small fix

added_tokens.json

@@ -0,0 +1,3 @@
+{
+  "[PAD]": 32000
+}

config.json

@@ -0,0 +1,24 @@
+{
+  "_name_or_path": "decapoda-research/llama-7b-hf",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "bos_token_id": 0,
+  "eos_token_id": 1,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_position_embeddings": 2048,
+  "max_sequence_length": 2048,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "pad_token_id": -1,
+  "rms_norm_eps": 1e-06,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.28.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32001
+}

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 1,
+  "pad_token_id": 0,
+  "transformers_version": "4.28.0.dev0"
+}

latest

@@ -0,0 +1 @@
+global_step222

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:275ee8dcae406d0cf4af5a06830a049419b441f034e370831a6a9ef90da84625
+size 13476958625

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e3c5cb412e12159a59afe5657ce4b5e0a06e7fb420bedbb5228fe1245702762
+size 14583

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:741230672078323886b763e522c728741456a587860909fc529ce815a7aca5ec
+size 14583

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ea587886b41579993bb5d20c79047b968ae2d71d22ba4c739b07ce31d7486a6
+size 14583

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ab727740f74dd67e60283d27b4339609a1dda888b067cc06520e2f1d7dc17db
+size 14583

rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85fbffd04f81c5419775f5b8507ac368aa0ff88b146755becf7b9cd26c139501
+size 14583

rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fcfb5fd9b58d3febef1eee1e52cb6997af12bac5fccd4be5e31c38721f4c3410
+size 14583

rng_state_6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f120fd79fe083c8bfcd736e4bd37c6fe37ba8dc492e550f309fb809413a218d
+size 14583

rng_state_7.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aed17a3c5698176f919ced6408c261f11e642658a3bdc526ead16625e7bb4a6f
+size 14583

special_tokens_map.json

@@ -0,0 +1 @@
+{}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

tokenizer_config.json

@@ -0,0 +1,9 @@
+{
+  "bos_token": "",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "",
+  "model_max_length": 1000000000000000019884624838656,
+  "special_tokens_map_file": "/mnt/data3/wendaxu/.cache/huggingface/hub/models--decapoda-research--llama-7b-hf/snapshots/5f98eefcc80e437ef68d457ad7bf167c2c6a1348/special_tokens_map.json",
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": ""
+}

trainer_state.json

@@ -0,0 +1,1348 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.9873843566021865,
+  "global_step": 222,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "learning_rate": 0,
+      "loss": 9.0082,
+      "step": 1
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0,
+      "loss": 8.6216,
+      "step": 2
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0,
+      "loss": 8.9495,
+      "step": 3
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0,
+      "loss": 9.3486,
+      "step": 4
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0,
+      "loss": 9.1724,
+      "step": 5
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0,
+      "loss": 8.6433,
+      "step": 6
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0,
+      "loss": 8.872,
+      "step": 7
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0,
+      "loss": 10.083,
+      "step": 8
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0,
+      "loss": 10.3374,
+      "step": 9
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 2e-05,
+      "loss": 9.8785,
+      "step": 10
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 2e-05,
+      "loss": 10.5295,
+      "step": 11
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 2e-05,
+      "loss": 7.6001,
+      "step": 12
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 2e-05,
+      "loss": 7.0393,
+      "step": 13
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 2e-05,
+      "loss": 6.9884,
+      "step": 14
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 2e-05,
+      "loss": 6.9941,
+      "step": 15
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 2e-05,
+      "loss": 7.235,
+      "step": 16
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 2e-05,
+      "loss": 6.7456,
+      "step": 17
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 2e-05,
+      "loss": 6.7632,
+      "step": 18
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 2e-05,
+      "loss": 6.3373,
+      "step": 19
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 2e-05,
+      "loss": 6.5851,
+      "step": 20
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 2e-05,
+      "loss": 6.8799,
+      "step": 21
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 2e-05,
+      "loss": 6.0525,
+      "step": 22
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 2e-05,
+      "loss": 6.178,
+      "step": 23
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 2e-05,
+      "loss": 6.6378,
+      "step": 24
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 2e-05,
+      "loss": 6.1607,
+      "step": 25
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 2e-05,
+      "loss": 6.2084,
+      "step": 26
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 2e-05,
+      "loss": 5.7216,
+      "step": 27
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 2e-05,
+      "loss": 5.5294,
+      "step": 28
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 2e-05,
+      "loss": 5.6123,
+      "step": 29
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 2e-05,
+      "loss": 5.3022,
+      "step": 30
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 2e-05,
+      "loss": 5.2346,
+      "step": 31
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 2e-05,
+      "loss": 5.268,
+      "step": 32
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 2e-05,
+      "loss": 4.8788,
+      "step": 33
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 2e-05,
+      "loss": 4.8944,
+      "step": 34
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 2e-05,
+      "loss": 4.487,
+      "step": 35
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 2e-05,
+      "loss": 4.3159,
+      "step": 36
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 2e-05,
+      "loss": 4.1472,
+      "step": 37
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 2e-05,
+      "loss": 4.2212,
+      "step": 38
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 2e-05,
+      "loss": 4.1312,
+      "step": 39
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 2e-05,
+      "loss": 3.9116,
+      "step": 40
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 2e-05,
+      "loss": 3.8727,
+      "step": 41
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 2e-05,
+      "loss": 3.4962,
+      "step": 42
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 2e-05,
+      "loss": 3.2736,
+      "step": 43
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 2e-05,
+      "loss": 3.1902,
+      "step": 44
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 2e-05,
+      "loss": 3.0781,
+      "step": 45
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 2e-05,
+      "loss": 2.9503,
+      "step": 46
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 2e-05,
+      "loss": 2.6742,
+      "step": 47
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 2e-05,
+      "loss": 2.5908,
+      "step": 48
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 2e-05,
+      "loss": 2.4428,
+      "step": 49
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 2e-05,
+      "loss": 2.1305,
+      "step": 50
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 2e-05,
+      "loss": 2.0908,
+      "step": 51
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 2e-05,
+      "loss": 1.8319,
+      "step": 52
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 2e-05,
+      "loss": 1.6784,
+      "step": 53
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 2e-05,
+      "loss": 1.553,
+      "step": 54
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 2e-05,
+      "loss": 1.3679,
+      "step": 55
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 2e-05,
+      "loss": 1.2306,
+      "step": 56
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 2e-05,
+      "loss": 1.0531,
+      "step": 57
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 2e-05,
+      "loss": 0.9721,
+      "step": 58
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 2e-05,
+      "loss": 0.8153,
+      "step": 59
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 2e-05,
+      "loss": 0.6768,
+      "step": 60
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 2e-05,
+      "loss": 0.5962,
+      "step": 61
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 2e-05,
+      "loss": 0.5061,
+      "step": 62
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 2e-05,
+      "loss": 0.4319,
+      "step": 63
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 2e-05,
+      "loss": 0.372,
+      "step": 64
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 2e-05,
+      "loss": 0.327,
+      "step": 65
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 2e-05,
+      "loss": 0.2799,
+      "step": 66
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 2e-05,
+      "loss": 0.2455,
+      "step": 67
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 2e-05,
+      "loss": 0.2267,
+      "step": 68
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 2e-05,
+      "loss": 0.2177,
+      "step": 69
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 2e-05,
+      "loss": 0.2029,
+      "step": 70
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 2e-05,
+      "loss": 0.1958,
+      "step": 71
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 2e-05,
+      "loss": 0.1748,
+      "step": 72
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 2e-05,
+      "loss": 0.1772,
+      "step": 73
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 2e-05,
+      "loss": 0.1639,
+      "step": 74
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 2e-05,
+      "loss": 0.1495,
+      "step": 75
+    },
+    {
+      "epoch": 1.02,
+      "learning_rate": 2e-05,
+      "loss": 0.1595,
+      "step": 76
+    },
+    {
+      "epoch": 1.04,
+      "learning_rate": 2e-05,
+      "loss": 0.141,
+      "step": 77
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 2e-05,
+      "loss": 0.1411,
+      "step": 78
+    },
+    {
+      "epoch": 1.06,
+      "learning_rate": 2e-05,
+      "loss": 0.1456,
+      "step": 79
+    },
+    {
+      "epoch": 1.08,
+      "learning_rate": 2e-05,
+      "loss": 0.1396,
+      "step": 80
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 2e-05,
+      "loss": 0.1276,
+      "step": 81
+    },
+    {
+      "epoch": 1.1,
+      "learning_rate": 2e-05,
+      "loss": 0.1285,
+      "step": 82
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 2e-05,
+      "loss": 0.1274,
+      "step": 83
+    },
+    {
+      "epoch": 1.13,
+      "learning_rate": 2e-05,
+      "loss": 0.1315,
+      "step": 84
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 2e-05,
+      "loss": 0.1283,
+      "step": 85
+    },
+    {
+      "epoch": 1.16,
+      "learning_rate": 2e-05,
+      "loss": 0.1055,
+      "step": 86
+    },
+    {
+      "epoch": 1.17,
+      "learning_rate": 2e-05,
+      "loss": 0.1164,
+      "step": 87
+    },
+    {
+      "epoch": 1.18,
+      "learning_rate": 2e-05,
+      "loss": 0.1306,
+      "step": 88
+    },
+    {
+      "epoch": 1.2,
+      "learning_rate": 2e-05,
+      "loss": 0.121,
+      "step": 89
+    },
+    {
+      "epoch": 1.21,
+      "learning_rate": 2e-05,
+      "loss": 0.1275,
+      "step": 90
+    },
+    {
+      "epoch": 1.22,
+      "learning_rate": 2e-05,
+      "loss": 0.1081,
+      "step": 91
+    },
+    {
+      "epoch": 1.24,
+      "learning_rate": 2e-05,
+      "loss": 0.1114,
+      "step": 92
+    },
+    {
+      "epoch": 1.25,
+      "learning_rate": 2e-05,
+      "loss": 0.1186,
+      "step": 93
+    },
+    {
+      "epoch": 1.26,
+      "learning_rate": 2e-05,
+      "loss": 0.1097,
+      "step": 94
+    },
+    {
+      "epoch": 1.28,
+      "learning_rate": 2e-05,
+      "loss": 0.1163,
+      "step": 95
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 2e-05,
+      "loss": 0.113,
+      "step": 96
+    },
+    {
+      "epoch": 1.31,
+      "learning_rate": 2e-05,
+      "loss": 0.1217,
+      "step": 97
+    },
+    {
+      "epoch": 1.32,
+      "learning_rate": 2e-05,
+      "loss": 0.1132,
+      "step": 98
+    },
+    {
+      "epoch": 1.33,
+      "learning_rate": 2e-05,
+      "loss": 0.1203,
+      "step": 99
+    },
+    {
+      "epoch": 1.35,
+      "learning_rate": 2e-05,
+      "loss": 0.1094,
+      "step": 100
+    },
+    {
+      "epoch": 1.36,
+      "learning_rate": 2e-05,
+      "loss": 0.1279,
+      "step": 101
+    },
+    {
+      "epoch": 1.37,
+      "learning_rate": 2e-05,
+      "loss": 0.1043,
+      "step": 102
+    },
+    {
+      "epoch": 1.39,
+      "learning_rate": 2e-05,
+      "loss": 0.1097,
+      "step": 103
+    },
+    {
+      "epoch": 1.4,
+      "learning_rate": 2e-05,
+      "loss": 0.1125,
+      "step": 104
+    },
+    {
+      "epoch": 1.41,
+      "learning_rate": 2e-05,
+      "loss": 0.118,
+      "step": 105
+    },
+    {
+      "epoch": 1.43,
+      "learning_rate": 2e-05,
+      "loss": 0.0962,
+      "step": 106
+    },
+    {
+      "epoch": 1.44,
+      "learning_rate": 2e-05,
+      "loss": 0.1093,
+      "step": 107
+    },
+    {
+      "epoch": 1.45,
+      "learning_rate": 2e-05,
+      "loss": 0.1219,
+      "step": 108
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 2e-05,
+      "loss": 0.1046,
+      "step": 109
+    },
+    {
+      "epoch": 1.48,
+      "learning_rate": 2e-05,
+      "loss": 0.1017,
+      "step": 110
+    },
+    {
+      "epoch": 1.49,
+      "learning_rate": 2e-05,
+      "loss": 0.1044,
+      "step": 111
+    },
+    {
+      "epoch": 1.51,
+      "learning_rate": 2e-05,
+      "loss": 0.1084,
+      "step": 112
+    },
+    {
+      "epoch": 1.52,
+      "learning_rate": 2e-05,
+      "loss": 0.1158,
+      "step": 113
+    },
+    {
+      "epoch": 1.53,
+      "learning_rate": 2e-05,
+      "loss": 0.1074,
+      "step": 114
+    },
+    {
+      "epoch": 1.55,
+      "learning_rate": 2e-05,
+      "loss": 0.1075,
+      "step": 115
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 2e-05,
+      "loss": 0.1052,
+      "step": 116
+    },
+    {
+      "epoch": 1.57,
+      "learning_rate": 2e-05,
+      "loss": 0.1017,
+      "step": 117
+    },
+    {
+      "epoch": 1.59,
+      "learning_rate": 2e-05,
+      "loss": 0.0992,
+      "step": 118
+    },
+    {
+      "epoch": 1.6,
+      "learning_rate": 2e-05,
+      "loss": 0.1096,
+      "step": 119
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 2e-05,
+      "loss": 0.1272,
+      "step": 120
+    },
+    {
+      "epoch": 1.63,
+      "learning_rate": 2e-05,
+      "loss": 0.1049,
+      "step": 121
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 2e-05,
+      "loss": 0.0998,
+      "step": 122
+    },
+    {
+      "epoch": 1.66,
+      "learning_rate": 2e-05,
+      "loss": 0.1052,
+      "step": 123
+    },
+    {
+      "epoch": 1.67,
+      "learning_rate": 2e-05,
+      "loss": 0.1177,
+      "step": 124
+    },
+    {
+      "epoch": 1.68,
+      "learning_rate": 2e-05,
+      "loss": 0.1052,
+      "step": 125
+    },
+    {
+      "epoch": 1.7,
+      "learning_rate": 2e-05,
+      "loss": 0.0997,
+      "step": 126
+    },
+    {
+      "epoch": 1.71,
+      "learning_rate": 2e-05,
+      "loss": 0.1115,
+      "step": 127
+    },
+    {
+      "epoch": 1.72,
+      "learning_rate": 2e-05,
+      "loss": 0.109,
+      "step": 128
+    },
+    {
+      "epoch": 1.74,
+      "learning_rate": 2e-05,
+      "loss": 0.1094,
+      "step": 129
+    },
+    {
+      "epoch": 1.75,
+      "learning_rate": 2e-05,
+      "loss": 0.0988,
+      "step": 130
+    },
+    {
+      "epoch": 1.76,
+      "learning_rate": 2e-05,
+      "loss": 0.1088,
+      "step": 131
+    },
+    {
+      "epoch": 1.78,
+      "learning_rate": 2e-05,
+      "loss": 0.0988,
+      "step": 132
+    },
+    {
+      "epoch": 1.79,
+      "learning_rate": 2e-05,
+      "loss": 0.0931,
+      "step": 133
+    },
+    {
+      "epoch": 1.8,
+      "learning_rate": 2e-05,
+      "loss": 0.0989,
+      "step": 134
+    },
+    {
+      "epoch": 1.82,
+      "learning_rate": 2e-05,
+      "loss": 0.1099,
+      "step": 135
+    },
+    {
+      "epoch": 1.83,
+      "learning_rate": 2e-05,
+      "loss": 0.107,
+      "step": 136
+    },
+    {
+      "epoch": 1.84,
+      "learning_rate": 2e-05,
+      "loss": 0.0991,
+      "step": 137
+    },
+    {
+      "epoch": 1.86,
+      "learning_rate": 2e-05,
+      "loss": 0.1045,
+      "step": 138
+    },
+    {
+      "epoch": 1.87,
+      "learning_rate": 2e-05,
+      "loss": 0.1112,
+      "step": 139
+    },
+    {
+      "epoch": 1.88,
+      "learning_rate": 2e-05,
+      "loss": 0.1128,
+      "step": 140
+    },
+    {
+      "epoch": 1.9,
+      "learning_rate": 2e-05,
+      "loss": 0.1106,
+      "step": 141
+    },
+    {
+      "epoch": 1.91,
+      "learning_rate": 2e-05,
+      "loss": 0.0934,
+      "step": 142
+    },
+    {
+      "epoch": 1.92,
+      "learning_rate": 2e-05,
+      "loss": 0.1105,
+      "step": 143
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 2e-05,
+      "loss": 0.1024,
+      "step": 144
+    },
+    {
+      "epoch": 1.95,
+      "learning_rate": 2e-05,
+      "loss": 0.0985,
+      "step": 145
+    },
+    {
+      "epoch": 1.96,
+      "learning_rate": 2e-05,
+      "loss": 0.1051,
+      "step": 146
+    },
+    {
+      "epoch": 1.98,
+      "learning_rate": 2e-05,
+      "loss": 0.0955,
+      "step": 147
+    },
+    {
+      "epoch": 1.99,
+      "learning_rate": 2e-05,
+      "loss": 0.0915,
+      "step": 148
+    },
+    {
+      "epoch": 2.01,
+      "learning_rate": 2e-05,
+      "loss": 0.0985,
+      "step": 149
+    },
+    {
+      "epoch": 2.02,
+      "learning_rate": 2e-05,
+      "loss": 0.0678,
+      "step": 150
+    },
+    {
+      "epoch": 2.03,
+      "learning_rate": 2e-05,
+      "loss": 0.0667,
+      "step": 151
+    },
+    {
+      "epoch": 2.05,
+      "learning_rate": 2e-05,
+      "loss": 0.0706,
+      "step": 152
+    },
+    {
+      "epoch": 2.06,
+      "learning_rate": 2e-05,
+      "loss": 0.0689,
+      "step": 153
+    },
+    {
+      "epoch": 2.07,
+      "learning_rate": 2e-05,
+      "loss": 0.0661,
+      "step": 154
+    },
+    {
+      "epoch": 2.09,
+      "learning_rate": 2e-05,
+      "loss": 0.0705,
+      "step": 155
+    },
+    {
+      "epoch": 2.1,
+      "learning_rate": 2e-05,
+      "loss": 0.077,
+      "step": 156
+    },
+    {
+      "epoch": 2.11,
+      "learning_rate": 2e-05,
+      "loss": 0.0605,
+      "step": 157
+    },
+    {
+      "epoch": 2.13,
+      "learning_rate": 2e-05,
+      "loss": 0.0653,
+      "step": 158
+    },
+    {
+      "epoch": 2.14,
+      "learning_rate": 2e-05,
+      "loss": 0.07,
+      "step": 159
+    },
+    {
+      "epoch": 2.15,
+      "learning_rate": 2e-05,
+      "loss": 0.0694,
+      "step": 160
+    },
+    {
+      "epoch": 2.17,
+      "learning_rate": 2e-05,
+      "loss": 0.0655,
+      "step": 161
+    },
+    {
+      "epoch": 2.18,
+      "learning_rate": 2e-05,
+      "loss": 0.061,
+      "step": 162
+    },
+    {
+      "epoch": 2.19,
+      "learning_rate": 2e-05,
+      "loss": 0.0799,
+      "step": 163
+    },
+    {
+      "epoch": 2.21,
+      "learning_rate": 2e-05,
+      "loss": 0.0637,
+      "step": 164
+    },
+    {
+      "epoch": 2.22,
+      "learning_rate": 2e-05,
+      "loss": 0.0711,
+      "step": 165
+    },
+    {
+      "epoch": 2.23,
+      "learning_rate": 2e-05,
+      "loss": 0.0668,
+      "step": 166
+    },
+    {
+      "epoch": 2.25,
+      "learning_rate": 2e-05,
+      "loss": 0.0699,
+      "step": 167
+    },
+    {
+      "epoch": 2.26,
+      "learning_rate": 2e-05,
+      "loss": 0.0748,
+      "step": 168
+    },
+    {
+      "epoch": 2.27,
+      "learning_rate": 2e-05,
+      "loss": 0.0614,
+      "step": 169
+    },
+    {
+      "epoch": 2.29,
+      "learning_rate": 2e-05,
+      "loss": 0.0676,
+      "step": 170
+    },
+    {
+      "epoch": 2.3,
+      "learning_rate": 2e-05,
+      "loss": 0.0628,
+      "step": 171
+    },
+    {
+      "epoch": 2.31,
+      "learning_rate": 2e-05,
+      "loss": 0.0642,
+      "step": 172
+    },
+    {
+      "epoch": 2.33,
+      "learning_rate": 2e-05,
+      "loss": 0.0678,
+      "step": 173
+    },
+    {
+      "epoch": 2.34,
+      "learning_rate": 2e-05,
+      "loss": 0.0629,
+      "step": 174
+    },
+    {
+      "epoch": 2.35,
+      "learning_rate": 2e-05,
+      "loss": 0.0645,
+      "step": 175
+    },
+    {
+      "epoch": 2.37,
+      "learning_rate": 2e-05,
+      "loss": 0.0612,
+      "step": 176
+    },
+    {
+      "epoch": 2.38,
+      "learning_rate": 2e-05,
+      "loss": 0.0575,
+      "step": 177
+    },
+    {
+      "epoch": 2.4,
+      "learning_rate": 2e-05,
+      "loss": 0.0651,
+      "step": 178
+    },
+    {
+      "epoch": 2.41,
+      "learning_rate": 2e-05,
+      "loss": 0.0679,
+      "step": 179
+    },
+    {
+      "epoch": 2.42,
+      "learning_rate": 2e-05,
+      "loss": 0.0648,
+      "step": 180
+    },
+    {
+      "epoch": 2.44,
+      "learning_rate": 2e-05,
+      "loss": 0.0685,
+      "step": 181
+    },
+    {
+      "epoch": 2.45,
+      "learning_rate": 2e-05,
+      "loss": 0.0646,
+      "step": 182
+    },
+    {
+      "epoch": 2.46,
+      "learning_rate": 2e-05,
+      "loss": 0.0688,
+      "step": 183
+    },
+    {
+      "epoch": 2.48,
+      "learning_rate": 2e-05,
+      "loss": 0.068,
+      "step": 184
+    },
+    {
+      "epoch": 2.49,
+      "learning_rate": 2e-05,
+      "loss": 0.067,
+      "step": 185
+    },
+    {
+      "epoch": 2.5,
+      "learning_rate": 2e-05,
+      "loss": 0.0683,
+      "step": 186
+    },
+    {
+      "epoch": 2.52,
+      "learning_rate": 2e-05,
+      "loss": 0.0668,
+      "step": 187
+    },
+    {
+      "epoch": 2.53,
+      "learning_rate": 2e-05,
+      "loss": 0.0663,
+      "step": 188
+    },
+    {
+      "epoch": 2.54,
+      "learning_rate": 2e-05,
+      "loss": 0.0636,
+      "step": 189
+    },
+    {
+      "epoch": 2.56,
+      "learning_rate": 2e-05,
+      "loss": 0.0667,
+      "step": 190
+    },
+    {
+      "epoch": 2.57,
+      "learning_rate": 2e-05,
+      "loss": 0.0688,
+      "step": 191
+    },
+    {
+      "epoch": 2.58,
+      "learning_rate": 2e-05,
+      "loss": 0.0708,
+      "step": 192
+    },
+    {
+      "epoch": 2.6,
+      "learning_rate": 2e-05,
+      "loss": 0.0666,
+      "step": 193
+    },
+    {
+      "epoch": 2.61,
+      "learning_rate": 2e-05,
+      "loss": 0.0679,
+      "step": 194
+    },
+    {
+      "epoch": 2.62,
+      "learning_rate": 2e-05,
+      "loss": 0.0675,
+      "step": 195
+    },
+    {
+      "epoch": 2.64,
+      "learning_rate": 2e-05,
+      "loss": 0.0688,
+      "step": 196
+    },
+    {
+      "epoch": 2.65,
+      "learning_rate": 2e-05,
+      "loss": 0.068,
+      "step": 197
+    },
+    {
+      "epoch": 2.66,
+      "learning_rate": 2e-05,
+      "loss": 0.0665,
+      "step": 198
+    },
+    {
+      "epoch": 2.68,
+      "learning_rate": 2e-05,
+      "loss": 0.0678,
+      "step": 199
+    },
+    {
+      "epoch": 2.69,
+      "learning_rate": 2e-05,
+      "loss": 0.0674,
+      "step": 200
+    },
+    {
+      "epoch": 2.7,
+      "learning_rate": 2e-05,
+      "loss": 0.0627,
+      "step": 201
+    },
+    {
+      "epoch": 2.72,
+      "learning_rate": 2e-05,
+      "loss": 0.0676,
+      "step": 202
+    },
+    {
+      "epoch": 2.73,
+      "learning_rate": 2e-05,
+      "loss": 0.0605,
+      "step": 203
+    },
+    {
+      "epoch": 2.75,
+      "learning_rate": 2e-05,
+      "loss": 0.0656,
+      "step": 204
+    },
+    {
+      "epoch": 2.76,
+      "learning_rate": 2e-05,
+      "loss": 0.072,
+      "step": 205
+    },
+    {
+      "epoch": 2.77,
+      "learning_rate": 2e-05,
+      "loss": 0.0635,
+      "step": 206
+    },
+    {
+      "epoch": 2.79,
+      "learning_rate": 2e-05,
+      "loss": 0.0624,
+      "step": 207
+    },
+    {
+      "epoch": 2.8,
+      "learning_rate": 2e-05,
+      "loss": 0.0628,
+      "step": 208
+    },
+    {
+      "epoch": 2.81,
+      "learning_rate": 2e-05,
+      "loss": 0.0704,
+      "step": 209
+    },
+    {
+      "epoch": 2.83,
+      "learning_rate": 2e-05,
+      "loss": 0.0647,
+      "step": 210
+    },
+    {
+      "epoch": 2.84,
+      "learning_rate": 2e-05,
+      "loss": 0.0611,
+      "step": 211
+    },
+    {
+      "epoch": 2.85,
+      "learning_rate": 2e-05,
+      "loss": 0.0686,
+      "step": 212
+    },
+    {
+      "epoch": 2.87,
+      "learning_rate": 2e-05,
+      "loss": 0.0686,
+      "step": 213
+    },
+    {
+      "epoch": 2.88,
+      "learning_rate": 2e-05,
+      "loss": 0.0723,
+      "step": 214
+    },
+    {
+      "epoch": 2.89,
+      "learning_rate": 2e-05,
+      "loss": 0.0644,
+      "step": 215
+    },
+    {
+      "epoch": 2.91,
+      "learning_rate": 2e-05,
+      "loss": 0.0702,
+      "step": 216
+    },
+    {
+      "epoch": 2.92,
+      "learning_rate": 2e-05,
+      "loss": 0.0624,
+      "step": 217
+    },
+    {
+      "epoch": 2.93,
+      "learning_rate": 2e-05,
+      "loss": 0.0664,
+      "step": 218
+    },
+    {
+      "epoch": 2.95,
+      "learning_rate": 2e-05,
+      "loss": 0.0592,
+      "step": 219
+    },
+    {
+      "epoch": 2.96,
+      "learning_rate": 2e-05,
+      "loss": 0.0565,
+      "step": 220
+    },
+    {
+      "epoch": 2.97,
+      "learning_rate": 2e-05,
+      "loss": 0.0679,
+      "step": 221
+    },
+    {
+      "epoch": 2.99,
+      "learning_rate": 2e-05,
+      "loss": 0.0617,
+      "step": 222
+    }
+  ],
+  "max_steps": 222,
+  "num_train_epochs": 3,
+  "total_flos": 31856813015040.0,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e04a090494b1b9ef7c1035d6a1a84d50dd4eedd9b6ceab3e135a3f316ba05776
+size 4923

zero_to_fp32.py

@@ -0,0 +1,483 @@
+#!/usr/bin/env python
+'''Copyright The Microsoft DeepSpeed Team'''
+
+# 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 math
+import os
+import re
+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.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION,
+                                            OPTIMIZER_STATE_DICT,
+                                            SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS,
+                                            ZERO_STAGE,
+                                            PARTITION_COUNT,
+                                            PARAM_SHAPES,
+                                            BUFFER_NAMES)
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+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")),
+                         key=natural_keys)
+
+    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
+    }
+    param_shapes = state_dict[PARAM_SHAPES]
+
+    ds_version = state_dict.get(DS_VERSION, None)
+
+    return buffers, param_shapes, ds_version
+
+
+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]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    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 zero_stage == 2:
+        fp32_flat_groups = [
+            state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
+            for i in range(len(state_dicts))
+        ]
+    elif zero_stage == 3:
+        # 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, fp32_flat_groups
+
+
+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, 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, param_shapes, ds_version = parse_model_state(model_file)
+    print(f'Parsing checkpoint created by deepspeed=={ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(
+                    f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum([
+        full_single_fp32_vector.numel()
+        for full_single_fp32_vector in merged_single_partition_of_fp32_groups
+    ])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum(
+            [sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # 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
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            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
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # 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 zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        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
+
+        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
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0,
+                                             offset,
+                                             partitioned_numel)
+                  for i in range(world_size)),
+            0).narrow(0,
+                      0,
+                      unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    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 usable 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 usable 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)