init

config.json

@@ -0,0 +1,30 @@
+{
+  "_name_or_path": "google/t5-large-lm-adapt",
+  "architectures": [
+    "T5ForConditionalGeneration"
+  ],
+  "d_ff": 2816,
+  "d_kv": 64,
+  "d_model": 1024,
+  "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": 16,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.9.1",
+  "use_cache": false,
+  "vocab_size": 32128
+}

latest

@@ -0,0 +1 @@
+global_step2000

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e37ea3374e90847d2f52fbeec66e4189afb8961c582f597ed237861a1d68406b
+size 3132793282

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33d4e867339985a01d3551db4e58b19992103c762b3ebb145698ca499afe8886
+size 14649

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07b9b533e43a86fd8d75da6e6dc2516c5a274ae1f30b985b8c6510bfe0cd7638
+size 14654

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e404ce3b064e6bb412f0226169d0deaf72d60d8d1935994e0e97d8bbfdbaebfd
+size 14654

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b9d1476a94159f17356260ee7fb2b4f98d4b6ec20ead888948f164e364a97eb
+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-large-lm-adapt", "special_tokens_map_file": "/home/patrick/.cache/huggingface/transformers/28fa6bb11d5fd637c4b67b299f5616169510dac4cd181efc8f70674c7872c874.c94798918c92ded6aeef2d2f0e666d2cc4145eca1aa6e1336fde07f2e13e2f46", "tokenizer_class": "T5Tokenizer"}

trainer_state.json

@@ -0,0 +1,1248 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 32.78688524590164,
+  "global_step": 2000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.16,
+      "learning_rate": 2.4999999999999998e-05,
+      "loss": 4.5271,
+      "step": 10
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 3.2525749891599525e-05,
+      "loss": 3.9027,
+      "step": 20
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 3.6928031367991554e-05,
+      "loss": 3.4367,
+      "step": 30
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 4.005149978319905e-05,
+      "loss": 3.3343,
+      "step": 40
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 4.247425010840046e-05,
+      "loss": 3.2186,
+      "step": 50
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 4.445378125959108e-05,
+      "loss": 3.1259,
+      "step": 60
+    },
+    {
+      "epoch": 1.15,
+      "learning_rate": 4.612745100035642e-05,
+      "loss": 2.9406,
+      "step": 70
+    },
+    {
+      "epoch": 1.31,
+      "learning_rate": 4.757724967479858e-05,
+      "loss": 2.8567,
+      "step": 80
+    },
+    {
+      "epoch": 1.48,
+      "learning_rate": 4.885606273598312e-05,
+      "loss": 2.8893,
+      "step": 90
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 4.9999999999999996e-05,
+      "loss": 2.9067,
+      "step": 100
+    },
+    {
+      "epoch": 1.8,
+      "learning_rate": 5.1034817128955624e-05,
+      "loss": 2.813,
+      "step": 110
+    },
+    {
+      "epoch": 1.97,
+      "learning_rate": 5.197953115119061e-05,
+      "loss": 2.8364,
+      "step": 120
+    },
+    {
+      "epoch": 2.13,
+      "learning_rate": 5.2848583807670914e-05,
+      "loss": 2.7439,
+      "step": 130
+    },
+    {
+      "epoch": 2.3,
+      "learning_rate": 5.365320089195593e-05,
+      "loss": 2.7048,
+      "step": 140
+    },
+    {
+      "epoch": 2.46,
+      "learning_rate": 5.4402281476392025e-05,
+      "loss": 2.6631,
+      "step": 150
+    },
+    {
+      "epoch": 2.62,
+      "learning_rate": 5.5102999566398106e-05,
+      "loss": 2.6804,
+      "step": 160
+    },
+    {
+      "epoch": 2.79,
+      "learning_rate": 5.5761223034456847e-05,
+      "loss": 2.6094,
+      "step": 170
+    },
+    {
+      "epoch": 2.95,
+      "learning_rate": 5.6381812627582644e-05,
+      "loss": 2.6055,
+      "step": 180
+    },
+    {
+      "epoch": 3.11,
+      "learning_rate": 5.696884002382071e-05,
+      "loss": 2.4945,
+      "step": 190
+    },
+    {
+      "epoch": 3.28,
+      "learning_rate": 5.752574989159952e-05,
+      "loss": 2.5523,
+      "step": 200
+    },
+    {
+      "epoch": 3.44,
+      "learning_rate": 5.805548236834797e-05,
+      "loss": 2.5406,
+      "step": 210
+    },
+    {
+      "epoch": 3.61,
+      "learning_rate": 5.856056702055516e-05,
+      "loss": 2.5444,
+      "step": 220
+    },
+    {
+      "epoch": 3.77,
+      "learning_rate": 5.9043195900439815e-05,
+      "loss": 2.568,
+      "step": 230
+    },
+    {
+      "epoch": 3.93,
+      "learning_rate": 5.950528104279014e-05,
+      "loss": 2.5305,
+      "step": 240
+    },
+    {
+      "epoch": 4.1,
+      "learning_rate": 5.9948500216800926e-05,
+      "loss": 2.4653,
+      "step": 250
+    },
+    {
+      "epoch": 4.26,
+      "learning_rate": 6.037433369927045e-05,
+      "loss": 2.4644,
+      "step": 260
+    },
+    {
+      "epoch": 4.43,
+      "learning_rate": 6.078409410397467e-05,
+      "loss": 2.4302,
+      "step": 270
+    },
+    {
+      "epoch": 4.59,
+      "learning_rate": 6.117895078355547e-05,
+      "loss": 2.4025,
+      "step": 280
+    },
+    {
+      "epoch": 4.75,
+      "learning_rate": 6.15599499474739e-05,
+      "loss": 2.4185,
+      "step": 290
+    },
+    {
+      "epoch": 4.92,
+      "learning_rate": 6.192803136799156e-05,
+      "loss": 2.4651,
+      "step": 300
+    },
+    {
+      "epoch": 5.08,
+      "learning_rate": 6.22840423458568e-05,
+      "loss": 2.4095,
+      "step": 310
+    },
+    {
+      "epoch": 5.25,
+      "learning_rate": 6.262874945799764e-05,
+      "loss": 2.3613,
+      "step": 320
+    },
+    {
+      "epoch": 5.41,
+      "learning_rate": 6.296284849694718e-05,
+      "loss": 2.381,
+      "step": 330
+    },
+    {
+      "epoch": 5.57,
+      "learning_rate": 6.328697292605637e-05,
+      "loss": 2.3387,
+      "step": 340
+    },
+    {
+      "epoch": 5.74,
+      "learning_rate": 6.360170110875688e-05,
+      "loss": 2.3331,
+      "step": 350
+    },
+    {
+      "epoch": 5.9,
+      "learning_rate": 6.390756251918218e-05,
+      "loss": 2.3831,
+      "step": 360
+    },
+    {
+      "epoch": 6.07,
+      "learning_rate": 6.420504310167487e-05,
+      "loss": 2.2977,
+      "step": 370
+    },
+    {
+      "epoch": 6.23,
+      "learning_rate": 6.449458991542025e-05,
+      "loss": 2.2816,
+      "step": 380
+    },
+    {
+      "epoch": 6.39,
+      "learning_rate": 6.477661517566246e-05,
+      "loss": 2.2826,
+      "step": 390
+    },
+    {
+      "epoch": 6.56,
+      "learning_rate": 6.505149978319905e-05,
+      "loss": 2.3021,
+      "step": 400
+    },
+    {
+      "epoch": 6.72,
+      "learning_rate": 6.531959641799339e-05,
+      "loss": 2.3166,
+      "step": 410
+    },
+    {
+      "epoch": 6.89,
+      "learning_rate": 6.55812322599475e-05,
+      "loss": 2.3421,
+      "step": 420
+    },
+    {
+      "epoch": 7.05,
+      "learning_rate": 6.583671138948966e-05,
+      "loss": 2.3094,
+      "step": 430
+    },
+    {
+      "epoch": 7.21,
+      "learning_rate": 6.608631691215467e-05,
+      "loss": 2.2268,
+      "step": 440
+    },
+    {
+      "epoch": 7.38,
+      "learning_rate": 6.633031284438359e-05,
+      "loss": 2.2722,
+      "step": 450
+    },
+    {
+      "epoch": 7.54,
+      "learning_rate": 6.656894579203935e-05,
+      "loss": 2.1933,
+      "step": 460
+    },
+    {
+      "epoch": 7.7,
+      "learning_rate": 6.680244644839293e-05,
+      "loss": 2.1991,
+      "step": 470
+    },
+    {
+      "epoch": 7.87,
+      "learning_rate": 6.703103093438967e-05,
+      "loss": 2.2637,
+      "step": 480
+    },
+    {
+      "epoch": 8.03,
+      "learning_rate": 6.725490200071283e-05,
+      "loss": 2.2277,
+      "step": 490
+    },
+    {
+      "epoch": 8.2,
+      "learning_rate": 6.747425010840046e-05,
+      "loss": 2.1819,
+      "step": 500
+    },
+    {
+      "epoch": 8.2,
+      "eval_loss": 2.786494255065918,
+      "eval_runtime": 13.6759,
+      "eval_samples_per_second": 47.602,
+      "eval_steps_per_second": 0.585,
+      "step": 500
+    },
+    {
+      "epoch": 8.36,
+      "learning_rate": 6.76892544024484e-05,
+      "loss": 2.1427,
+      "step": 510
+    },
+    {
+      "epoch": 8.52,
+      "learning_rate": 6.790008359086997e-05,
+      "loss": 2.173,
+      "step": 520
+    },
+    {
+      "epoch": 8.69,
+      "learning_rate": 6.810689674001973e-05,
+      "loss": 2.1895,
+      "step": 530
+    },
+    {
+      "epoch": 8.85,
+      "learning_rate": 6.830984399557421e-05,
+      "loss": 2.2101,
+      "step": 540
+    },
+    {
+      "epoch": 9.02,
+      "learning_rate": 6.850906723735608e-05,
+      "loss": 2.1926,
+      "step": 550
+    },
+    {
+      "epoch": 9.18,
+      "learning_rate": 6.870470067515499e-05,
+      "loss": 2.0861,
+      "step": 560
+    },
+    {
+      "epoch": 9.34,
+      "learning_rate": 6.889687139181228e-05,
+      "loss": 2.1092,
+      "step": 570
+    },
+    {
+      "epoch": 9.51,
+      "learning_rate": 6.908569983907343e-05,
+      "loss": 2.129,
+      "step": 580
+    },
+    {
+      "epoch": 9.67,
+      "learning_rate": 6.92713002910536e-05,
+      "loss": 2.113,
+      "step": 590
+    },
+    {
+      "epoch": 9.84,
+      "learning_rate": 6.945378125959108e-05,
+      "loss": 2.1234,
+      "step": 600
+    },
+    {
+      "epoch": 10.0,
+      "learning_rate": 6.963324587526918e-05,
+      "loss": 2.1028,
+      "step": 610
+    },
+    {
+      "epoch": 10.16,
+      "learning_rate": 6.980979223745634e-05,
+      "loss": 2.0476,
+      "step": 620
+    },
+    {
+      "epoch": 10.33,
+      "learning_rate": 6.998351373633953e-05,
+      "loss": 2.0879,
+      "step": 630
+    },
+    {
+      "epoch": 10.49,
+      "learning_rate": 7.015449934959717e-05,
+      "loss": 2.0547,
+      "step": 640
+    },
+    {
+      "epoch": 10.66,
+      "learning_rate": 7.032283391607138e-05,
+      "loss": 2.0791,
+      "step": 650
+    },
+    {
+      "epoch": 10.82,
+      "learning_rate": 7.048859838854671e-05,
+      "loss": 2.1454,
+      "step": 660
+    },
+    {
+      "epoch": 10.98,
+      "learning_rate": 7.065187006752065e-05,
+      "loss": 2.0957,
+      "step": 670
+    },
+    {
+      "epoch": 11.15,
+      "learning_rate": 7.08127228176559e-05,
+      "loss": 2.0561,
+      "step": 680
+    },
+    {
+      "epoch": 11.31,
+      "learning_rate": 7.097122726843138e-05,
+      "loss": 2.0563,
+      "step": 690
+    },
+    {
+      "epoch": 11.48,
+      "learning_rate": 7.112745100035642e-05,
+      "loss": 2.027,
+      "step": 700
+    },
+    {
+      "epoch": 11.64,
+      "learning_rate": 7.128145871797688e-05,
+      "loss": 2.0495,
+      "step": 710
+    },
+    {
+      "epoch": 11.8,
+      "learning_rate": 7.143331241078171e-05,
+      "loss": 2.019,
+      "step": 720
+    },
+    {
+      "epoch": 11.97,
+      "learning_rate": 7.158307150301139e-05,
+      "loss": 2.0242,
+      "step": 730
+    },
+    {
+      "epoch": 12.13,
+      "learning_rate": 7.17307929932744e-05,
+      "loss": 1.9962,
+      "step": 740
+    },
+    {
+      "epoch": 12.3,
+      "learning_rate": 7.187653158479249e-05,
+      "loss": 1.9971,
+      "step": 750
+    },
+    {
+      "epoch": 12.46,
+      "learning_rate": 7.202033980701978e-05,
+      "loss": 2.0236,
+      "step": 760
+    },
+    {
+      "epoch": 12.62,
+      "learning_rate": 7.216226812931204e-05,
+      "loss": 1.9923,
+      "step": 770
+    },
+    {
+      "epoch": 12.79,
+      "learning_rate": 7.2302365067262e-05,
+      "loss": 2.0244,
+      "step": 780
+    },
+    {
+      "epoch": 12.95,
+      "learning_rate": 7.244067728226103e-05,
+      "loss": 1.9846,
+      "step": 790
+    },
+    {
+      "epoch": 13.11,
+      "learning_rate": 7.257724967479857e-05,
+      "loss": 1.9811,
+      "step": 800
+    },
+    {
+      "epoch": 13.28,
+      "learning_rate": 7.271212547196624e-05,
+      "loss": 1.9709,
+      "step": 810
+    },
+    {
+      "epoch": 13.44,
+      "learning_rate": 7.284534630959291e-05,
+      "loss": 1.9652,
+      "step": 820
+    },
+    {
+      "epoch": 13.61,
+      "learning_rate": 7.297695230940184e-05,
+      "loss": 1.9605,
+      "step": 830
+    },
+    {
+      "epoch": 13.77,
+      "learning_rate": 7.310698215154704e-05,
+      "loss": 1.9692,
+      "step": 840
+    },
+    {
+      "epoch": 13.93,
+      "learning_rate": 7.323547314285732e-05,
+      "loss": 1.9945,
+      "step": 850
+    },
+    {
+      "epoch": 14.1,
+      "learning_rate": 7.336246128108918e-05,
+      "loss": 1.9222,
+      "step": 860
+    },
+    {
+      "epoch": 14.26,
+      "learning_rate": 7.348798131546546e-05,
+      "loss": 1.9283,
+      "step": 870
+    },
+    {
+      "epoch": 14.43,
+      "learning_rate": 7.36120668037542e-05,
+      "loss": 1.9376,
+      "step": 880
+    },
+    {
+      "epoch": 14.59,
+      "learning_rate": 7.37347501661228e-05,
+      "loss": 1.9247,
+      "step": 890
+    },
+    {
+      "epoch": 14.75,
+      "learning_rate": 7.385606273598311e-05,
+      "loss": 1.9218,
+      "step": 900
+    },
+    {
+      "epoch": 14.92,
+      "learning_rate": 7.397603480802732e-05,
+      "loss": 1.9492,
+      "step": 910
+    },
+    {
+      "epoch": 15.08,
+      "learning_rate": 7.409469568363888e-05,
+      "loss": 1.9235,
+      "step": 920
+    },
+    {
+      "epoch": 15.25,
+      "learning_rate": 7.421207371384837e-05,
+      "loss": 1.8671,
+      "step": 930
+    },
+    {
+      "epoch": 15.41,
+      "learning_rate": 7.432819633999247e-05,
+      "loss": 1.909,
+      "step": 940
+    },
+    {
+      "epoch": 15.57,
+      "learning_rate": 7.444309013222118e-05,
+      "loss": 1.8568,
+      "step": 950
+    },
+    {
+      "epoch": 15.74,
+      "learning_rate": 7.45567808259892e-05,
+      "loss": 1.9199,
+      "step": 960
+    },
+    {
+      "epoch": 15.9,
+      "learning_rate": 7.46692933566561e-05,
+      "loss": 1.9247,
+      "step": 970
+    },
+    {
+      "epoch": 16.07,
+      "learning_rate": 7.478065189231236e-05,
+      "loss": 1.895,
+      "step": 980
+    },
+    {
+      "epoch": 16.23,
+      "learning_rate": 7.489087986493874e-05,
+      "loss": 1.8821,
+      "step": 990
+    },
+    {
+      "epoch": 16.39,
+      "learning_rate": 7.5e-05,
+      "loss": 1.8423,
+      "step": 1000
+    },
+    {
+      "epoch": 16.39,
+      "eval_loss": 2.925347328186035,
+      "eval_runtime": 13.9926,
+      "eval_samples_per_second": 46.524,
+      "eval_steps_per_second": 0.572,
+      "step": 1000
+    },
+    {
+      "epoch": 16.56,
+      "learning_rate": 7.510803434456605e-05,
+      "loss": 1.8519,
+      "step": 1010
+    },
+    {
+      "epoch": 16.72,
+      "learning_rate": 7.521500429404794e-05,
+      "loss": 1.8578,
+      "step": 1020
+    },
+    {
+      "epoch": 16.89,
+      "learning_rate": 7.532093061762931e-05,
+      "loss": 1.8676,
+      "step": 1030
+    },
+    {
+      "epoch": 17.05,
+      "learning_rate": 7.54258334824695e-05,
+      "loss": 1.8492,
+      "step": 1040
+    },
+    {
+      "epoch": 17.21,
+      "learning_rate": 7.552973247674843e-05,
+      "loss": 1.8542,
+      "step": 1050
+    },
+    {
+      "epoch": 17.38,
+      "learning_rate": 7.563264663161926e-05,
+      "loss": 1.8312,
+      "step": 1060
+    },
+    {
+      "epoch": 17.54,
+      "learning_rate": 7.573459444213023e-05,
+      "loss": 1.8554,
+      "step": 1070
+    },
+    {
+      "epoch": 17.7,
+      "learning_rate": 7.583559388717374e-05,
+      "loss": 1.8484,
+      "step": 1080
+    },
+    {
+      "epoch": 17.87,
+      "learning_rate": 7.593566244851558e-05,
+      "loss": 1.8485,
+      "step": 1090
+    },
+    {
+      "epoch": 18.03,
+      "learning_rate": 7.603481712895562e-05,
+      "loss": 1.8505,
+      "step": 1100
+    },
+    {
+      "epoch": 18.2,
+      "learning_rate": 7.613307446966643e-05,
+      "loss": 1.8163,
+      "step": 1110
+    },
+    {
+      "epoch": 18.36,
+      "learning_rate": 7.623045056675453e-05,
+      "loss": 1.8382,
+      "step": 1120
+    },
+    {
+      "epoch": 18.52,
+      "learning_rate": 7.632696108708549e-05,
+      "loss": 1.8251,
+      "step": 1130
+    },
+    {
+      "epoch": 18.69,
+      "learning_rate": 7.642262128341181e-05,
+      "loss": 1.8252,
+      "step": 1140
+    },
+    {
+      "epoch": 18.85,
+      "learning_rate": 7.651744600884029e-05,
+      "loss": 1.849,
+      "step": 1150
+    },
+    {
+      "epoch": 19.02,
+      "learning_rate": 7.661144973067295e-05,
+      "loss": 1.8202,
+      "step": 1160
+    },
+    {
+      "epoch": 19.18,
+      "learning_rate": 7.670464654365404e-05,
+      "loss": 1.8013,
+      "step": 1170
+    },
+    {
+      "epoch": 19.34,
+      "learning_rate": 7.679705018265312e-05,
+      "loss": 1.8149,
+      "step": 1180
+    },
+    {
+      "epoch": 19.51,
+      "learning_rate": 7.688867403481326e-05,
+      "loss": 1.7919,
+      "step": 1190
+    },
+    {
+      "epoch": 19.67,
+      "learning_rate": 7.697953115119061e-05,
+      "loss": 1.801,
+      "step": 1200
+    },
+    {
+      "epoch": 19.84,
+      "learning_rate": 7.706963425791124e-05,
+      "loss": 1.8286,
+      "step": 1210
+    },
+    {
+      "epoch": 20.0,
+      "learning_rate": 7.71589957668687e-05,
+      "loss": 1.7945,
+      "step": 1220
+    },
+    {
+      "epoch": 20.16,
+      "learning_rate": 7.724762778598493e-05,
+      "loss": 1.7619,
+      "step": 1230
+    },
+    {
+      "epoch": 20.33,
+      "learning_rate": 7.733554212905587e-05,
+      "loss": 1.7693,
+      "step": 1240
+    },
+    {
+      "epoch": 20.49,
+      "learning_rate": 7.74227503252014e-05,
+      "loss": 1.7689,
+      "step": 1250
+    },
+    {
+      "epoch": 20.66,
+      "learning_rate": 7.750926362793907e-05,
+      "loss": 1.77,
+      "step": 1260
+    },
+    {
+      "epoch": 20.82,
+      "learning_rate": 7.759509302389892e-05,
+      "loss": 1.7765,
+      "step": 1270
+    },
+    {
+      "epoch": 20.98,
+      "learning_rate": 7.768024924119671e-05,
+      "loss": 1.7791,
+      "step": 1280
+    },
+    {
+      "epoch": 21.15,
+      "learning_rate": 7.776474275748121e-05,
+      "loss": 1.7514,
+      "step": 1290
+    },
+    {
+      "epoch": 21.31,
+      "learning_rate": 7.784858380767091e-05,
+      "loss": 1.7564,
+      "step": 1300
+    },
+    {
+      "epoch": 21.48,
+      "learning_rate": 7.793178239139409e-05,
+      "loss": 1.7541,
+      "step": 1310
+    },
+    {
+      "epoch": 21.64,
+      "learning_rate": 7.801434828014625e-05,
+      "loss": 1.7519,
+      "step": 1320
+    },
+    {
+      "epoch": 21.8,
+      "learning_rate": 7.809629102417713e-05,
+      "loss": 1.7862,
+      "step": 1330
+    },
+    {
+      "epoch": 21.97,
+      "learning_rate": 7.817761995912018e-05,
+      "loss": 1.7724,
+      "step": 1340
+    },
+    {
+      "epoch": 22.13,
+      "learning_rate": 7.825834421237515e-05,
+      "loss": 1.7565,
+      "step": 1350
+    },
+    {
+      "epoch": 22.3,
+      "learning_rate": 7.833847270925543e-05,
+      "loss": 1.7346,
+      "step": 1360
+    },
+    {
+      "epoch": 22.46,
+      "learning_rate": 7.841801417891016e-05,
+      "loss": 1.7238,
+      "step": 1370
+    },
+    {
+      "epoch": 22.62,
+      "learning_rate": 7.84969771600309e-05,
+      "loss": 1.738,
+      "step": 1380
+    },
+    {
+      "epoch": 22.79,
+      "learning_rate": 7.857537000635237e-05,
+      "loss": 1.7446,
+      "step": 1390
+    },
+    {
+      "epoch": 22.95,
+      "learning_rate": 7.865320089195594e-05,
+      "loss": 1.7395,
+      "step": 1400
+    },
+    {
+      "epoch": 23.11,
+      "learning_rate": 7.87304778163845e-05,
+      "loss": 1.7533,
+      "step": 1410
+    },
+    {
+      "epoch": 23.28,
+      "learning_rate": 7.880720860957641e-05,
+      "loss": 1.7101,
+      "step": 1420
+    },
+    {
+      "epoch": 23.44,
+      "learning_rate": 7.888340093662653e-05,
+      "loss": 1.7145,
+      "step": 1430
+    },
+    {
+      "epoch": 23.61,
+      "learning_rate": 7.895906230238123e-05,
+      "loss": 1.7496,
+      "step": 1440
+    },
+    {
+      "epoch": 23.77,
+      "learning_rate": 7.903420005587436e-05,
+      "loss": 1.7416,
+      "step": 1450
+    },
+    {
+      "epoch": 23.93,
+      "learning_rate": 7.910882139461093e-05,
+      "loss": 1.7315,
+      "step": 1460
+    },
+    {
+      "epoch": 24.1,
+      "learning_rate": 7.918293336870439e-05,
+      "loss": 1.7224,
+      "step": 1470
+    },
+    {
+      "epoch": 24.26,
+      "learning_rate": 7.925654288487392e-05,
+      "loss": 1.716,
+      "step": 1480
+    },
+    {
+      "epoch": 24.43,
+      "learning_rate": 7.932965671030685e-05,
+      "loss": 1.704,
+      "step": 1490
+    },
+    {
+      "epoch": 24.59,
+      "learning_rate": 7.940228147639202e-05,
+      "loss": 1.6873,
+      "step": 1500
+    },
+    {
+      "epoch": 24.59,
+      "eval_loss": 3.106438159942627,
+      "eval_runtime": 13.9459,
+      "eval_samples_per_second": 46.681,
+      "eval_steps_per_second": 0.574,
+      "step": 1500
+    },
+    {
+      "epoch": 24.75,
+      "learning_rate": 7.947442368232923e-05,
+      "loss": 1.7098,
+      "step": 1510
+    },
+    {
+      "epoch": 24.92,
+      "learning_rate": 7.954608969861931e-05,
+      "loss": 1.7058,
+      "step": 1520
+    },
+    {
+      "epoch": 25.08,
+      "learning_rate": 7.961728577043997e-05,
+      "loss": 1.7189,
+      "step": 1530
+    },
+    {
+      "epoch": 25.25,
+      "learning_rate": 7.968801802091157e-05,
+      "loss": 1.6689,
+      "step": 1540
+    },
+    {
+      "epoch": 25.41,
+      "learning_rate": 7.975829245425728e-05,
+      "loss": 1.709,
+      "step": 1550
+    },
+    {
+      "epoch": 25.57,
+      "learning_rate": 7.982811495886153e-05,
+      "loss": 1.6881,
+      "step": 1560
+    },
+    {
+      "epoch": 25.74,
+      "learning_rate": 7.989749131023083e-05,
+      "loss": 1.7032,
+      "step": 1570
+    },
+    {
+      "epoch": 25.9,
+      "learning_rate": 7.996642717386056e-05,
+      "loss": 1.6887,
+      "step": 1580
+    },
+    {
+      "epoch": 26.07,
+      "learning_rate": 8.003492810801127e-05,
+      "loss": 1.6961,
+      "step": 1590
+    },
+    {
+      "epoch": 26.23,
+      "learning_rate": 8.01029995663981e-05,
+      "loss": 1.6701,
+      "step": 1600
+    },
+    {
+      "epoch": 26.39,
+      "learning_rate": 8.017064690079624e-05,
+      "loss": 1.69,
+      "step": 1610
+    },
+    {
+      "epoch": 26.56,
+      "learning_rate": 8.023787536356576e-05,
+      "loss": 1.7125,
+      "step": 1620
+    },
+    {
+      "epoch": 26.72,
+      "learning_rate": 8.030469011009893e-05,
+      "loss": 1.6606,
+      "step": 1630
+    },
+    {
+      "epoch": 26.89,
+      "learning_rate": 8.037109620119243e-05,
+      "loss": 1.6649,
+      "step": 1640
+    },
+    {
+      "epoch": 27.05,
+      "learning_rate": 8.043709860534764e-05,
+      "loss": 1.6699,
+      "step": 1650
+    },
+    {
+      "epoch": 27.21,
+      "learning_rate": 8.050270220100136e-05,
+      "loss": 1.645,
+      "step": 1660
+    },
+    {
+      "epoch": 27.38,
+      "learning_rate": 8.056791177868957e-05,
+      "loss": 1.65,
+      "step": 1670
+    },
+    {
+      "epoch": 27.54,
+      "learning_rate": 8.063273204314657e-05,
+      "loss": 1.6552,
+      "step": 1680
+    },
+    {
+      "epoch": 27.7,
+      "learning_rate": 8.069716761534183e-05,
+      "loss": 1.6772,
+      "step": 1690
+    },
+    {
+      "epoch": 27.87,
+      "learning_rate": 8.076122303445684e-05,
+      "loss": 1.6664,
+      "step": 1700
+    },
+    {
+      "epoch": 28.03,
+      "learning_rate": 8.082490275980384e-05,
+      "loss": 1.6539,
+      "step": 1710
+    },
+    {
+      "epoch": 28.2,
+      "learning_rate": 8.088821117268871e-05,
+      "loss": 1.6616,
+      "step": 1720
+    },
+    {
+      "epoch": 28.36,
+      "learning_rate": 8.095115257821987e-05,
+      "loss": 1.6379,
+      "step": 1730
+    },
+    {
+      "epoch": 28.52,
+      "learning_rate": 8.1013731207065e-05,
+      "loss": 1.6536,
+      "step": 1740
+    },
+    {
+      "epoch": 28.69,
+      "learning_rate": 8.107595121715735e-05,
+      "loss": 1.6506,
+      "step": 1750
+    },
+    {
+      "epoch": 28.85,
+      "learning_rate": 8.113781669535373e-05,
+      "loss": 1.66,
+      "step": 1760
+    },
+    {
+      "epoch": 29.02,
+      "learning_rate": 8.119933165904515e-05,
+      "loss": 1.6548,
+      "step": 1770
+    },
+    {
+      "epoch": 29.18,
+      "learning_rate": 8.126050005772234e-05,
+      "loss": 1.6408,
+      "step": 1780
+    },
+    {
+      "epoch": 29.34,
+      "learning_rate": 8.132132577449732e-05,
+      "loss": 1.6533,
+      "step": 1790
+    },
+    {
+      "epoch": 29.51,
+      "learning_rate": 8.138181262758264e-05,
+      "loss": 1.6508,
+      "step": 1800
+    },
+    {
+      "epoch": 29.67,
+      "learning_rate": 8.144196437172959e-05,
+      "loss": 1.6302,
+      "step": 1810
+    },
+    {
+      "epoch": 29.84,
+      "learning_rate": 8.150178469962686e-05,
+      "loss": 1.6319,
+      "step": 1820
+    },
+    {
+      "epoch": 30.0,
+      "learning_rate": 8.156127724326073e-05,
+      "loss": 1.623,
+      "step": 1830
+    },
+    {
+      "epoch": 30.16,
+      "learning_rate": 8.16204455752384e-05,
+      "loss": 1.6299,
+      "step": 1840
+    },
+    {
+      "epoch": 30.33,
+      "learning_rate": 8.167929321007533e-05,
+      "loss": 1.6187,
+      "step": 1850
+    },
+    {
+      "epoch": 30.49,
+      "learning_rate": 8.17378236054479e-05,
+      "loss": 1.6138,
+      "step": 1860
+    },
+    {
+      "epoch": 30.66,
+      "learning_rate": 8.179604016341247e-05,
+      "loss": 1.6418,
+      "step": 1870
+    },
+    {
+      "epoch": 30.82,
+      "learning_rate": 8.1853946231592e-05,
+      "loss": 1.6433,
+      "step": 1880
+    },
+    {
+      "epoch": 30.98,
+      "learning_rate": 8.19115451043311e-05,
+      "loss": 1.6416,
+      "step": 1890
+    },
+    {
+      "epoch": 31.15,
+      "learning_rate": 8.196884002382071e-05,
+      "loss": 1.6244,
+      "step": 1900
+    },
+    {
+      "epoch": 31.31,
+      "learning_rate": 8.202583418119318e-05,
+      "loss": 1.6141,
+      "step": 1910
+    },
+    {
+      "epoch": 31.48,
+      "learning_rate": 8.208253071758874e-05,
+      "loss": 1.6033,
+      "step": 1920
+    },
+    {
+      "epoch": 31.64,
+      "learning_rate": 8.213893272519434e-05,
+      "loss": 1.6287,
+      "step": 1930
+    },
+    {
+      "epoch": 31.8,
+      "learning_rate": 8.219504324825564e-05,
+      "loss": 1.6268,
+      "step": 1940
+    },
+    {
+      "epoch": 31.97,
+      "learning_rate": 8.225086528406294e-05,
+      "loss": 1.6297,
+      "step": 1950
+    },
+    {
+      "epoch": 32.13,
+      "learning_rate": 8.23064017839119e-05,
+      "loss": 1.6012,
+      "step": 1960
+    },
+    {
+      "epoch": 32.3,
+      "learning_rate": 8.236165565403982e-05,
+      "loss": 1.6203,
+      "step": 1970
+    },
+    {
+      "epoch": 32.46,
+      "learning_rate": 8.241662975653826e-05,
+      "loss": 1.6107,
+      "step": 1980
+    },
+    {
+      "epoch": 32.62,
+      "learning_rate": 8.247132691024267e-05,
+      "loss": 1.6107,
+      "step": 1990
+    },
+    {
+      "epoch": 32.79,
+      "learning_rate": 8.252574989159953e-05,
+      "loss": 1.5886,
+      "step": 2000
+    },
+    {
+      "epoch": 32.79,
+      "eval_loss": 3.1871249675750732,
+      "eval_runtime": 13.1185,
+      "eval_samples_per_second": 49.624,
+      "eval_steps_per_second": 0.61,
+      "step": 2000
+    }
+  ],
+  "max_steps": 100000,
+  "num_train_epochs": 1640,
+  "total_flos": 394015948800.0,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:429c0dc53ed2ff5e8b7c901bbcf2bb07691a9d6ef79fc60cc81ec69a50afe4f7
+size 4015

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)