inwit

config.json

@@ -0,0 +1,30 @@
+{
+  "_name_or_path": "google/t5-xl-lm-adapt",
+  "architectures": [
+    "T5ForConditionalGeneration"
+  ],
+  "d_ff": 5120,
+  "d_kv": 64,
+  "d_model": 2048,
+  "decoder_start_token_id": 0,
+  "dropout_rate": 0.1,
+  "eos_token_id": 1,
+  "feed_forward_proj": "gated-gelu",
+  "gradient_checkpointing": true,
+  "initializer_factor": 1.0,
+  "is_encoder_decoder": true,
+  "layer_norm_epsilon": 1e-06,
+  "max_length": 512,
+  "model_type": "t5",
+  "num_decoder_layers": 24,
+  "num_heads": 32,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.9.1",
+  "use_cache": false,
+  "vocab_size": 32128
+}

latest

@@ -0,0 +1 @@
+global_step1335

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bdbd622b49608641ce70036489106a835e03342686028c1686c5987fa1e00f9
+size 5699709115

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:879b3d44968be70784fb13d5e18ff98a9044e7d9d20da97ee1347d0ca11d4376
+size 14654

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ceadd58a1b326b4481eac4a5ea8dcc47f3859a8196c57df16720076976b2a9e7
+size 14654

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22412da0ce046caaade62bf68fa7a4bee9643a441b80dc0009133b3997dcc1bd
+size 14654

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54f6cf989efd1cb00d28079e567953eea428deda220c8b710a0a654777b56834
+size 14654

special_tokens_map.json

@@ -0,0 +1 @@
+{"eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>", "additional_special_tokens": ["<extra_id_0>", "<extra_id_1>", "<extra_id_2>", "<extra_id_3>", "<extra_id_4>", "<extra_id_5>", "<extra_id_6>", "<extra_id_7>", "<extra_id_8>", "<extra_id_9>", "<extra_id_10>", "<extra_id_11>", "<extra_id_12>", "<extra_id_13>", "<extra_id_14>", "<extra_id_15>", "<extra_id_16>", "<extra_id_17>", "<extra_id_18>", "<extra_id_19>", "<extra_id_20>", "<extra_id_21>", "<extra_id_22>", "<extra_id_23>", "<extra_id_24>", "<extra_id_25>", "<extra_id_26>", "<extra_id_27>", "<extra_id_28>", "<extra_id_29>", "<extra_id_30>", "<extra_id_31>", "<extra_id_32>", "<extra_id_33>", "<extra_id_34>", "<extra_id_35>", "<extra_id_36>", "<extra_id_37>", "<extra_id_38>", "<extra_id_39>", "<extra_id_40>", "<extra_id_41>", "<extra_id_42>", "<extra_id_43>", "<extra_id_44>", "<extra_id_45>", "<extra_id_46>", "<extra_id_47>", "<extra_id_48>", "<extra_id_49>", "<extra_id_50>", "<extra_id_51>", "<extra_id_52>", "<extra_id_53>", "<extra_id_54>", "<extra_id_55>", "<extra_id_56>", "<extra_id_57>", "<extra_id_58>", "<extra_id_59>", "<extra_id_60>", "<extra_id_61>", "<extra_id_62>", "<extra_id_63>", "<extra_id_64>", "<extra_id_65>", "<extra_id_66>", "<extra_id_67>", "<extra_id_68>", "<extra_id_69>", "<extra_id_70>", "<extra_id_71>", "<extra_id_72>", "<extra_id_73>", "<extra_id_74>", "<extra_id_75>", "<extra_id_76>", "<extra_id_77>", "<extra_id_78>", "<extra_id_79>", "<extra_id_80>", "<extra_id_81>", "<extra_id_82>", "<extra_id_83>", "<extra_id_84>", "<extra_id_85>", "<extra_id_86>", "<extra_id_87>", "<extra_id_88>", "<extra_id_89>", "<extra_id_90>", "<extra_id_91>", "<extra_id_92>", "<extra_id_93>", "<extra_id_94>", "<extra_id_95>", "<extra_id_96>", "<extra_id_97>", "<extra_id_98>", "<extra_id_99>"]}

tokenizer_config.json

@@ -0,0 +1 @@
+{"eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>", "extra_ids": 100, "additional_special_tokens": ["<extra_id_0>", "<extra_id_1>", "<extra_id_2>", "<extra_id_3>", "<extra_id_4>", "<extra_id_5>", "<extra_id_6>", "<extra_id_7>", "<extra_id_8>", "<extra_id_9>", "<extra_id_10>", "<extra_id_11>", "<extra_id_12>", "<extra_id_13>", "<extra_id_14>", "<extra_id_15>", "<extra_id_16>", "<extra_id_17>", "<extra_id_18>", "<extra_id_19>", "<extra_id_20>", "<extra_id_21>", "<extra_id_22>", "<extra_id_23>", "<extra_id_24>", "<extra_id_25>", "<extra_id_26>", "<extra_id_27>", "<extra_id_28>", "<extra_id_29>", "<extra_id_30>", "<extra_id_31>", "<extra_id_32>", "<extra_id_33>", "<extra_id_34>", "<extra_id_35>", "<extra_id_36>", "<extra_id_37>", "<extra_id_38>", "<extra_id_39>", "<extra_id_40>", "<extra_id_41>", "<extra_id_42>", "<extra_id_43>", "<extra_id_44>", "<extra_id_45>", "<extra_id_46>", "<extra_id_47>", "<extra_id_48>", "<extra_id_49>", "<extra_id_50>", "<extra_id_51>", "<extra_id_52>", "<extra_id_53>", "<extra_id_54>", "<extra_id_55>", "<extra_id_56>", "<extra_id_57>", "<extra_id_58>", "<extra_id_59>", "<extra_id_60>", "<extra_id_61>", "<extra_id_62>", "<extra_id_63>", "<extra_id_64>", "<extra_id_65>", "<extra_id_66>", "<extra_id_67>", "<extra_id_68>", "<extra_id_69>", "<extra_id_70>", "<extra_id_71>", "<extra_id_72>", "<extra_id_73>", "<extra_id_74>", "<extra_id_75>", "<extra_id_76>", "<extra_id_77>", "<extra_id_78>", "<extra_id_79>", "<extra_id_80>", "<extra_id_81>", "<extra_id_82>", "<extra_id_83>", "<extra_id_84>", "<extra_id_85>", "<extra_id_86>", "<extra_id_87>", "<extra_id_88>", "<extra_id_89>", "<extra_id_90>", "<extra_id_91>", "<extra_id_92>", "<extra_id_93>", "<extra_id_94>", "<extra_id_95>", "<extra_id_96>", "<extra_id_97>", "<extra_id_98>", "<extra_id_99>"], "sp_model_kwargs": {}, "model_max_length": 512, "name_or_path": "google/t5-xl-lm-adapt", "special_tokens_map_file": "/home/patrick/.cache/huggingface/transformers/e88f2448cc299b3d5844700ccb67f86e37caa0873ebab334ad4e881fd84f1abf.c94798918c92ded6aeef2d2f0e666d2cc4145eca1aa6e1336fde07f2e13e2f46", "tokenizer_class": "T5Tokenizer"}

trainer_state.json

@@ -0,0 +1,944 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 106.63897763578275,
+  "global_step": 1280,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.8,
+      "learning_rate": 9.99969999399988e-05,
+      "loss": 3.2772,
+      "step": 10
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 9.99869997399948e-05,
+      "loss": 0.587,
+      "step": 20
+    },
+    {
+      "epoch": 2.48,
+      "learning_rate": 9.99759995199904e-05,
+      "loss": 0.2432,
+      "step": 30
+    },
+    {
+      "epoch": 3.32,
+      "learning_rate": 9.99659993199864e-05,
+      "loss": 0.1511,
+      "step": 40
+    },
+    {
+      "epoch": 4.16,
+      "learning_rate": 9.9954999099982e-05,
+      "loss": 0.1104,
+      "step": 50
+    },
+    {
+      "epoch": 4.96,
+      "learning_rate": 9.994499889997801e-05,
+      "loss": 0.0839,
+      "step": 60
+    },
+    {
+      "epoch": 5.32,
+      "eval_loss": 0.116455078125,
+      "eval_runtime": 85.2162,
+      "eval_samples_per_second": 38.044,
+      "eval_steps_per_second": 0.481,
+      "step": 64
+    },
+    {
+      "epoch": 5.8,
+      "learning_rate": 9.9934998699974e-05,
+      "loss": 0.0716,
+      "step": 70
+    },
+    {
+      "epoch": 6.64,
+      "learning_rate": 9.99239984799696e-05,
+      "loss": 0.0665,
+      "step": 80
+    },
+    {
+      "epoch": 7.48,
+      "learning_rate": 9.99139982799656e-05,
+      "loss": 0.057,
+      "step": 90
+    },
+    {
+      "epoch": 8.32,
+      "learning_rate": 9.990299805996121e-05,
+      "loss": 0.0521,
+      "step": 100
+    },
+    {
+      "epoch": 9.16,
+      "learning_rate": 9.98929978599572e-05,
+      "loss": 0.0466,
+      "step": 110
+    },
+    {
+      "epoch": 9.96,
+      "learning_rate": 9.98829976599532e-05,
+      "loss": 0.0417,
+      "step": 120
+    },
+    {
+      "epoch": 10.64,
+      "eval_loss": 0.12420654296875,
+      "eval_runtime": 85.7837,
+      "eval_samples_per_second": 37.793,
+      "eval_steps_per_second": 0.478,
+      "step": 128
+    },
+    {
+      "epoch": 10.8,
+      "learning_rate": 9.987199743994881e-05,
+      "loss": 0.0388,
+      "step": 130
+    },
+    {
+      "epoch": 11.64,
+      "learning_rate": 9.986199723994481e-05,
+      "loss": 0.0371,
+      "step": 140
+    },
+    {
+      "epoch": 12.48,
+      "learning_rate": 9.98509970199404e-05,
+      "loss": 0.0342,
+      "step": 150
+    },
+    {
+      "epoch": 13.32,
+      "learning_rate": 9.98409968199364e-05,
+      "loss": 0.0303,
+      "step": 160
+    },
+    {
+      "epoch": 14.16,
+      "learning_rate": 9.982999659993201e-05,
+      "loss": 0.0287,
+      "step": 170
+    },
+    {
+      "epoch": 14.96,
+      "learning_rate": 9.981999639992801e-05,
+      "loss": 0.0262,
+      "step": 180
+    },
+    {
+      "epoch": 15.8,
+      "learning_rate": 9.9809996199924e-05,
+      "loss": 0.0242,
+      "step": 190
+    },
+    {
+      "epoch": 15.96,
+      "eval_loss": 0.13671875,
+      "eval_runtime": 71.3468,
+      "eval_samples_per_second": 45.44,
+      "eval_steps_per_second": 0.575,
+      "step": 192
+    },
+    {
+      "epoch": 16.64,
+      "learning_rate": 9.979899597991961e-05,
+      "loss": 0.0239,
+      "step": 200
+    },
+    {
+      "epoch": 17.48,
+      "learning_rate": 9.978899577991561e-05,
+      "loss": 0.0216,
+      "step": 210
+    },
+    {
+      "epoch": 18.32,
+      "learning_rate": 9.97779955599112e-05,
+      "loss": 0.0204,
+      "step": 220
+    },
+    {
+      "epoch": 19.16,
+      "learning_rate": 9.97679953599072e-05,
+      "loss": 0.0194,
+      "step": 230
+    },
+    {
+      "epoch": 19.96,
+      "learning_rate": 9.97579951599032e-05,
+      "loss": 0.0167,
+      "step": 240
+    },
+    {
+      "epoch": 20.8,
+      "learning_rate": 9.974699493989881e-05,
+      "loss": 0.0177,
+      "step": 250
+    },
+    {
+      "epoch": 21.32,
+      "eval_loss": 0.15576171875,
+      "eval_runtime": 73.465,
+      "eval_samples_per_second": 44.13,
+      "eval_steps_per_second": 0.558,
+      "step": 256
+    },
+    {
+      "epoch": 21.64,
+      "learning_rate": 9.973699473989481e-05,
+      "loss": 0.0155,
+      "step": 260
+    },
+    {
+      "epoch": 22.48,
+      "learning_rate": 9.97259945198904e-05,
+      "loss": 0.015,
+      "step": 270
+    },
+    {
+      "epoch": 23.32,
+      "learning_rate": 9.971599431988641e-05,
+      "loss": 0.0148,
+      "step": 280
+    },
+    {
+      "epoch": 24.16,
+      "learning_rate": 9.9704994099882e-05,
+      "loss": 0.0137,
+      "step": 290
+    },
+    {
+      "epoch": 24.96,
+      "learning_rate": 9.969499389987799e-05,
+      "loss": 0.0117,
+      "step": 300
+    },
+    {
+      "epoch": 25.8,
+      "learning_rate": 9.96839936798736e-05,
+      "loss": 0.0118,
+      "step": 310
+    },
+    {
+      "epoch": 26.64,
+      "learning_rate": 9.96739934798696e-05,
+      "loss": 0.0111,
+      "step": 320
+    },
+    {
+      "epoch": 26.64,
+      "eval_loss": 0.1729736328125,
+      "eval_runtime": 73.6,
+      "eval_samples_per_second": 44.049,
+      "eval_steps_per_second": 0.557,
+      "step": 320
+    },
+    {
+      "epoch": 27.48,
+      "learning_rate": 9.96629932598652e-05,
+      "loss": 0.0109,
+      "step": 330
+    },
+    {
+      "epoch": 28.32,
+      "learning_rate": 9.96529930598612e-05,
+      "loss": 0.0102,
+      "step": 340
+    },
+    {
+      "epoch": 29.16,
+      "learning_rate": 9.96419928398568e-05,
+      "loss": 0.0094,
+      "step": 350
+    },
+    {
+      "epoch": 29.96,
+      "learning_rate": 9.96319926398528e-05,
+      "loss": 0.0088,
+      "step": 360
+    },
+    {
+      "epoch": 30.8,
+      "learning_rate": 9.96219924398488e-05,
+      "loss": 0.0084,
+      "step": 370
+    },
+    {
+      "epoch": 31.64,
+      "learning_rate": 9.96109922198444e-05,
+      "loss": 0.0082,
+      "step": 380
+    },
+    {
+      "epoch": 31.96,
+      "eval_loss": 0.1884765625,
+      "eval_runtime": 72.563,
+      "eval_samples_per_second": 44.678,
+      "eval_steps_per_second": 0.565,
+      "step": 384
+    },
+    {
+      "epoch": 32.48,
+      "learning_rate": 9.96009920198404e-05,
+      "loss": 0.0082,
+      "step": 390
+    },
+    {
+      "epoch": 33.32,
+      "learning_rate": 9.9589991799836e-05,
+      "loss": 0.0078,
+      "step": 400
+    },
+    {
+      "epoch": 34.16,
+      "learning_rate": 9.9579991599832e-05,
+      "loss": 0.0068,
+      "step": 410
+    },
+    {
+      "epoch": 34.96,
+      "learning_rate": 9.9569991399828e-05,
+      "loss": 0.0063,
+      "step": 420
+    },
+    {
+      "epoch": 35.8,
+      "learning_rate": 9.95589911798236e-05,
+      "loss": 0.0061,
+      "step": 430
+    },
+    {
+      "epoch": 36.64,
+      "learning_rate": 9.95489909798196e-05,
+      "loss": 0.0064,
+      "step": 440
+    },
+    {
+      "epoch": 37.32,
+      "eval_loss": 0.191162109375,
+      "eval_runtime": 72.3445,
+      "eval_samples_per_second": 44.813,
+      "eval_steps_per_second": 0.567,
+      "step": 448
+    },
+    {
+      "epoch": 37.48,
+      "learning_rate": 9.95379907598152e-05,
+      "loss": 0.0058,
+      "step": 450
+    },
+    {
+      "epoch": 38.32,
+      "learning_rate": 9.95279905598112e-05,
+      "loss": 0.0051,
+      "step": 460
+    },
+    {
+      "epoch": 39.16,
+      "learning_rate": 9.95169903398068e-05,
+      "loss": 0.0053,
+      "step": 470
+    },
+    {
+      "epoch": 39.96,
+      "learning_rate": 9.950699013980281e-05,
+      "loss": 0.0049,
+      "step": 480
+    },
+    {
+      "epoch": 40.8,
+      "learning_rate": 9.94969899397988e-05,
+      "loss": 0.0048,
+      "step": 490
+    },
+    {
+      "epoch": 41.64,
+      "learning_rate": 9.94859897197944e-05,
+      "loss": 0.0047,
+      "step": 500
+    },
+    {
+      "epoch": 42.48,
+      "learning_rate": 9.94759895197904e-05,
+      "loss": 0.0047,
+      "step": 510
+    },
+    {
+      "epoch": 42.64,
+      "eval_loss": 0.195556640625,
+      "eval_runtime": 71.2606,
+      "eval_samples_per_second": 45.495,
+      "eval_steps_per_second": 0.575,
+      "step": 512
+    },
+    {
+      "epoch": 43.32,
+      "learning_rate": 9.946498929978601e-05,
+      "loss": 0.0044,
+      "step": 520
+    },
+    {
+      "epoch": 44.16,
+      "learning_rate": 9.9454989099782e-05,
+      "loss": 0.0044,
+      "step": 530
+    },
+    {
+      "epoch": 44.96,
+      "learning_rate": 9.9444988899778e-05,
+      "loss": 0.0039,
+      "step": 540
+    },
+    {
+      "epoch": 45.8,
+      "learning_rate": 9.943398867977361e-05,
+      "loss": 0.0036,
+      "step": 550
+    },
+    {
+      "epoch": 46.64,
+      "learning_rate": 9.942398847976961e-05,
+      "loss": 0.0037,
+      "step": 560
+    },
+    {
+      "epoch": 47.48,
+      "learning_rate": 9.94129882597652e-05,
+      "loss": 0.0034,
+      "step": 570
+    },
+    {
+      "epoch": 47.96,
+      "eval_loss": 0.2215576171875,
+      "eval_runtime": 71.3132,
+      "eval_samples_per_second": 45.461,
+      "eval_steps_per_second": 0.575,
+      "step": 576
+    },
+    {
+      "epoch": 48.32,
+      "learning_rate": 9.94029880597612e-05,
+      "loss": 0.0032,
+      "step": 580
+    },
+    {
+      "epoch": 49.16,
+      "learning_rate": 9.93919878397568e-05,
+      "loss": 0.0035,
+      "step": 590
+    },
+    {
+      "epoch": 49.96,
+      "learning_rate": 9.938198763975281e-05,
+      "loss": 0.0032,
+      "step": 600
+    },
+    {
+      "epoch": 50.8,
+      "learning_rate": 9.93709874197484e-05,
+      "loss": 0.003,
+      "step": 610
+    },
+    {
+      "epoch": 51.64,
+      "learning_rate": 9.93609872197444e-05,
+      "loss": 0.0031,
+      "step": 620
+    },
+    {
+      "epoch": 52.48,
+      "learning_rate": 9.934998699974e-05,
+      "loss": 0.0029,
+      "step": 630
+    },
+    {
+      "epoch": 53.32,
+      "learning_rate": 9.933998679973599e-05,
+      "loss": 0.0029,
+      "step": 640
+    },
+    {
+      "epoch": 53.32,
+      "eval_loss": 0.221923828125,
+      "eval_runtime": 71.4852,
+      "eval_samples_per_second": 45.352,
+      "eval_steps_per_second": 0.574,
+      "step": 640
+    },
+    {
+      "epoch": 54.16,
+      "learning_rate": 9.93289865797316e-05,
+      "loss": 0.0028,
+      "step": 650
+    },
+    {
+      "epoch": 54.96,
+      "learning_rate": 9.93189863797276e-05,
+      "loss": 0.0027,
+      "step": 660
+    },
+    {
+      "epoch": 55.8,
+      "learning_rate": 9.93089861797236e-05,
+      "loss": 0.0026,
+      "step": 670
+    },
+    {
+      "epoch": 56.64,
+      "learning_rate": 9.929798595971919e-05,
+      "loss": 0.0022,
+      "step": 680
+    },
+    {
+      "epoch": 57.48,
+      "learning_rate": 9.92879857597152e-05,
+      "loss": 0.0022,
+      "step": 690
+    },
+    {
+      "epoch": 58.32,
+      "learning_rate": 9.92769855397108e-05,
+      "loss": 0.0022,
+      "step": 700
+    },
+    {
+      "epoch": 58.64,
+      "eval_loss": 0.2276611328125,
+      "eval_runtime": 71.4916,
+      "eval_samples_per_second": 45.348,
+      "eval_steps_per_second": 0.573,
+      "step": 704
+    },
+    {
+      "epoch": 59.16,
+      "learning_rate": 9.92669853397068e-05,
+      "loss": 0.0025,
+      "step": 710
+    },
+    {
+      "epoch": 59.96,
+      "learning_rate": 9.925698513970279e-05,
+      "loss": 0.0021,
+      "step": 720
+    },
+    {
+      "epoch": 60.8,
+      "learning_rate": 9.92459849196984e-05,
+      "loss": 0.0024,
+      "step": 730
+    },
+    {
+      "epoch": 61.64,
+      "learning_rate": 9.92359847196944e-05,
+      "loss": 0.0022,
+      "step": 740
+    },
+    {
+      "epoch": 62.48,
+      "learning_rate": 9.922498449968999e-05,
+      "loss": 0.0018,
+      "step": 750
+    },
+    {
+      "epoch": 63.32,
+      "learning_rate": 9.9214984299686e-05,
+      "loss": 0.0017,
+      "step": 760
+    },
+    {
+      "epoch": 63.96,
+      "eval_loss": 0.24072265625,
+      "eval_runtime": 72.0268,
+      "eval_samples_per_second": 45.011,
+      "eval_steps_per_second": 0.569,
+      "step": 768
+    },
+    {
+      "epoch": 64.16,
+      "learning_rate": 9.92039840796816e-05,
+      "loss": 0.0021,
+      "step": 770
+    },
+    {
+      "epoch": 64.96,
+      "learning_rate": 9.91939838796776e-05,
+      "loss": 0.0017,
+      "step": 780
+    },
+    {
+      "epoch": 65.8,
+      "learning_rate": 9.91839836796736e-05,
+      "loss": 0.0022,
+      "step": 790
+    },
+    {
+      "epoch": 66.64,
+      "learning_rate": 9.91729834596692e-05,
+      "loss": 0.0019,
+      "step": 800
+    },
+    {
+      "epoch": 67.48,
+      "learning_rate": 9.91629832596652e-05,
+      "loss": 0.0017,
+      "step": 810
+    },
+    {
+      "epoch": 68.32,
+      "learning_rate": 9.91519830396608e-05,
+      "loss": 0.0017,
+      "step": 820
+    },
+    {
+      "epoch": 69.16,
+      "learning_rate": 9.91419828396568e-05,
+      "loss": 0.0018,
+      "step": 830
+    },
+    {
+      "epoch": 69.32,
+      "eval_loss": 0.24853515625,
+      "eval_runtime": 73.1565,
+      "eval_samples_per_second": 44.316,
+      "eval_steps_per_second": 0.56,
+      "step": 832
+    },
+    {
+      "epoch": 69.96,
+      "learning_rate": 9.91319826396528e-05,
+      "loss": 0.0018,
+      "step": 840
+    },
+    {
+      "epoch": 70.8,
+      "learning_rate": 9.91209824196484e-05,
+      "loss": 0.0016,
+      "step": 850
+    },
+    {
+      "epoch": 71.64,
+      "learning_rate": 9.91109822196444e-05,
+      "loss": 0.0015,
+      "step": 860
+    },
+    {
+      "epoch": 72.48,
+      "learning_rate": 9.909998199964e-05,
+      "loss": 0.0016,
+      "step": 870
+    },
+    {
+      "epoch": 73.32,
+      "learning_rate": 9.9089981799636e-05,
+      "loss": 0.0016,
+      "step": 880
+    },
+    {
+      "epoch": 74.16,
+      "learning_rate": 9.90789815796316e-05,
+      "loss": 0.0015,
+      "step": 890
+    },
+    {
+      "epoch": 74.64,
+      "eval_loss": 0.23681640625,
+      "eval_runtime": 71.4778,
+      "eval_samples_per_second": 45.357,
+      "eval_steps_per_second": 0.574,
+      "step": 896
+    },
+    {
+      "epoch": 74.96,
+      "learning_rate": 9.90689813796276e-05,
+      "loss": 0.0016,
+      "step": 900
+    },
+    {
+      "epoch": 75.8,
+      "learning_rate": 9.905798115962319e-05,
+      "loss": 0.0016,
+      "step": 910
+    },
+    {
+      "epoch": 76.64,
+      "learning_rate": 9.90479809596192e-05,
+      "loss": 0.0016,
+      "step": 920
+    },
+    {
+      "epoch": 77.48,
+      "learning_rate": 9.90369807396148e-05,
+      "loss": 0.0014,
+      "step": 930
+    },
+    {
+      "epoch": 78.32,
+      "learning_rate": 9.90269805396108e-05,
+      "loss": 0.0011,
+      "step": 940
+    },
+    {
+      "epoch": 79.16,
+      "learning_rate": 9.901598031960639e-05,
+      "loss": 0.0015,
+      "step": 950
+    },
+    {
+      "epoch": 79.96,
+      "learning_rate": 9.900598011960239e-05,
+      "loss": 0.0012,
+      "step": 960
+    },
+    {
+      "epoch": 79.96,
+      "eval_loss": 0.254150390625,
+      "eval_runtime": 76.7721,
+      "eval_samples_per_second": 42.229,
+      "eval_steps_per_second": 0.534,
+      "step": 960
+    },
+    {
+      "epoch": 80.8,
+      "learning_rate": 9.89959799195984e-05,
+      "loss": 0.0013,
+      "step": 970
+    },
+    {
+      "epoch": 81.64,
+      "learning_rate": 9.898497969959399e-05,
+      "loss": 0.0014,
+      "step": 980
+    },
+    {
+      "epoch": 82.48,
+      "learning_rate": 9.897497949958999e-05,
+      "loss": 0.0015,
+      "step": 990
+    },
+    {
+      "epoch": 83.32,
+      "learning_rate": 9.89639792795856e-05,
+      "loss": 0.0013,
+      "step": 1000
+    },
+    {
+      "epoch": 84.16,
+      "learning_rate": 9.89539790795816e-05,
+      "loss": 0.0014,
+      "step": 1010
+    },
+    {
+      "epoch": 84.96,
+      "learning_rate": 9.894397887957759e-05,
+      "loss": 0.0012,
+      "step": 1020
+    },
+    {
+      "epoch": 85.32,
+      "eval_loss": 0.253662109375,
+      "eval_runtime": 76.0139,
+      "eval_samples_per_second": 42.65,
+      "eval_steps_per_second": 0.539,
+      "step": 1024
+    },
+    {
+      "epoch": 85.8,
+      "learning_rate": 9.893297865957319e-05,
+      "loss": 0.0014,
+      "step": 1030
+    },
+    {
+      "epoch": 86.64,
+      "learning_rate": 9.89229784595692e-05,
+      "loss": 0.0011,
+      "step": 1040
+    },
+    {
+      "epoch": 87.48,
+      "learning_rate": 9.89119782395648e-05,
+      "loss": 0.0012,
+      "step": 1050
+    },
+    {
+      "epoch": 88.32,
+      "learning_rate": 9.890197803956079e-05,
+      "loss": 0.0012,
+      "step": 1060
+    },
+    {
+      "epoch": 89.16,
+      "learning_rate": 9.88909778195564e-05,
+      "loss": 0.0013,
+      "step": 1070
+    },
+    {
+      "epoch": 89.96,
+      "learning_rate": 9.88809776195524e-05,
+      "loss": 0.0011,
+      "step": 1080
+    },
+    {
+      "epoch": 90.64,
+      "eval_loss": 0.268798828125,
+      "eval_runtime": 75.8685,
+      "eval_samples_per_second": 42.732,
+      "eval_steps_per_second": 0.54,
+      "step": 1088
+    },
+    {
+      "epoch": 90.8,
+      "learning_rate": 9.88709774195484e-05,
+      "loss": 0.0011,
+      "step": 1090
+    },
+    {
+      "epoch": 91.64,
+      "learning_rate": 9.885997719954399e-05,
+      "loss": 0.0013,
+      "step": 1100
+    },
+    {
+      "epoch": 92.48,
+      "learning_rate": 9.884997699954e-05,
+      "loss": 0.001,
+      "step": 1110
+    },
+    {
+      "epoch": 93.32,
+      "learning_rate": 9.88389767795356e-05,
+      "loss": 0.0011,
+      "step": 1120
+    },
+    {
+      "epoch": 94.16,
+      "learning_rate": 9.88289765795316e-05,
+      "loss": 0.0013,
+      "step": 1130
+    },
+    {
+      "epoch": 94.96,
+      "learning_rate": 9.881897637952759e-05,
+      "loss": 0.001,
+      "step": 1140
+    },
+    {
+      "epoch": 95.8,
+      "learning_rate": 9.88079761595232e-05,
+      "loss": 0.0016,
+      "step": 1150
+    },
+    {
+      "epoch": 95.96,
+      "eval_loss": 0.24560546875,
+      "eval_runtime": 75.5498,
+      "eval_samples_per_second": 42.912,
+      "eval_steps_per_second": 0.543,
+      "step": 1152
+    },
+    {
+      "epoch": 96.64,
+      "learning_rate": 9.87979759595192e-05,
+      "loss": 0.0011,
+      "step": 1160
+    },
+    {
+      "epoch": 97.48,
+      "learning_rate": 9.878697573951479e-05,
+      "loss": 0.0011,
+      "step": 1170
+    },
+    {
+      "epoch": 98.32,
+      "learning_rate": 9.87769755395108e-05,
+      "loss": 0.0013,
+      "step": 1180
+    },
+    {
+      "epoch": 99.16,
+      "learning_rate": 9.87659753195064e-05,
+      "loss": 0.0011,
+      "step": 1190
+    },
+    {
+      "epoch": 99.96,
+      "learning_rate": 9.87559751195024e-05,
+      "loss": 0.0011,
+      "step": 1200
+    },
+    {
+      "epoch": 100.8,
+      "learning_rate": 9.8744974899498e-05,
+      "loss": 0.0008,
+      "step": 1210
+    },
+    {
+      "epoch": 101.32,
+      "eval_loss": 0.278076171875,
+      "eval_runtime": 74.8396,
+      "eval_samples_per_second": 43.319,
+      "eval_steps_per_second": 0.548,
+      "step": 1216
+    },
+    {
+      "epoch": 101.64,
+      "learning_rate": 9.8734974699494e-05,
+      "loss": 0.001,
+      "step": 1220
+    },
+    {
+      "epoch": 102.48,
+      "learning_rate": 9.872397447948958e-05,
+      "loss": 0.0008,
+      "step": 1230
+    },
+    {
+      "epoch": 103.32,
+      "learning_rate": 9.87139742794856e-05,
+      "loss": 0.0009,
+      "step": 1240
+    },
+    {
+      "epoch": 104.16,
+      "learning_rate": 9.870297405948119e-05,
+      "loss": 0.0008,
+      "step": 1250
+    },
+    {
+      "epoch": 104.96,
+      "learning_rate": 9.869297385947719e-05,
+      "loss": 0.0008,
+      "step": 1260
+    },
+    {
+      "epoch": 105.8,
+      "learning_rate": 9.86829736594732e-05,
+      "loss": 0.0011,
+      "step": 1270
+    },
+    {
+      "epoch": 106.64,
+      "learning_rate": 9.86719734394688e-05,
+      "loss": 0.0009,
+      "step": 1280
+    },
+    {
+      "epoch": 106.64,
+      "eval_loss": 0.2666015625,
+      "eval_runtime": 75.7897,
+      "eval_samples_per_second": 42.776,
+      "eval_steps_per_second": 0.541,
+      "step": 1280
+    }
+  ],
+  "max_steps": 100000,
+  "num_train_epochs": 8334,
+  "total_flos": 6810132464640.0,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1063828ba2e9ddcd06c922117754b55edeba8f0e5dbe1457111b25089b27d900
+size 4207

zero_to_fp32.py

@@ -0,0 +1,348 @@
+#!/usr/bin/env python
+
+# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import os
+from collections import OrderedDict
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+import deepspeed
+from deepspeed.utils import logger
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage == 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_optim_files(checkpoint_dir):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    optim_files = sorted(glob.glob(os.path.join(checkpoint_dir, "*_optim_states.pt")))
+
+    if len(optim_files) == 0:
+        raise FileNotFoundError(
+            f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
+
+    return optim_files
+
+
+def parse_model_state(file):
+    state_dict = torch.load(file, map_location=device)
+
+    if "buffer_names" not in state_dict:
+        raise ValueError(f"{file} is not a model state checkpoint")
+    buffer_names = state_dict["buffer_names"]
+    if debug:
+        print("Found buffers:", buffer_names)
+
+    # recover just the buffers while restoring them to fp32 if they were saved in fp16
+    buffers = {
+        k: v.float()
+        for k,
+        v in state_dict["module"].items() if k in buffer_names
+    }
+    return buffers
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    if not "zero_stage" in state_dicts[0]['optimizer_state_dict']:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0]['optimizer_state_dict']["zero_stage"]
+    world_size = state_dicts[0]['optimizer_state_dict']["partition_count"]
+    param_shapes = state_dicts[0]["param_shapes"]
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage == 2:
+        fp32_groups_key = "single_partition_of_fp32_groups"
+    elif zero_stage == 3:
+        fp32_groups_key = "fp32_flat_groups"
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    # if there is more than one param group, there will be multiple flattened tensors - one
+    # flattened tensor per group - for simplicity merge them into a single tensor
+    #
+    # XXX: could make the script more memory efficient for when there are multiple groups - it
+    # will require matching the sub-lists of param_shapes for each param group flattened tensor
+    fp32_flat_groups = [
+        torch.cat(state_dicts[i]['optimizer_state_dict'][fp32_groups_key],
+                  0) for i in range(len(state_dicts))
+    ]
+
+    return zero_stage, world_size, param_shapes, fp32_flat_groups
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = int(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, param_shapes, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(
+        f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
+    buffers = parse_model_state(model_file)
+
+    # Reconstruction protocol:
+    #
+    # - for zero2 we just need to concat the partitions back to back and reconsolidate over one huge
+    # flat buffer - no need to deal with padding since if there is any it will be only in the tail
+    # of the last partition so there it will be just left out
+    #
+    # - for zero3 we need to zip the partitions together at boundary of each param, re-consolidating
+    # each param, while dealing with padding if any
+
+    if debug:
+        for i in range(world_size):
+            print(f"fp32_flat_groups[i].shape={fp32_flat_groups[i].shape}")
+
+    if zero_stage == 2:
+        # XXX: memory usage doubles here (zero2)
+        full_single_fp32_vector = torch.cat(fp32_flat_groups, 0)
+        avail_numel = full_single_fp32_vector.numel()
+    elif zero_stage == 3:
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+
+    if debug:
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        if zero_stage == 2:
+            if debug:
+                print(
+                    f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
+                )
+            state_dict[name] = full_single_fp32_vector.narrow(
+                0,
+                offset,
+                unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        elif zero_stage == 3:
+            partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+            if debug:
+                print(
+                    f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+                )
+
+            # XXX: memory usage doubles here (zero3)
+            state_dict[name] = torch.cat(
+                tuple(fp32_flat_groups[i].narrow(0,
+                                                 offset,
+                                                 partitioned_numel)
+                      for i in range(world_size)),
+                0).view(shape)
+            offset += partitioned_numel + partitioned_padding_numel
+
+    if zero_stage == 3:
+        offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(
+            f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(
+        f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be useable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be useable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "checkpoint_dir",
+        type=str,
+        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help=
+        "path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
+    )
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)