init

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": "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:bc5df05419d84d845c164ec380b04a5799fa7b38df5a13817d71e73bcc28f838
+size 11399226056

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0f99ed1ad363d23589e6f3284efb9cef791b4528e7db754ceff6790404a8b55
+size 14649

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7a2c754c62b0b84bfdd4902636921fe37de7b6948a04b7b1e52f029b76d40f2
+size 14654

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:897adda85809ba97e1738114f1832b64a43e9f592de3f8213e34180e00cb2154
+size 14654

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0681d28a3315be950b5ea7df44bc55e0676e5fb1f348828ee6633e41accf24ea
+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 +1,1248 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 29.850746268656717,
+  "global_step": 2000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.15,
+      "learning_rate": 2.7034552830322406e-05,
+      "loss": 4.2127,
+      "step": 10
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 3.5172764151612024e-05,
+      "loss": 3.3881,
+      "step": 20
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 3.993331259751083e-05,
+      "loss": 3.0315,
+      "step": 30
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 4.331097547290165e-05,
+      "loss": 2.8793,
+      "step": 40
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 4.5930894339355186e-05,
+      "loss": 2.7883,
+      "step": 50
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 4.8071523918800455e-05,
+      "loss": 2.7364,
+      "step": 60
+    },
+    {
+      "epoch": 1.04,
+      "learning_rate": 4.9881400439889756e-05,
+      "loss": 2.6109,
+      "step": 70
+    },
+    {
+      "epoch": 1.19,
+      "learning_rate": 5.1449186794191275e-05,
+      "loss": 2.5014,
+      "step": 80
+    },
+    {
+      "epoch": 1.34,
+      "learning_rate": 5.283207236469926e-05,
+      "loss": 2.4571,
+      "step": 90
+    },
+    {
+      "epoch": 1.49,
+      "learning_rate": 5.406910566064481e-05,
+      "loss": 2.5222,
+      "step": 100
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 5.518813839434375e-05,
+      "loss": 2.4918,
+      "step": 110
+    },
+    {
+      "epoch": 1.79,
+      "learning_rate": 5.620973524009008e-05,
+      "loss": 2.4835,
+      "step": 120
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 5.714951323824802e-05,
+      "loss": 2.4628,
+      "step": 130
+    },
+    {
+      "epoch": 2.09,
+      "learning_rate": 5.8019611761179374e-05,
+      "loss": 2.3972,
+      "step": 140
+    },
+    {
+      "epoch": 2.24,
+      "learning_rate": 5.882965410654361e-05,
+      "loss": 2.3233,
+      "step": 150
+    },
+    {
+      "epoch": 2.39,
+      "learning_rate": 5.95873981154809e-05,
+      "loss": 2.3075,
+      "step": 160
+    },
+    {
+      "epoch": 2.54,
+      "learning_rate": 6.029918920033657e-05,
+      "loss": 2.321,
+      "step": 170
+    },
+    {
+      "epoch": 2.69,
+      "learning_rate": 6.0970283685988885e-05,
+      "loss": 2.2946,
+      "step": 180
+    },
+    {
+      "epoch": 2.84,
+      "learning_rate": 6.160508461224668e-05,
+      "loss": 2.2593,
+      "step": 190
+    },
+    {
+      "epoch": 2.99,
+      "learning_rate": 6.220731698193443e-05,
+      "loss": 2.2308,
+      "step": 200
+    },
+    {
+      "epoch": 3.13,
+      "learning_rate": 6.278016020707817e-05,
+      "loss": 2.0929,
+      "step": 210
+    },
+    {
+      "epoch": 3.28,
+      "learning_rate": 6.332634971563337e-05,
+      "loss": 2.156,
+      "step": 220
+    },
+    {
+      "epoch": 3.43,
+      "learning_rate": 6.384825595366063e-05,
+      "loss": 2.1504,
+      "step": 230
+    },
+    {
+      "epoch": 3.58,
+      "learning_rate": 6.43479465613797e-05,
+      "loss": 2.182,
+      "step": 240
+    },
+    {
+      "epoch": 3.73,
+      "learning_rate": 6.482723584838796e-05,
+      "loss": 2.1635,
+      "step": 250
+    },
+    {
+      "epoch": 3.88,
+      "learning_rate": 6.528772455953764e-05,
+      "loss": 2.1302,
+      "step": 260
+    },
+    {
+      "epoch": 4.03,
+      "learning_rate": 6.573083213188768e-05,
+      "loss": 2.1364,
+      "step": 270
+    },
+    {
+      "epoch": 4.18,
+      "learning_rate": 6.6157823082469e-05,
+      "loss": 2.015,
+      "step": 280
+    },
+    {
+      "epoch": 4.33,
+      "learning_rate": 6.656982876347945e-05,
+      "loss": 2.0479,
+      "step": 290
+    },
+    {
+      "epoch": 4.48,
+      "learning_rate": 6.696786542783324e-05,
+      "loss": 1.9968,
+      "step": 300
+    },
+    {
+      "epoch": 4.63,
+      "learning_rate": 6.735284933140416e-05,
+      "loss": 2.012,
+      "step": 310
+    },
+    {
+      "epoch": 4.78,
+      "learning_rate": 6.772560943677052e-05,
+      "loss": 2.0458,
+      "step": 320
+    },
+    {
+      "epoch": 4.93,
+      "learning_rate": 6.808689816153217e-05,
+      "loss": 2.0991,
+      "step": 330
+    },
+    {
+      "epoch": 5.07,
+      "learning_rate": 6.84374005216262e-05,
+      "loss": 1.9845,
+      "step": 340
+    },
+    {
+      "epoch": 5.22,
+      "learning_rate": 6.877774194892253e-05,
+      "loss": 1.9325,
+      "step": 350
+    },
+    {
+      "epoch": 5.37,
+      "learning_rate": 6.910849500727851e-05,
+      "loss": 1.9603,
+      "step": 360
+    },
+    {
+      "epoch": 5.52,
+      "learning_rate": 6.943018518821426e-05,
+      "loss": 1.9095,
+      "step": 370
+    },
+    {
+      "epoch": 5.67,
+      "learning_rate": 6.97432959335363e-05,
+      "loss": 1.9443,
+      "step": 380
+    },
+    {
+      "epoch": 5.82,
+      "learning_rate": 7.004827300543644e-05,
+      "loss": 1.9461,
+      "step": 390
+    },
+    {
+      "epoch": 5.97,
+      "learning_rate": 7.034552830322405e-05,
+      "loss": 1.9462,
+      "step": 400
+    },
+    {
+      "epoch": 6.12,
+      "learning_rate": 7.063544320870321e-05,
+      "loss": 1.8685,
+      "step": 410
+    },
+    {
+      "epoch": 6.27,
+      "learning_rate": 7.09183715283678e-05,
+      "loss": 1.8694,
+      "step": 420
+    },
+    {
+      "epoch": 6.42,
+      "learning_rate": 7.119464208935388e-05,
+      "loss": 1.8429,
+      "step": 430
+    },
+    {
+      "epoch": 6.57,
+      "learning_rate": 7.146456103692298e-05,
+      "loss": 1.8458,
+      "step": 440
+    },
+    {
+      "epoch": 6.72,
+      "learning_rate": 7.172841387373204e-05,
+      "loss": 1.9065,
+      "step": 450
+    },
+    {
+      "epoch": 6.87,
+      "learning_rate": 7.198646727495026e-05,
+      "loss": 1.911,
+      "step": 460
+    },
+    {
+      "epoch": 7.01,
+      "learning_rate": 7.223897070815449e-05,
+      "loss": 1.8694,
+      "step": 470
+    },
+    {
+      "epoch": 7.16,
+      "learning_rate": 7.248615788266932e-05,
+      "loss": 1.7973,
+      "step": 480
+    },
+    {
+      "epoch": 7.31,
+      "learning_rate": 7.272824804945709e-05,
+      "loss": 1.8114,
+      "step": 490
+    },
+    {
+      "epoch": 7.46,
+      "learning_rate": 7.296544716967758e-05,
+      "loss": 1.7718,
+      "step": 500
+    },
+    {
+      "epoch": 7.46,
+      "eval_loss": 2.8234732151031494,
+      "eval_runtime": 35.5202,
+      "eval_samples_per_second": 18.328,
+      "eval_steps_per_second": 0.253,
+      "step": 500
+    },
+    {
+      "epoch": 7.61,
+      "learning_rate": 7.319794896752499e-05,
+      "loss": 1.7947,
+      "step": 510
+    },
+    {
+      "epoch": 7.76,
+      "learning_rate": 7.342593588082727e-05,
+      "loss": 1.8117,
+      "step": 520
+    },
+    {
+      "epoch": 7.91,
+      "learning_rate": 7.364957992109503e-05,
+      "loss": 1.8188,
+      "step": 530
+    },
+    {
+      "epoch": 8.06,
+      "learning_rate": 7.386904345317732e-05,
+      "loss": 1.8015,
+      "step": 540
+    },
+    {
+      "epoch": 8.21,
+      "learning_rate": 7.408447990337652e-05,
+      "loss": 1.734,
+      "step": 550
+    },
+    {
+      "epoch": 8.36,
+      "learning_rate": 7.429603440375862e-05,
+      "loss": 1.7217,
+      "step": 560
+    },
+    {
+      "epoch": 8.51,
+      "learning_rate": 7.450384437943511e-05,
+      "loss": 1.7398,
+      "step": 570
+    },
+    {
+      "epoch": 8.66,
+      "learning_rate": 7.470804008476907e-05,
+      "loss": 1.7452,
+      "step": 580
+    },
+    {
+      "epoch": 8.81,
+      "learning_rate": 7.490874509374465e-05,
+      "loss": 1.778,
+      "step": 590
+    },
+    {
+      "epoch": 8.96,
+      "learning_rate": 7.510607674912285e-05,
+      "loss": 1.7332,
+      "step": 600
+    },
+    {
+      "epoch": 9.1,
+      "learning_rate": 7.530014657447177e-05,
+      "loss": 1.6922,
+      "step": 610
+    },
+    {
+      "epoch": 9.25,
+      "learning_rate": 7.549106065269378e-05,
+      "loss": 1.6818,
+      "step": 620
+    },
+    {
+      "epoch": 9.4,
+      "learning_rate": 7.567891997426661e-05,
+      "loss": 1.6757,
+      "step": 630
+    },
+    {
+      "epoch": 9.55,
+      "learning_rate": 7.586382075806015e-05,
+      "loss": 1.752,
+      "step": 640
+    },
+    {
+      "epoch": 9.7,
+      "learning_rate": 7.604585474728082e-05,
+      "loss": 1.7074,
+      "step": 650
+    },
+    {
+      "epoch": 9.85,
+      "learning_rate": 7.62251094828218e-05,
+      "loss": 1.7052,
+      "step": 660
+    },
+    {
+      "epoch": 10.0,
+      "learning_rate": 7.640166855605846e-05,
+      "loss": 1.742,
+      "step": 670
+    },
+    {
+      "epoch": 10.15,
+      "learning_rate": 7.65756118429158e-05,
+      "loss": 1.6759,
+      "step": 680
+    },
+    {
+      "epoch": 10.3,
+      "learning_rate": 7.674701572084905e-05,
+      "loss": 1.6935,
+      "step": 690
+    },
+    {
+      "epoch": 10.45,
+      "learning_rate": 7.691595327021215e-05,
+      "loss": 1.6563,
+      "step": 700
+    },
+    {
+      "epoch": 10.6,
+      "learning_rate": 7.708249446134367e-05,
+      "loss": 1.6941,
+      "step": 710
+    },
+    {
+      "epoch": 10.75,
+      "learning_rate": 7.724670632856813e-05,
+      "loss": 1.676,
+      "step": 720
+    },
+    {
+      "epoch": 10.9,
+      "learning_rate": 7.740865313219632e-05,
+      "loss": 1.6948,
+      "step": 730
+    },
+    {
+      "epoch": 11.04,
+      "learning_rate": 7.756839650950389e-05,
+      "loss": 1.6687,
+      "step": 740
+    },
+    {
+      "epoch": 11.19,
+      "learning_rate": 7.772599561557638e-05,
+      "loss": 1.6469,
+      "step": 750
+    },
+    {
+      "epoch": 11.34,
+      "learning_rate": 7.788150725482592e-05,
+      "loss": 1.6783,
+      "step": 760
+    },
+    {
+      "epoch": 11.49,
+      "learning_rate": 7.803498600391108e-05,
+      "loss": 1.6408,
+      "step": 770
+    },
+    {
+      "epoch": 11.64,
+      "learning_rate": 7.818648432672608e-05,
+      "loss": 1.6521,
+      "step": 780
+    },
+    {
+      "epoch": 11.79,
+      "learning_rate": 7.833605268206489e-05,
+      "loss": 1.6451,
+      "step": 790
+    },
+    {
+      "epoch": 11.94,
+      "learning_rate": 7.848373962451368e-05,
+      "loss": 1.6504,
+      "step": 800
+    },
+    {
+      "epoch": 12.09,
+      "learning_rate": 7.862959189907611e-05,
+      "loss": 1.6431,
+      "step": 810
+    },
+    {
+      "epoch": 12.24,
+      "learning_rate": 7.877365452999284e-05,
+      "loss": 1.6131,
+      "step": 820
+    },
+    {
+      "epoch": 12.39,
+      "learning_rate": 7.89159709041777e-05,
+      "loss": 1.6256,
+      "step": 830
+    },
+    {
+      "epoch": 12.54,
+      "learning_rate": 7.905658284965742e-05,
+      "loss": 1.6257,
+      "step": 840
+    },
+    {
+      "epoch": 12.69,
+      "learning_rate": 7.919553070936936e-05,
+      "loss": 1.6143,
+      "step": 850
+    },
+    {
+      "epoch": 12.84,
+      "learning_rate": 7.933285341064351e-05,
+      "loss": 1.6383,
+      "step": 860
+    },
+    {
+      "epoch": 12.99,
+      "learning_rate": 7.946858853066788e-05,
+      "loss": 1.6234,
+      "step": 870
+    },
+    {
+      "epoch": 13.13,
+      "learning_rate": 7.960277235821263e-05,
+      "loss": 1.5871,
+      "step": 880
+    },
+    {
+      "epoch": 13.28,
+      "learning_rate": 7.973543995186684e-05,
+      "loss": 1.6028,
+      "step": 890
+    },
+    {
+      "epoch": 13.43,
+      "learning_rate": 7.986662519502166e-05,
+      "loss": 1.5723,
+      "step": 900
+    },
+    {
+      "epoch": 13.58,
+      "learning_rate": 7.999636084781537e-05,
+      "loss": 1.5936,
+      "step": 910
+    },
+    {
+      "epoch": 13.73,
+      "learning_rate": 8.012467859623988e-05,
+      "loss": 1.5869,
+      "step": 920
+    },
+    {
+      "epoch": 13.88,
+      "learning_rate": 8.025160909859258e-05,
+      "loss": 1.6018,
+      "step": 930
+    },
+    {
+      "epoch": 14.03,
+      "learning_rate": 8.037718202944411e-05,
+      "loss": 1.5926,
+      "step": 940
+    },
+    {
+      "epoch": 14.18,
+      "learning_rate": 8.050142612127945e-05,
+      "loss": 1.5546,
+      "step": 950
+    },
+    {
+      "epoch": 14.33,
+      "learning_rate": 8.062436920395896e-05,
+      "loss": 1.5601,
+      "step": 960
+    },
+    {
+      "epoch": 14.48,
+      "learning_rate": 8.074603824213446e-05,
+      "loss": 1.5668,
+      "step": 970
+    },
+    {
+      "epoch": 14.63,
+      "learning_rate": 8.086645937074672e-05,
+      "loss": 1.5623,
+      "step": 980
+    },
+    {
+      "epoch": 14.78,
+      "learning_rate": 8.09856579287206e-05,
+      "loss": 1.579,
+      "step": 990
+    },
+    {
+      "epoch": 14.93,
+      "learning_rate": 8.110365849096721e-05,
+      "loss": 1.5637,
+      "step": 1000
+    },
+    {
+      "epoch": 14.93,
+      "eval_loss": 3.1092050075531006,
+      "eval_runtime": 35.1461,
+      "eval_samples_per_second": 18.523,
+      "eval_steps_per_second": 0.256,
+      "step": 1000
+    },
+    {
+      "epoch": 15.07,
+      "learning_rate": 8.122048489879363e-05,
+      "loss": 1.5647,
+      "step": 1010
+    },
+    {
+      "epoch": 15.22,
+      "learning_rate": 8.133616028881462e-05,
+      "loss": 1.5349,
+      "step": 1020
+    },
+    {
+      "epoch": 15.37,
+      "learning_rate": 8.145070712045392e-05,
+      "loss": 1.542,
+      "step": 1030
+    },
+    {
+      "epoch": 15.52,
+      "learning_rate": 8.15641472021169e-05,
+      "loss": 1.5345,
+      "step": 1040
+    },
+    {
+      "epoch": 15.67,
+      "learning_rate": 8.167650171611095e-05,
+      "loss": 1.5491,
+      "step": 1050
+    },
+    {
+      "epoch": 15.82,
+      "learning_rate": 8.178779124238466e-05,
+      "loss": 1.5469,
+      "step": 1060
+    },
+    {
+      "epoch": 15.97,
+      "learning_rate": 8.189803578115246e-05,
+      "loss": 1.5825,
+      "step": 1070
+    },
+    {
+      "epoch": 16.12,
+      "learning_rate": 8.200725477446693e-05,
+      "loss": 1.5314,
+      "step": 1080
+    },
+    {
+      "epoch": 16.27,
+      "learning_rate": 8.211546712679696e-05,
+      "loss": 1.5126,
+      "step": 1090
+    },
+    {
+      "epoch": 16.42,
+      "learning_rate": 8.222269122466616e-05,
+      "loss": 1.5194,
+      "step": 1100
+    },
+    {
+      "epoch": 16.57,
+      "learning_rate": 8.232894495540269e-05,
+      "loss": 1.5276,
+      "step": 1110
+    },
+    {
+      "epoch": 16.72,
+      "learning_rate": 8.243424572504824e-05,
+      "loss": 1.5376,
+      "step": 1120
+    },
+    {
+      "epoch": 16.87,
+      "learning_rate": 8.2538610475471e-05,
+      "loss": 1.5393,
+      "step": 1130
+    },
+    {
+      "epoch": 17.01,
+      "learning_rate": 8.264205570072473e-05,
+      "loss": 1.5298,
+      "step": 1140
+    },
+    {
+      "epoch": 17.16,
+      "learning_rate": 8.27445974626934e-05,
+      "loss": 1.5135,
+      "step": 1150
+    },
+    {
+      "epoch": 17.31,
+      "learning_rate": 8.284625140605869e-05,
+      "loss": 1.5175,
+      "step": 1160
+    },
+    {
+      "epoch": 17.46,
+      "learning_rate": 8.294703277262488e-05,
+      "loss": 1.5106,
+      "step": 1170
+    },
+    {
+      "epoch": 17.61,
+      "learning_rate": 8.304695641503428e-05,
+      "loss": 1.5276,
+      "step": 1180
+    },
+    {
+      "epoch": 17.76,
+      "learning_rate": 8.31460368099039e-05,
+      "loss": 1.5227,
+      "step": 1190
+    },
+    {
+      "epoch": 17.91,
+      "learning_rate": 8.324428807041249e-05,
+      "loss": 1.5241,
+      "step": 1200
+    },
+    {
+      "epoch": 18.06,
+      "learning_rate": 8.334172395836509e-05,
+      "loss": 1.5187,
+      "step": 1210
+    },
+    {
+      "epoch": 18.21,
+      "learning_rate": 8.34383578957614e-05,
+      "loss": 1.4929,
+      "step": 1220
+    },
+    {
+      "epoch": 18.36,
+      "learning_rate": 8.353420297589165e-05,
+      "loss": 1.4934,
+      "step": 1230
+    },
+    {
+      "epoch": 18.51,
+      "learning_rate": 8.362927197398341e-05,
+      "loss": 1.5061,
+      "step": 1240
+    },
+    {
+      "epoch": 18.66,
+      "learning_rate": 8.372357735742074e-05,
+      "loss": 1.5068,
+      "step": 1250
+    },
+    {
+      "epoch": 18.81,
+      "learning_rate": 8.381713129555623e-05,
+      "loss": 1.5058,
+      "step": 1260
+    },
+    {
+      "epoch": 18.96,
+      "learning_rate": 8.390994566913507e-05,
+      "loss": 1.4944,
+      "step": 1270
+    },
+    {
+      "epoch": 19.1,
+      "learning_rate": 8.400203207934977e-05,
+      "loss": 1.4905,
+      "step": 1280
+    },
+    {
+      "epoch": 19.25,
+      "learning_rate": 8.409340185654231e-05,
+      "loss": 1.4908,
+      "step": 1290
+    },
+    {
+      "epoch": 19.4,
+      "learning_rate": 8.418406606857043e-05,
+      "loss": 1.4788,
+      "step": 1300
+    },
+    {
+      "epoch": 19.55,
+      "learning_rate": 8.427403552885332e-05,
+      "loss": 1.4851,
+      "step": 1310
+    },
+    {
+      "epoch": 19.7,
+      "learning_rate": 8.436332080411142e-05,
+      "loss": 1.4934,
+      "step": 1320
+    },
+    {
+      "epoch": 19.85,
+      "learning_rate": 8.445193222181402e-05,
+      "loss": 1.4862,
+      "step": 1330
+    },
+    {
+      "epoch": 20.0,
+      "learning_rate": 8.453987987734808e-05,
+      "loss": 1.4922,
+      "step": 1340
+    },
+    {
+      "epoch": 20.15,
+      "learning_rate": 8.462717364092046e-05,
+      "loss": 1.48,
+      "step": 1350
+    },
+    {
+      "epoch": 20.3,
+      "learning_rate": 8.471382316420545e-05,
+      "loss": 1.4731,
+      "step": 1360
+    },
+    {
+      "epoch": 20.45,
+      "learning_rate": 8.479983788674874e-05,
+      "loss": 1.4746,
+      "step": 1370
+    },
+    {
+      "epoch": 20.6,
+      "learning_rate": 8.488522704213867e-05,
+      "loss": 1.48,
+      "step": 1380
+    },
+    {
+      "epoch": 20.75,
+      "learning_rate": 8.496999966395455e-05,
+      "loss": 1.4743,
+      "step": 1390
+    },
+    {
+      "epoch": 20.9,
+      "learning_rate": 8.505416459150177e-05,
+      "loss": 1.4758,
+      "step": 1400
+    },
+    {
+      "epoch": 21.04,
+      "learning_rate": 8.513773047534291e-05,
+      "loss": 1.4738,
+      "step": 1410
+    },
+    {
+      "epoch": 21.19,
+      "learning_rate": 8.522070578263329e-05,
+      "loss": 1.4589,
+      "step": 1420
+    },
+    {
+      "epoch": 21.34,
+      "learning_rate": 8.530309880226936e-05,
+      "loss": 1.4783,
+      "step": 1430
+    },
+    {
+      "epoch": 21.49,
+      "learning_rate": 8.538491764985775e-05,
+      "loss": 1.4656,
+      "step": 1440
+    },
+    {
+      "epoch": 21.64,
+      "learning_rate": 8.546617027251222e-05,
+      "loss": 1.4702,
+      "step": 1450
+    },
+    {
+      "epoch": 21.79,
+      "learning_rate": 8.554686445348594e-05,
+      "loss": 1.4768,
+      "step": 1460
+    },
+    {
+      "epoch": 21.94,
+      "learning_rate": 8.562700781664552e-05,
+      "loss": 1.4802,
+      "step": 1470
+    },
+    {
+      "epoch": 22.09,
+      "learning_rate": 8.57066078307935e-05,
+      "loss": 1.463,
+      "step": 1480
+    },
+    {
+      "epoch": 22.24,
+      "learning_rate": 8.578567181384524e-05,
+      "loss": 1.4582,
+      "step": 1490
+    },
+    {
+      "epoch": 22.39,
+      "learning_rate": 8.586420693686602e-05,
+      "loss": 1.4588,
+      "step": 1500
+    },
+    {
+      "epoch": 22.39,
+      "eval_loss": 3.2750725746154785,
+      "eval_runtime": 34.9007,
+      "eval_samples_per_second": 18.653,
+      "eval_steps_per_second": 0.258,
+      "step": 1500
+    },
+    {
+      "epoch": 22.54,
+      "learning_rate": 8.594222022797423e-05,
+      "loss": 1.462,
+      "step": 1510
+    },
+    {
+      "epoch": 22.69,
+      "learning_rate": 8.601971857611555e-05,
+      "loss": 1.4671,
+      "step": 1520
+    },
+    {
+      "epoch": 22.84,
+      "learning_rate": 8.609670873471342e-05,
+      "loss": 1.4637,
+      "step": 1530
+    },
+    {
+      "epoch": 22.99,
+      "learning_rate": 8.617319732520071e-05,
+      "loss": 1.4661,
+      "step": 1540
+    },
+    {
+      "epoch": 23.13,
+      "learning_rate": 8.624919084043694e-05,
+      "loss": 1.4601,
+      "step": 1550
+    },
+    {
+      "epoch": 23.28,
+      "learning_rate": 8.632469564801571e-05,
+      "loss": 1.4553,
+      "step": 1560
+    },
+    {
+      "epoch": 23.43,
+      "learning_rate": 8.639971799346644e-05,
+      "loss": 1.4543,
+      "step": 1570
+    },
+    {
+      "epoch": 23.58,
+      "learning_rate": 8.647426400335451e-05,
+      "loss": 1.4667,
+      "step": 1580
+    },
+    {
+      "epoch": 23.73,
+      "learning_rate": 8.654833968828348e-05,
+      "loss": 1.4622,
+      "step": 1590
+    },
+    {
+      "epoch": 23.88,
+      "learning_rate": 8.66219509458033e-05,
+      "loss": 1.4654,
+      "step": 1600
+    },
+    {
+      "epoch": 24.03,
+      "learning_rate": 8.669510356322798e-05,
+      "loss": 1.4532,
+      "step": 1610
+    },
+    {
+      "epoch": 24.18,
+      "learning_rate": 8.676780322036573e-05,
+      "loss": 1.4525,
+      "step": 1620
+    },
+    {
+      "epoch": 24.33,
+      "learning_rate": 8.684005549216557e-05,
+      "loss": 1.4508,
+      "step": 1630
+    },
+    {
+      "epoch": 24.48,
+      "learning_rate": 8.691186585128246e-05,
+      "loss": 1.4526,
+      "step": 1640
+    },
+    {
+      "epoch": 24.63,
+      "learning_rate": 8.698323967056495e-05,
+      "loss": 1.4499,
+      "step": 1650
+    },
+    {
+      "epoch": 24.78,
+      "learning_rate": 8.705418222546732e-05,
+      "loss": 1.4633,
+      "step": 1660
+    },
+    {
+      "epoch": 24.93,
+      "learning_rate": 8.712469869638952e-05,
+      "loss": 1.4513,
+      "step": 1670
+    },
+    {
+      "epoch": 25.07,
+      "learning_rate": 8.719479417094704e-05,
+      "loss": 1.4543,
+      "step": 1680
+    },
+    {
+      "epoch": 25.22,
+      "learning_rate": 8.726447364617366e-05,
+      "loss": 1.4454,
+      "step": 1690
+    },
+    {
+      "epoch": 25.37,
+      "learning_rate": 8.733374203065898e-05,
+      "loss": 1.4462,
+      "step": 1700
+    },
+    {
+      "epoch": 25.52,
+      "learning_rate": 8.740260414662352e-05,
+      "loss": 1.4561,
+      "step": 1710
+    },
+    {
+      "epoch": 25.67,
+      "learning_rate": 8.747106473193313e-05,
+      "loss": 1.4503,
+      "step": 1720
+    },
+    {
+      "epoch": 25.82,
+      "learning_rate": 8.753912844205501e-05,
+      "loss": 1.453,
+      "step": 1730
+    },
+    {
+      "epoch": 25.97,
+      "learning_rate": 8.76067998519575e-05,
+      "loss": 1.4593,
+      "step": 1740
+    },
+    {
+      "epoch": 26.12,
+      "learning_rate": 8.76740834579553e-05,
+      "loss": 1.4412,
+      "step": 1750
+    },
+    {
+      "epoch": 26.27,
+      "learning_rate": 8.774098367950224e-05,
+      "loss": 1.4476,
+      "step": 1760
+    },
+    {
+      "epoch": 26.42,
+      "learning_rate": 8.780750486093308e-05,
+      "loss": 1.4412,
+      "step": 1770
+    },
+    {
+      "epoch": 26.57,
+      "learning_rate": 8.787365127315646e-05,
+      "loss": 1.4481,
+      "step": 1780
+    },
+    {
+      "epoch": 26.72,
+      "learning_rate": 8.79394271153003e-05,
+      "loss": 1.4471,
+      "step": 1790
+    },
+    {
+      "epoch": 26.87,
+      "learning_rate": 8.800483651631128e-05,
+      "loss": 1.447,
+      "step": 1800
+    },
+    {
+      "epoch": 27.01,
+      "learning_rate": 8.806988353651037e-05,
+      "loss": 1.4507,
+      "step": 1810
+    },
+    {
+      "epoch": 27.16,
+      "learning_rate": 8.813457216910499e-05,
+      "loss": 1.435,
+      "step": 1820
+    },
+    {
+      "epoch": 27.31,
+      "learning_rate": 8.81989063416602e-05,
+      "loss": 1.4361,
+      "step": 1830
+    },
+    {
+      "epoch": 27.46,
+      "learning_rate": 8.82628899175295e-05,
+      "loss": 1.4359,
+      "step": 1840
+    },
+    {
+      "epoch": 27.61,
+      "learning_rate": 8.832652669724704e-05,
+      "loss": 1.4379,
+      "step": 1850
+    },
+    {
+      "epoch": 27.76,
+      "learning_rate": 8.838982041988221e-05,
+      "loss": 1.4476,
+      "step": 1860
+    },
+    {
+      "epoch": 27.91,
+      "learning_rate": 8.845277476435792e-05,
+      "loss": 1.4395,
+      "step": 1870
+    },
+    {
+      "epoch": 28.06,
+      "learning_rate": 8.851539335073373e-05,
+      "loss": 1.4403,
+      "step": 1880
+    },
+    {
+      "epoch": 28.21,
+      "learning_rate": 8.857767974145503e-05,
+      "loss": 1.4387,
+      "step": 1890
+    },
+    {
+      "epoch": 28.36,
+      "learning_rate": 8.863963744256908e-05,
+      "loss": 1.4388,
+      "step": 1900
+    },
+    {
+      "epoch": 28.51,
+      "learning_rate": 8.87012699049093e-05,
+      "loss": 1.4377,
+      "step": 1910
+    },
+    {
+      "epoch": 28.66,
+      "learning_rate": 8.876258052524857e-05,
+      "loss": 1.4367,
+      "step": 1920
+    },
+    {
+      "epoch": 28.81,
+      "learning_rate": 8.882357264742258e-05,
+      "loss": 1.4482,
+      "step": 1930
+    },
+    {
+      "epoch": 28.96,
+      "learning_rate": 8.88842495634241e-05,
+      "loss": 1.4354,
+      "step": 1940
+    },
+    {
+      "epoch": 29.1,
+      "learning_rate": 8.894461451446924e-05,
+      "loss": 1.4333,
+      "step": 1950
+    },
+    {
+      "epoch": 29.25,
+      "learning_rate": 8.900467069203634e-05,
+      "loss": 1.4334,
+      "step": 1960
+    },
+    {
+      "epoch": 29.4,
+      "learning_rate": 8.906442123887845e-05,
+      "loss": 1.4454,
+      "step": 1970
+    },
+    {
+      "epoch": 29.55,
+      "learning_rate": 8.912386925001022e-05,
+      "loss": 1.4368,
+      "step": 1980
+    },
+    {
+      "epoch": 29.7,
+      "learning_rate": 8.918301777366981e-05,
+      "loss": 1.4319,
+      "step": 1990
+    },
+    {
+      "epoch": 29.85,
+      "learning_rate": 8.924186981225684e-05,
+      "loss": 1.4337,
+      "step": 2000
+    },
+    {
+      "epoch": 29.85,
+      "eval_loss": 3.362933874130249,
+      "eval_runtime": 35.3655,
+      "eval_samples_per_second": 18.408,
+      "eval_steps_per_second": 0.254,
+      "step": 2000
+    }
+  ],
+  "max_steps": 50000,
+  "num_train_epochs": 747,
+  "total_flos": 348961395840.0,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54e9565836fceb2464f6ef4d1760033dc4de56642205b8404884d9b18cef7623
+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)