 4.9914
+[2025-09-04 13:34:15] [Rank 0]   Group 9 Loss: 4.9914
+[2025-09-04 13:34:15] [Rank 0]   Group 10 Loss: 5.1748
+[2025-09-04 13:34:15] [Rank 0]   Group 10 Loss: 5.1748
+[2025-09-04 13:34:15] [Rank 0]   Group 11 Loss: 5.2088
+[2025-09-04 13:34:15] [Rank 0]   Group 11 Loss: 5.2088
+[2025-09-04 13:34:15] [Rank 0]   Group 12 Loss: 5.1110
+[2025-09-04 13:34:15] [Rank 0]   Group 12 Loss: 5.1110
+[2025-09-04 13:34:15] [Rank 0]   Group 13 Loss: 5.1916
+[2025-09-04 13:34:15] [Rank 0]   Group 13 Loss: 5.1916
+[2025-09-04 13:34:15] [Rank 0]   Group 14 Loss: 5.2206
+[2025-09-04 13:34:15] [Rank 0]   Group 14 Loss: 5.2206
+[2025-09-04 13:34:15] [Rank 0]   Group 15 Loss: 5.1842
+[2025-09-04 13:34:15] [Rank 0]   Group 15 Loss: 5.1842
+[2025-09-04 13:34:15] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:34:15] [Rank 0]   Group 14 FTA: 0.9200
+[2025-09-04 13:34:15] [Rank 0]   Group 14 FTA: 0.9200
+[2025-09-04 13:34:15] [Rank 0]   Group 15 FTA: 0.7300
+[2025-09-04 13:34:15] [Rank 0]   Group 15 FTA: 0.7300
+[2025-09-04 13:34:16] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:34:16] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:34:16] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:34:16] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:34:16] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:34:16] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:34:17] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:34:17] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:34:17] [Rank 0] step:8001/10000 train_time:342096ms step_avg:42.76ms
+[2025-09-04 13:34:17] [Rank 0] step:8001/10000 train_time:342096ms step_avg:42.76ms
+[2025-09-04 13:34:18] [Rank 0] step:8021/10000 train_time:342930ms step_avg:42.75ms
+[2025-09-04 13:34:18] [Rank 0] step:8021/10000 train_time:342930ms step_avg:42.75ms
+[2025-09-04 13:34:18] [Rank 0] step:8041/10000 train_time:343685ms step_avg:42.74ms
+[2025-09-04 13:34:18] [Rank 0] step:8041/10000 train_time:343685ms step_avg:42.74ms
+[2025-09-04 13:34:19] [Rank 0] step:8061/10000 train_time:344439ms step_avg:42.73ms
+[2025-09-04 13:34:19] [Rank 0] step:8061/10000 train_time:344439ms step_avg:42.73ms
+[2025-09-04 13:34:20] [Rank 0] step:8081/10000 train_time:345192ms step_avg:42.72ms
+[2025-09-04 13:34:20] [Rank 0] step:8081/10000 train_time:345192ms step_avg:42.72ms
+[2025-09-04 13:34:21] [Rank 0] step:8101/10000 train_time:345946ms step_avg:42.70ms
+[2025-09-04 13:34:21] [Rank 0] step:8101/10000 train_time:345946ms step_avg:42.70ms
+[2025-09-04 13:34:21] [Rank 0] step:8121/10000 train_time:346700ms step_avg:42.69ms
+[2025-09-04 13:34:21] [Rank 0] step:8121/10000 train_time:346700ms step_avg:42.69ms
+[2025-09-04 13:34:22] [Rank 0] step:8141/10000 train_time:347455ms step_avg:42.68ms
+[2025-09-04 13:34:22] [Rank 0] step:8141/10000 train_time:347455ms step_avg:42.68ms
+[2025-09-04 13:34:23] [Rank 0] step:8161/10000 train_time:348208ms step_avg:42.67ms
+[2025-09-04 13:34:23] [Rank 0] step:8161/10000 train_time:348208ms step_avg:42.67ms
+[2025-09-04 13:34:24] [Rank 0] step:8181/10000 train_time:348962ms step_avg:42.66ms
+[2025-09-04 13:34:24] [Rank 0] step:8181/10000 train_time:348962ms step_avg:42.66ms
+[2025-09-04 13:34:24] [Rank 0] step:8201/10000 train_time:349716ms step_avg:42.64ms
+[2025-09-04 13:34:24] [Rank 0] step:8201/10000 train_time:349716ms step_avg:42.64ms
+[2025-09-04 13:34:25] [Rank 0] step:8221/10000 train_time:350470ms step_avg:42.63ms
+[2025-09-04 13:34:25] [Rank 0] step:8221/10000 train_time:350470ms step_avg:42.63ms
+[2025-09-04 13:34:26] [Rank 0] step:8241/10000 train_time:351224ms step_avg:42.62ms
+[2025-09-04 13:34:26] [Rank 0] step:8241/10000 train_time:351224ms step_avg:42.62ms
+[2025-09-04 13:34:27] [Rank 0] step:8261/10000 train_time:351978ms step_avg:42.61ms
+[2025-09-04 13:34:27] [Rank 0] step:8261/10000 train_time:351978ms step_avg:42.61ms
+[2025-09-04 13:34:27] [Rank 0] step:8281/10000 train_time:352732ms step_avg:42.60ms
+[2025-09-04 13:34:27] [Rank 0] step:8281/10000 train_time:352732ms step_avg:42.60ms
+[2025-09-04 13:34:28] [Rank 0] step:8301/10000 train_time:353489ms step_avg:42.58ms
+[2025-09-04 13:34:28] [Rank 0] step:8301/10000 train_time:353489ms step_avg:42.58ms
+[2025-09-04 13:34:29] [Rank 0] step:8321/10000 train_time:354242ms step_avg:42.57ms
+[2025-09-04 13:34:29] [Rank 0] step:8321/10000 train_time:354242ms step_avg:42.57ms
+[2025-09-04 13:34:30] [Rank 0] step:8341/10000 train_time:354996ms step_avg:42.56ms
+[2025-09-04 13:34:30] [Rank 0] step:8341/10000 train_time:354996ms step_avg:42.56ms
+[2025-09-04 13:34:30] [Rank 0] step:8361/10000 train_time:355751ms step_avg:42.55ms
+[2025-09-04 13:34:30] [Rank 0] step:8361/10000 train_time:355751ms step_avg:42.55ms
+[2025-09-04 13:34:31] [Rank 0] step:8381/10000 train_time:356505ms step_avg:42.54ms
+[2025-09-04 13:34:31] [Rank 0] step:8381/10000 train_time:356505ms step_avg:42.54ms
+[2025-09-04 13:34:32] [Rank 0] step:8401/10000 train_time:357259ms step_avg:42.53ms
+[2025-09-04 13:34:32] [Rank 0] step:8401/10000 train_time:357259ms step_avg:42.53ms
+[2025-09-04 13:34:33] [Rank 0] step:8421/10000 train_time:358013ms step_avg:42.51ms
+[2025-09-04 13:34:33] [Rank 0] step:8421/10000 train_time:358013ms step_avg:42.51ms
+[2025-09-04 13:34:33] [Rank 0] step:8441/10000 train_time:358767ms step_avg:42.50ms
+[2025-09-04 13:34:33] [Rank 0] step:8441/10000 train_time:358767ms step_avg:42.50ms
+[2025-09-04 13:34:34] [Rank 0] step:8461/10000 train_time:359522ms step_avg:42.49ms
+[2025-09-04 13:34:34] [Rank 0] step:8461/10000 train_time:359522ms step_avg:42.49ms
+[2025-09-04 13:34:35] [Rank 0] step:8481/10000 train_time:360276ms step_avg:42.48ms
+[2025-09-04 13:34:35] [Rank 0] step:8481/10000 train_time:360276ms step_avg:42.48ms
+[2025-09-04 13:34:36] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:34:36] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:34:36] [Rank 0] PRINT: step:8500/10000 train_loss:0.6178 val_loss:0.6101 train_time:361034ms step_avg:42.47ms
+[2025-09-04 13:34:36] [Rank 0] PRINT: step:8500/10000 train_loss:0.6178 val_loss:0.6101 train_time:361034ms step_avg:42.47ms
+[2025-09-04 13:34:36] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:34:36] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:34:36] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:34:36] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:36:12] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:36:12] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:36:12] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:36:12] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:36:12] [Rank 0]   Total Loss: 4.9368
+[2025-09-04 13:36:12] [Rank 0]   Total Loss: 4.9368
+[2025-09-04 13:36:12] [Rank 0]   Total FTA (Unweighted): 0.9881
+[2025-09-04 13:36:12] [Rank 0]   Total FTA (Unweighted): 0.9881
+[2025-09-04 13:36:12] [Rank 0]   Total FTA (Weighted):   0.9881
+[2025-09-04 13:36:12] [Rank 0]   Total FTA (Weighted):   0.9881
+[2025-09-04 13:36:12] [Rank 0]   Group 0 Loss: 4.8392
+[2025-09-04 13:36:12] [Rank 0]   Group 0 Loss: 4.8392
+[2025-09-04 13:36:12] [Rank 0]   Group 1 Loss: 4.5194
+[2025-09-04 13:36:12] [Rank 0]   Group 1 Loss: 4.5194
+[2025-09-04 13:36:12] [Rank 0]   Group 2 Loss: 4.4365
+[2025-09-04 13:36:12] [Rank 0]   Group 2 Loss: 4.4365
+[2025-09-04 13:36:12] [Rank 0]   Group 3 Loss: 4.8297
+[2025-09-04 13:36:12] [Rank 0]   Group 3 Loss: 4.8297
+[2025-09-04 13:36:12] [Rank 0]   Group 4 Loss: 4.8615
+[2025-09-04 13:36:12] [Rank 0]   Group 4 Loss: 4.8615
+[2025-09-04 13:36:12] [Rank 0]   Group 5 Loss: 4.8879
+[2025-09-04 13:36:12] [Rank 0]   Group 5 Loss: 4.8879
+[2025-09-04 13:36:12] [Rank 0]   Group 6 Loss: 4.7933
+[2025-09-04 13:36:12] [Rank 0]   Group 6 Loss: 4.7933
+[2025-09-04 13:36:12] [Rank 0]   Group 7 Loss: 4.8866
+[2025-09-04 13:36:12] [Rank 0]   Group 7 Loss: 4.8866
+[2025-09-04 13:36:12] [Rank 0]   Group 8 Loss: 5.0196
+[2025-09-04 13:36:12] [Rank 0]   Group 8 Loss: 5.0196
+[2025-09-04 13:36:12] [Rank 0]   Group 9 Loss: 4.9591
+[2025-09-04 13:36:12] [Rank 0]   Group 9 Loss: 4.9591
+[2025-09-04 13:36:12] [Rank 0]   Group 10 Loss: 5.1504
+[2025-09-04 13:36:12] [Rank 0]   Group 10 Loss: 5.1504
+[2025-09-04 13:36:12] [Rank 0]   Group 11 Loss: 5.1769
+[2025-09-04 13:36:12] [Rank 0]   Group 11 Loss: 5.1769
+[2025-09-04 13:36:12] [Rank 0]   Group 12 Loss: 5.0932
+[2025-09-04 13:36:12] [Rank 0]   Group 12 Loss: 5.0932
+[2025-09-04 13:36:12] [Rank 0]   Group 13 Loss: 5.1657
+[2025-09-04 13:36:12] [Rank 0]   Group 13 Loss: 5.1657
+[2025-09-04 13:36:12] [Rank 0]   Group 14 Loss: 5.2120
+[2025-09-04 13:36:12] [Rank 0]   Group 14 Loss: 5.2120
+[2025-09-04 13:36:12] [Rank 0]   Group 15 Loss: 5.1587
+[2025-09-04 13:36:12] [Rank 0]   Group 15 Loss: 5.1587
+[2025-09-04 13:36:12] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:36:12] [Rank 0]   Group 14 FTA: 0.9700
+[2025-09-04 13:36:12] [Rank 0]   Group 14 FTA: 0.9700
+[2025-09-04 13:36:12] [Rank 0]   Group 15 FTA: 0.8400
+[2025-09-04 13:36:12] [Rank 0]   Group 15 FTA: 0.8400
+[2025-09-04 13:36:13] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:36:13] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:36:13] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:36:13] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:36:13] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:36:13] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:36:14] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:36:14] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:36:14] [Rank 0] step:8501/10000 train_time:361050ms step_avg:42.47ms
+[2025-09-04 13:36:14] [Rank 0] step:8501/10000 train_time:361050ms step_avg:42.47ms
+[2025-09-04 13:36:15] [Rank 0] step:8521/10000 train_time:362027ms step_avg:42.49ms
+[2025-09-04 13:36:15] [Rank 0] step:8521/10000 train_time:362027ms step_avg:42.49ms
+[2025-09-04 13:36:15] [Rank 0] step:8541/10000 train_time:362781ms step_avg:42.48ms
+[2025-09-04 13:36:15] [Rank 0] step:8541/10000 train_time:362781ms step_avg:42.48ms
+[2025-09-04 13:36:16] [Rank 0] step:8561/10000 train_time:363535ms step_avg:42.46ms
+[2025-09-04 13:36:16] [Rank 0] step:8561/10000 train_time:363535ms step_avg:42.46ms
+[2025-09-04 13:36:17] [Rank 0] step:8581/10000 train_time:364486ms step_avg:42.48ms
+[2025-09-04 13:36:17] [Rank 0] step:8581/10000 train_time:364486ms step_avg:42.48ms
+[2025-09-04 13:36:18] [Rank 0] step:8601/10000 train_time:365239ms step_avg:42.46ms
+[2025-09-04 13:36:18] [Rank 0] step:8601/10000 train_time:365239ms step_avg:42.46ms
+[2025-09-04 13:36:19] [Rank 0] step:8621/10000 train_time:365993ms step_avg:42.45ms
+[2025-09-04 13:36:19] [Rank 0] step:8621/10000 train_time:365993ms step_avg:42.45ms
+[2025-09-04 13:36:19] [Rank 0] step:8641/10000 train_time:366748ms step_avg:42.44ms
+[2025-09-04 13:36:19] [Rank 0] step:8641/10000 train_time:366748ms step_avg:42.44ms
+[2025-09-04 13:36:20] [Rank 0] step:8661/10000 train_time:367502ms step_avg:42.43ms
+[2025-09-04 13:36:20] [Rank 0] step:8661/10000 train_time:367502ms step_avg:42.43ms
+[2025-09-04 13:36:21] [Rank 0] step:8681/10000 train_time:368256ms step_avg:42.42ms
+[2025-09-04 13:36:21] [Rank 0] step:8681/10000 train_time:368256ms step_avg:42.42ms
+[2025-09-04 13:36:22] [Rank 0] step:8701/10000 train_time:369010ms step_avg:42.41ms
+[2025-09-04 13:36:22] [Rank 0] step:8701/10000 train_time:369010ms step_avg:42.41ms
+[2025-09-04 13:36:22] [Rank 0] step:8721/10000 train_time:369765ms step_avg:42.40ms
+[2025-09-04 13:36:22] [Rank 0] step:8721/10000 train_time:369765ms step_avg:42.40ms
+[2025-09-04 13:36:23] [Rank 0] step:8741/10000 train_time:370519ms step_avg:42.39ms
+[2025-09-04 13:36:23] [Rank 0] step:8741/10000 train_time:370519ms step_avg:42.39ms
+[2025-09-04 13:36:24] [Rank 0] step:8761/10000 train_time:371273ms step_avg:42.38ms
+[2025-09-04 13:36:24] [Rank 0] step:8761/10000 train_time:371273ms step_avg:42.38ms
+[2025-09-04 13:36:25] [Rank 0] step:8781/10000 train_time:372027ms step_avg:42.37ms
+[2025-09-04 13:36:25] [Rank 0] step:8781/10000 train_time:372027ms step_avg:42.37ms
+[2025-09-04 13:36:25] [Rank 0] step:8801/10000 train_time:372781ms step_avg:42.36ms
+[2025-09-04 13:36:25] [Rank 0] step:8801/10000 train_time:372781ms step_avg:42.36ms
+[2025-09-04 13:36:26] [Rank 0] step:8821/10000 train_time:373534ms step_avg:42.35ms
+[2025-09-04 13:36:26] [Rank 0] step:8821/10000 train_time:373534ms step_avg:42.35ms
+[2025-09-04 13:36:27] [Rank 0] step:8841/10000 train_time:374559ms step_avg:42.37ms
+[2025-09-04 13:36:27] [Rank 0] step:8841/10000 train_time:374559ms step_avg:42.37ms
+[2025-09-04 13:36:28] [Rank 0] step:8861/10000 train_time:375314ms step_avg:42.36ms
+[2025-09-04 13:36:28] [Rank 0] step:8861/10000 train_time:375314ms step_avg:42.36ms
+[2025-09-04 13:36:29] [Rank 0] step:8881/10000 train_time:376068ms step_avg:42.35ms
+[2025-09-04 13:36:29] [Rank 0] step:8881/10000 train_time:376068ms step_avg:42.35ms
+[2025-09-04 13:36:29] [Rank 0] step:8901/10000 train_time:376823ms step_avg:42.33ms
+[2025-09-04 13:36:29] [Rank 0] step:8901/10000 train_time:376823ms step_avg:42.33ms
+[2025-09-04 13:36:30] [Rank 0] step:8921/10000 train_time:377577ms step_avg:42.32ms
+[2025-09-04 13:36:30] [Rank 0] step:8921/10000 train_time:377577ms step_avg:42.32ms
+[2025-09-04 13:36:31] [Rank 0] step:8941/10000 train_time:378331ms step_avg:42.31ms
+[2025-09-04 13:36:31] [Rank 0] step:8941/10000 train_time:378331ms step_avg:42.31ms
+[2025-09-04 13:36:32] [Rank 0] step:8961/10000 train_time:379085ms step_avg:42.30ms
+[2025-09-04 13:36:32] [Rank 0] step:8961/10000 train_time:379085ms step_avg:42.30ms
+[2025-09-04 13:36:32] [Rank 0] step:8981/10000 train_time:379839ms step_avg:42.29ms
+[2025-09-04 13:36:32] [Rank 0] step:8981/10000 train_time:379839ms step_avg:42.29ms
+[2025-09-04 13:36:33] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:36:33] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:36:34] [Rank 0] PRINT: step:9000/10000 train_loss:0.6141 val_loss:0.6080 train_time:380598ms step_avg:42.29ms
+[2025-09-04 13:36:34] [Rank 0] PRINT: step:9000/10000 train_loss:0.6141 val_loss:0.6080 train_time:380598ms step_avg:42.29ms
+[2025-09-04 13:36:34] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:36:34] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:36:34] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:36:34] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:38:11] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:38:11] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:38:11] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:38:11] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:38:11] [Rank 0]   Total Loss: 4.9701
+[2025-09-04 13:38:11] [Rank 0]   Total Loss: 4.9701
+[2025-09-04 13:38:11] [Rank 0]   Total FTA (Unweighted): 0.9925
+[2025-09-04 13:38:11] [Rank 0]   Total FTA (Unweighted): 0.9925
+[2025-09-04 13:38:11] [Rank 0]   Total FTA (Weighted):   0.9925
+[2025-09-04 13:38:11] [Rank 0]   Total FTA (Weighted):   0.9925
+[2025-09-04 13:38:11] [Rank 0]   Group 0 Loss: 4.8504
+[2025-09-04 13:38:11] [Rank 0]   Group 0 Loss: 4.8504
+[2025-09-04 13:38:11] [Rank 0]   Group 1 Loss: 4.5433
+[2025-09-04 13:38:11] [Rank 0]   Group 1 Loss: 4.5433
+[2025-09-04 13:38:11] [Rank 0]   Group 2 Loss: 4.4269
+[2025-09-04 13:38:11] [Rank 0]   Group 2 Loss: 4.4269
+[2025-09-04 13:38:11] [Rank 0]   Group 3 Loss: 4.8548
+[2025-09-04 13:38:11] [Rank 0]   Group 3 Loss: 4.8548
+[2025-09-04 13:38:11] [Rank 0]   Group 4 Loss: 4.8700
+[2025-09-04 13:38:11] [Rank 0]   Group 4 Loss: 4.8700
+[2025-09-04 13:38:11] [Rank 0]   Group 5 Loss: 4.9241
+[2025-09-04 13:38:11] [Rank 0]   Group 5 Loss: 4.9241
+[2025-09-04 13:38:11] [Rank 0]   Group 6 Loss: 4.8414
+[2025-09-04 13:38:11] [Rank 0]   Group 6 Loss: 4.8414
+[2025-09-04 13:38:11] [Rank 0]   Group 7 Loss: 4.9211
+[2025-09-04 13:38:11] [Rank 0]   Group 7 Loss: 4.9211
+[2025-09-04 13:38:11] [Rank 0]   Group 8 Loss: 5.0679
+[2025-09-04 13:38:11] [Rank 0]   Group 8 Loss: 5.0679
+[2025-09-04 13:38:11] [Rank 0]   Group 9 Loss: 4.9993
+[2025-09-04 13:38:11] [Rank 0]   Group 9 Loss: 4.9993
+[2025-09-04 13:38:11] [Rank 0]   Group 10 Loss: 5.2027
+[2025-09-04 13:38:11] [Rank 0]   Group 10 Loss: 5.2027
+[2025-09-04 13:38:11] [Rank 0]   Group 11 Loss: 5.2625
+[2025-09-04 13:38:11] [Rank 0]   Group 11 Loss: 5.2625
+[2025-09-04 13:38:11] [Rank 0]   Group 12 Loss: 5.1184
+[2025-09-04 13:38:11] [Rank 0]   Group 12 Loss: 5.1184
+[2025-09-04 13:38:11] [Rank 0]   Group 13 Loss: 5.2123
+[2025-09-04 13:38:11] [Rank 0]   Group 13 Loss: 5.2123
+[2025-09-04 13:38:11] [Rank 0]   Group 14 Loss: 5.2462
+[2025-09-04 13:38:11] [Rank 0]   Group 14 Loss: 5.2462
+[2025-09-04 13:38:11] [Rank 0]   Group 15 Loss: 5.1798
+[2025-09-04 13:38:11] [Rank 0]   Group 15 Loss: 5.1798
+[2025-09-04 13:38:11] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 13 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:38:11] [Rank 0]   Group 15 FTA: 0.8800
+[2025-09-04 13:38:11] [Rank 0]   Group 15 FTA: 0.8800
+[2025-09-04 13:38:11] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:38:11] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:38:12] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:38:12] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:38:12] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:38:12] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:38:12] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:38:12] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:38:12] [Rank 0] step:9001/10000 train_time:380613ms step_avg:42.29ms
+[2025-09-04 13:38:12] [Rank 0] step:9001/10000 train_time:380613ms step_avg:42.29ms
+[2025-09-04 13:38:13] [Rank 0] step:9021/10000 train_time:381385ms step_avg:42.28ms
+[2025-09-04 13:38:13] [Rank 0] step:9021/10000 train_time:381385ms step_avg:42.28ms
+[2025-09-04 13:38:14] [Rank 0] step:9041/10000 train_time:382139ms step_avg:42.27ms
+[2025-09-04 13:38:14] [Rank 0] step:9041/10000 train_time:382139ms step_avg:42.27ms
+[2025-09-04 13:38:15] [Rank 0] step:9061/10000 train_time:382892ms step_avg:42.26ms
+[2025-09-04 13:38:15] [Rank 0] step:9061/10000 train_time:382892ms step_avg:42.26ms
+[2025-09-04 13:38:15] [Rank 0] step:9081/10000 train_time:383645ms step_avg:42.25ms
+[2025-09-04 13:38:15] [Rank 0] step:9081/10000 train_time:383645ms step_avg:42.25ms
+[2025-09-04 13:38:16] [Rank 0] step:9101/10000 train_time:384399ms step_avg:42.24ms
+[2025-09-04 13:38:16] [Rank 0] step:9101/10000 train_time:384399ms step_avg:42.24ms
+[2025-09-04 13:38:17] [Rank 0] step:9121/10000 train_time:385153ms step_avg:42.23ms
+[2025-09-04 13:38:17] [Rank 0] step:9121/10000 train_time:385153ms step_avg:42.23ms
+[2025-09-04 13:38:18] [Rank 0] step:9141/10000 train_time:385906ms step_avg:42.22ms
+[2025-09-04 13:38:18] [Rank 0] step:9141/10000 train_time:385906ms step_avg:42.22ms
+[2025-09-04 13:38:19] [Rank 0] step:9161/10000 train_time:386660ms step_avg:42.21ms
+[2025-09-04 13:38:19] [Rank 0] step:9161/10000 train_time:386660ms step_avg:42.21ms
+[2025-09-04 13:38:19] [Rank 0] step:9181/10000 train_time:387414ms step_avg:42.20ms
+[2025-09-04 13:38:19] [Rank 0] step:9181/10000 train_time:387414ms step_avg:42.20ms
+[2025-09-04 13:38:20] [Rank 0] step:9201/10000 train_time:388167ms step_avg:42.19ms
+[2025-09-04 13:38:20] [Rank 0] step:9201/10000 train_time:388167ms step_avg:42.19ms
+[2025-09-04 13:38:21] [Rank 0] step:9221/10000 train_time:389211ms step_avg:42.21ms
+[2025-09-04 13:38:21] [Rank 0] step:9221/10000 train_time:389211ms step_avg:42.21ms
+[2025-09-04 13:38:22] [Rank 0] step:9241/10000 train_time:389965ms step_avg:42.20ms
+[2025-09-04 13:38:22] [Rank 0] step:9241/10000 train_time:389965ms step_avg:42.20ms
+[2025-09-04 13:38:23] [Rank 0] step:9261/10000 train_time:390719ms step_avg:42.19ms
+[2025-09-04 13:38:23] [Rank 0] step:9261/10000 train_time:390719ms step_avg:42.19ms
+[2025-09-04 13:38:24] [Rank 0] step:9281/10000 train_time:391744ms step_avg:42.21ms
+[2025-09-04 13:38:24] [Rank 0] step:9281/10000 train_time:391744ms step_avg:42.21ms
+[2025-09-04 13:38:24] [Rank 0] step:9301/10000 train_time:392498ms step_avg:42.20ms
+[2025-09-04 13:38:24] [Rank 0] step:9301/10000 train_time:392498ms step_avg:42.20ms
+[2025-09-04 13:38:25] [Rank 0] step:9321/10000 train_time:393251ms step_avg:42.19ms
+[2025-09-04 13:38:25] [Rank 0] step:9321/10000 train_time:393251ms step_avg:42.19ms
+[2025-09-04 13:38:26] [Rank 0] step:9341/10000 train_time:394005ms step_avg:42.18ms
+[2025-09-04 13:38:26] [Rank 0] step:9341/10000 train_time:394005ms step_avg:42.18ms
+[2025-09-04 13:38:27] [Rank 0] step:9361/10000 train_time:394762ms step_avg:42.17ms
+[2025-09-04 13:38:27] [Rank 0] step:9361/10000 train_time:394762ms step_avg:42.17ms
+[2025-09-04 13:38:27] [Rank 0] step:9381/10000 train_time:395516ms step_avg:42.16ms
+[2025-09-04 13:38:27] [Rank 0] step:9381/10000 train_time:395516ms step_avg:42.16ms
+[2025-09-04 13:38:28] [Rank 0] step:9401/10000 train_time:396270ms step_avg:42.15ms
+[2025-09-04 13:38:28] [Rank 0] step:9401/10000 train_time:396270ms step_avg:42.15ms
+[2025-09-04 13:38:29] [Rank 0] step:9421/10000 train_time:397024ms step_avg:42.14ms
+[2025-09-04 13:38:29] [Rank 0] step:9421/10000 train_time:397024ms step_avg:42.14ms
+[2025-09-04 13:38:30] [Rank 0] step:9441/10000 train_time:397778ms step_avg:42.13ms
+[2025-09-04 13:38:30] [Rank 0] step:9441/10000 train_time:397778ms step_avg:42.13ms
+[2025-09-04 13:38:30] [Rank 0] step:9461/10000 train_time:398532ms step_avg:42.12ms
+[2025-09-04 13:38:30] [Rank 0] step:9461/10000 train_time:398532ms step_avg:42.12ms
+[2025-09-04 13:38:31] [Rank 0] step:9481/10000 train_time:399286ms step_avg:42.11ms
+[2025-09-04 13:38:31] [Rank 0] step:9481/10000 train_time:399286ms step_avg:42.11ms
+[2025-09-04 13:38:32] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:38:32] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:38:32] [Rank 0] PRINT: step:9500/10000 train_loss:0.6113 val_loss:0.6065 train_time:400045ms step_avg:42.11ms
+[2025-09-04 13:38:32] [Rank 0] PRINT: step:9500/10000 train_loss:0.6113 val_loss:0.6065 train_time:400045ms step_avg:42.11ms
+[2025-09-04 13:38:32] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:38:32] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:38:33] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:38:33] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:40:09] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:40:09] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:40:09] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:40:09] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:40:09] [Rank 0]   Total Loss: 4.9504
+[2025-09-04 13:40:09] [Rank 0]   Total Loss: 4.9504
+[2025-09-04 13:40:09] [Rank 0]   Total FTA (Unweighted): 0.9931
+[2025-09-04 13:40:09] [Rank 0]   Total FTA (Unweighted): 0.9931
+[2025-09-04 13:40:09] [Rank 0]   Total FTA (Weighted):   0.9931
+[2025-09-04 13:40:09] [Rank 0]   Total FTA (Weighted):   0.9931
+[2025-09-04 13:40:09] [Rank 0]   Group 0 Loss: 4.9126
+[2025-09-04 13:40:09] [Rank 0]   Group 0 Loss: 4.9126
+[2025-09-04 13:40:09] [Rank 0]   Group 1 Loss: 4.5150
+[2025-09-04 13:40:09] [Rank 0]   Group 1 Loss: 4.5150
+[2025-09-04 13:40:09] [Rank 0]   Group 2 Loss: 4.4068
+[2025-09-04 13:40:09] [Rank 0]   Group 2 Loss: 4.4068
+[2025-09-04 13:40:09] [Rank 0]   Group 3 Loss: 4.8486
+[2025-09-04 13:40:09] [Rank 0]   Group 3 Loss: 4.8486
+[2025-09-04 13:40:09] [Rank 0]   Group 4 Loss: 4.8378
+[2025-09-04 13:40:09] [Rank 0]   Group 4 Loss: 4.8378
+[2025-09-04 13:40:09] [Rank 0]   Group 5 Loss: 4.8919
+[2025-09-04 13:40:09] [Rank 0]   Group 5 Loss: 4.8919
+[2025-09-04 13:40:09] [Rank 0]   Group 6 Loss: 4.8268
+[2025-09-04 13:40:09] [Rank 0]   Group 6 Loss: 4.8268
+[2025-09-04 13:40:09] [Rank 0]   Group 7 Loss: 4.8994
+[2025-09-04 13:40:09] [Rank 0]   Group 7 Loss: 4.8994
+[2025-09-04 13:40:09] [Rank 0]   Group 8 Loss: 5.0190
+[2025-09-04 13:40:09] [Rank 0]   Group 8 Loss: 5.0190
+[2025-09-04 13:40:09] [Rank 0]   Group 9 Loss: 4.9583
+[2025-09-04 13:40:09] [Rank 0]   Group 9 Loss: 4.9583
+[2025-09-04 13:40:09] [Rank 0]   Group 10 Loss: 5.1651
+[2025-09-04 13:40:09] [Rank 0]   Group 10 Loss: 5.1651
+[2025-09-04 13:40:09] [Rank 0]   Group 11 Loss: 5.2182
+[2025-09-04 13:40:09] [Rank 0]   Group 11 Loss: 5.2182
+[2025-09-04 13:40:09] [Rank 0]   Group 12 Loss: 5.1073
+[2025-09-04 13:40:09] [Rank 0]   Group 12 Loss: 5.1073
+[2025-09-04 13:40:09] [Rank 0]   Group 13 Loss: 5.1916
+[2025-09-04 13:40:09] [Rank 0]   Group 13 Loss: 5.1916
+[2025-09-04 13:40:09] [Rank 0]   Group 14 Loss: 5.2255
+[2025-09-04 13:40:09] [Rank 0]   Group 14 Loss: 5.2255
+[2025-09-04 13:40:09] [Rank 0]   Group 15 Loss: 5.1825
+[2025-09-04 13:40:09] [Rank 0]   Group 15 Loss: 5.1825
+[2025-09-04 13:40:09] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:40:09] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:40:09] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:40:09] [Rank 0]   Group 15 FTA: 0.9000
+[2025-09-04 13:40:09] [Rank 0]   Group 15 FTA: 0.9000
+[2025-09-04 13:40:09] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:40:09] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:40:10] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:40:10] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:40:10] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:40:10] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:40:10] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:40:10] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:40:10] [Rank 0] step:9501/10000 train_time:400061ms step_avg:42.11ms
+[2025-09-04 13:40:10] [Rank 0] step:9501/10000 train_time:400061ms step_avg:42.11ms
+[2025-09-04 13:40:11] [Rank 0] step:9521/10000 train_time:400818ms step_avg:42.10ms
+[2025-09-04 13:40:11] [Rank 0] step:9521/10000 train_time:400818ms step_avg:42.10ms
+[2025-09-04 13:40:12] [Rank 0] step:9541/10000 train_time:401572ms step_avg:42.09ms
+[2025-09-04 13:40:12] [Rank 0] step:9541/10000 train_time:401572ms step_avg:42.09ms
+[2025-09-04 13:40:13] [Rank 0] step:9561/10000 train_time:402326ms step_avg:42.08ms
+[2025-09-04 13:40:13] [Rank 0] step:9561/10000 train_time:402326ms step_avg:42.08ms
+[2025-09-04 13:40:13] [Rank 0] step:9581/10000 train_time:403080ms step_avg:42.07ms
+[2025-09-04 13:40:13] [Rank 0] step:9581/10000 train_time:403080ms step_avg:42.07ms
+[2025-09-04 13:40:14] [Rank 0] step:9601/10000 train_time:403835ms step_avg:42.06ms
+[2025-09-04 13:40:14] [Rank 0] step:9601/10000 train_time:403835ms step_avg:42.06ms
+[2025-09-04 13:40:15] [Rank 0] step:9621/10000 train_time:404590ms step_avg:42.05ms
+[2025-09-04 13:40:15] [Rank 0] step:9621/10000 train_time:404590ms step_avg:42.05ms
+[2025-09-04 13:40:16] [Rank 0] step:9641/10000 train_time:405344ms step_avg:42.04ms
+[2025-09-04 13:40:16] [Rank 0] step:9641/10000 train_time:405344ms step_avg:42.04ms
+[2025-09-04 13:40:17] [Rank 0] step:9661/10000 train_time:406383ms step_avg:42.06ms
+[2025-09-04 13:40:17] [Rank 0] step:9661/10000 train_time:406383ms step_avg:42.06ms
+[2025-09-04 13:40:17] [Rank 0] step:9681/10000 train_time:407138ms step_avg:42.06ms
+[2025-09-04 13:40:17] [Rank 0] step:9681/10000 train_time:407138ms step_avg:42.06ms
+[2025-09-04 13:40:18] [Rank 0] step:9701/10000 train_time:407892ms step_avg:42.05ms
+[2025-09-04 13:40:18] [Rank 0] step:9701/10000 train_time:407892ms step_avg:42.05ms
+[2025-09-04 13:40:19] [Rank 0] step:9721/10000 train_time:408648ms step_avg:42.04ms
+[2025-09-04 13:40:19] [Rank 0] step:9721/10000 train_time:408648ms step_avg:42.04ms
+[2025-09-04 13:40:20] [Rank 0] step:9741/10000 train_time:409403ms step_avg:42.03ms
+[2025-09-04 13:40:20] [Rank 0] step:9741/10000 train_time:409403ms step_avg:42.03ms
+[2025-09-04 13:40:20] [Rank 0] step:9761/10000 train_time:410158ms step_avg:42.02ms
+[2025-09-04 13:40:20] [Rank 0] step:9761/10000 train_time:410158ms step_avg:42.02ms
+[2025-09-04 13:40:21] [Rank 0] step:9781/10000 train_time:410912ms step_avg:42.01ms
+[2025-09-04 13:40:21] [Rank 0] step:9781/10000 train_time:410912ms step_avg:42.01ms
+[2025-09-04 13:40:22] [Rank 0] step:9801/10000 train_time:411667ms step_avg:42.00ms
+[2025-09-04 13:40:22] [Rank 0] step:9801/10000 train_time:411667ms step_avg:42.00ms
+[2025-09-04 13:40:23] [Rank 0] step:9821/10000 train_time:412421ms step_avg:41.99ms
+[2025-09-04 13:40:23] [Rank 0] step:9821/10000 train_time:412421ms step_avg:41.99ms
+[2025-09-04 13:40:23] [Rank 0] step:9841/10000 train_time:413176ms step_avg:41.99ms
+[2025-09-04 13:40:23] [Rank 0] step:9841/10000 train_time:413176ms step_avg:41.99ms
+[2025-09-04 13:40:24] [Rank 0] step:9861/10000 train_time:413931ms step_avg:41.98ms
+[2025-09-04 13:40:24] [Rank 0] step:9861/10000 train_time:413931ms step_avg:41.98ms
+[2025-09-04 13:40:25] [Rank 0] step:9881/10000 train_time:414685ms step_avg:41.97ms
+[2025-09-04 13:40:25] [Rank 0] step:9881/10000 train_time:414685ms step_avg:41.97ms
+[2025-09-04 13:40:26] [Rank 0] step:9901/10000 train_time:415440ms step_avg:41.96ms
+[2025-09-04 13:40:26] [Rank 0] step:9901/10000 train_time:415440ms step_avg:41.96ms
+[2025-09-04 13:40:26] [Rank 0] step:9921/10000 train_time:416194ms step_avg:41.95ms
+[2025-09-04 13:40:26] [Rank 0] step:9921/10000 train_time:416194ms step_avg:41.95ms
+[2025-09-04 13:40:27] [Rank 0] step:9941/10000 train_time:417170ms step_avg:41.96ms
+[2025-09-04 13:40:27] [Rank 0] step:9941/10000 train_time:417170ms step_avg:41.96ms
+[2025-09-04 13:40:28] [Rank 0] step:9961/10000 train_time:417925ms step_avg:41.96ms
+[2025-09-04 13:40:28] [Rank 0] step:9961/10000 train_time:417925ms step_avg:41.96ms
+[2025-09-04 13:40:29] [Rank 0] step:9981/10000 train_time:418679ms step_avg:41.95ms
+[2025-09-04 13:40:29] [Rank 0] step:9981/10000 train_time:418679ms step_avg:41.95ms
+[2025-09-04 13:40:30] [Rank 0] step:10000/10000 train_time:419627ms step_avg:41.96ms
+[2025-09-04 13:40:30] [Rank 0] step:10000/10000 train_time:419627ms step_avg:41.96ms
+[2025-09-04 13:40:30] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:40:30] [Rank 0] PRINT: Warning: val_tokens (491520) not perfectly divisible by val_batch_size (65536). Some tokens might be missed.
+[2025-09-04 13:40:30] [Rank 0] PRINT: step:10000/10000 train_loss:0.6090 val_loss:0.6053 train_time:419675ms step_avg:41.97ms
+[2025-09-04 13:40:30] [Rank 0] PRINT: step:10000/10000 train_loss:0.6090 val_loss:0.6053 train_time:419675ms step_avg:41.97ms
+[2025-09-04 13:40:30] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:40:30] [Rank 0] 
+--- Starting Detailed Evaluation (Loss & FTA) ---
+[2025-09-04 13:40:31] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:40:31] [Rank 0] PRINT: Fixed-eval set loaded with 1600 samples.
+[2025-09-04 13:42:07] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:42:07] [Rank 0] --- Detailed Evaluation Complete ---
+[2025-09-04 13:42:07] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:42:07] [Rank 0] --- Detailed Evaluation Results (This Step) ---
+[2025-09-04 13:42:07] [Rank 0]   Total Loss: 4.9476
+[2025-09-04 13:42:07] [Rank 0]   Total Loss: 4.9476
+[2025-09-04 13:42:07] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 13:42:07] [Rank 0]   Total FTA (Unweighted): 0.9956
+[2025-09-04 13:42:07] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 13:42:07] [Rank 0]   Total FTA (Weighted):   0.9956
+[2025-09-04 13:42:07] [Rank 0]   Group 0 Loss: 4.8594
+[2025-09-04 13:42:07] [Rank 0]   Group 0 Loss: 4.8594
+[2025-09-04 13:42:07] [Rank 0]   Group 1 Loss: 4.5234
+[2025-09-04 13:42:07] [Rank 0]   Group 1 Loss: 4.5234
+[2025-09-04 13:42:07] [Rank 0]   Group 2 Loss: 4.4084
+[2025-09-04 13:42:07] [Rank 0]   Group 2 Loss: 4.4084
+[2025-09-04 13:42:07] [Rank 0]   Group 3 Loss: 4.8340
+[2025-09-04 13:42:07] [Rank 0]   Group 3 Loss: 4.8340
+[2025-09-04 13:42:07] [Rank 0]   Group 4 Loss: 4.8416
+[2025-09-04 13:42:07] [Rank 0]   Group 4 Loss: 4.8416
+[2025-09-04 13:42:07] [Rank 0]   Group 5 Loss: 4.8860
+[2025-09-04 13:42:07] [Rank 0]   Group 5 Loss: 4.8860
+[2025-09-04 13:42:07] [Rank 0]   Group 6 Loss: 4.8238
+[2025-09-04 13:42:07] [Rank 0]   Group 6 Loss: 4.8238
+[2025-09-04 13:42:07] [Rank 0]   Group 7 Loss: 4.8914
+[2025-09-04 13:42:07] [Rank 0]   Group 7 Loss: 4.8914
+[2025-09-04 13:42:07] [Rank 0]   Group 8 Loss: 5.0117
+[2025-09-04 13:42:07] [Rank 0]   Group 8 Loss: 5.0117
+[2025-09-04 13:42:07] [Rank 0]   Group 9 Loss: 4.9700
+[2025-09-04 13:42:07] [Rank 0]   Group 9 Loss: 4.9700
+[2025-09-04 13:42:07] [Rank 0]   Group 10 Loss: 5.1662
+[2025-09-04 13:42:07] [Rank 0]   Group 10 Loss: 5.1662
+[2025-09-04 13:42:07] [Rank 0]   Group 11 Loss: 5.2174
+[2025-09-04 13:42:07] [Rank 0]   Group 11 Loss: 5.2174
+[2025-09-04 13:42:07] [Rank 0]   Group 12 Loss: 5.0927
+[2025-09-04 13:42:07] [Rank 0]   Group 12 Loss: 5.0927
+[2025-09-04 13:42:07] [Rank 0]   Group 13 Loss: 5.1973
+[2025-09-04 13:42:07] [Rank 0]   Group 13 Loss: 5.1973
+[2025-09-04 13:42:07] [Rank 0]   Group 14 Loss: 5.2399
+[2025-09-04 13:42:07] [Rank 0]   Group 14 Loss: 5.2399
+[2025-09-04 13:42:07] [Rank 0]   Group 15 Loss: 5.1980
+[2025-09-04 13:42:07] [Rank 0]   Group 15 Loss: 5.1980
+[2025-09-04 13:42:07] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 0 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 1 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 2 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 3 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 4 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 5 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 6 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 7 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 8 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 9 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 10 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 11 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 12 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:42:07] [Rank 0]   Group 13 FTA: 0.9900
+[2025-09-04 13:42:07] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 14 FTA: 1.0000
+[2025-09-04 13:42:07] [Rank 0]   Group 15 FTA: 0.9400
+[2025-09-04 13:42:07] [Rank 0]   Group 15 FTA: 0.9400
+[2025-09-04 13:42:07] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:42:07] [Rank 0] [✓] Per-Class Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_loss_curves.png
+[2025-09-04 13:42:08] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:42:08] [Rank 0] [✓] Per-Class FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/per_class_acc_curves.png
+[2025-09-04 13:42:08] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:42:08] [Rank 0] [✓] Total Detailed Loss curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_loss_curve.png
+[2025-09-04 13:42:08] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:42:08] [Rank 0] [✓] Total Detailed FTA curve updated and saved to: logs_qa_muon/diff_modes/mode_10_param_qkvo_seed_46/total_acc_curve.png
+[2025-09-04 13:42:08] [Rank 0] step:10001/10000 train_time:419691ms step_avg:41.96ms
+[2025-09-04 13:42:08] [Rank 0] step:10001/10000 train_time:419691ms step_avg:41.96ms
+[2025-09-04 13:42:08] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 13:42:08 2025 ---
+[2025-09-04 13:42:08] [Rank 0] PRINT: --- Training Finished: Thu Sep  4 13:42:08 2025 ---
+[2025-09-04 13:42:08] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB
+[2025-09-04 13:42:08] [Rank 0] PRINT: Peak memory allocated: 3888 MiB reserved: 4768 MiB