First model version

added_tokens.json

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

config.json

@@ -0,0 +1,23 @@
+{
+  "_name_or_path": "../../../PLMs/ft_local/LLAMA/hf/llama-7b",
+  "architectures": [
+    "LLaMAForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_sequence_length": 2048,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.27.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32001
+}

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 0,
+  "transformers_version": "4.27.0.dev0"
+}

latest

@@ -0,0 +1 @@
+global_step234

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da612e2a757293805d3d904876bb6b134a00fe0ef32b15f042631abb3266cdb4
+size 26953810889

special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "</s>",
+  "eos_token": "</s>",
+  "pad_token": "[PAD]",
+  "unk_token": "</s>"
+}

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": "",
+  "eos_token": "",
+  "model_max_length": 512,
+  "padding_side": "right",
+  "special_tokens_map_file": "../../../PLMs/ft_local/LLAMA/hf/llama-7b/special_tokens_map.json",
+  "tokenizer_class": "LLaMATokenizer",
+  "unk_token": ""
+}

trainer_state.json

@@ -0,0 +1,1429 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.9984,
+  "global_step": 234,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0,
+      "loss": 1.0183,
+      "step": 1
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 6.666666666666668e-06,
+      "loss": 0.8252,
+      "step": 2
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 1.0566416671474378e-05,
+      "loss": 0.8853,
+      "step": 3
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 1.3333333333333337e-05,
+      "loss": 0.6674,
+      "step": 4
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 1.5479520632582417e-05,
+      "loss": 0.7124,
+      "step": 5
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 1.7233083338141044e-05,
+      "loss": 0.6497,
+      "step": 6
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1.8715699480384028e-05,
+      "loss": 0.6226,
+      "step": 7
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 2e-05,
+      "loss": 0.6205,
+      "step": 8
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 2e-05,
+      "loss": 0.6256,
+      "step": 9
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1.991150442477876e-05,
+      "loss": 0.6366,
+      "step": 10
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1.9823008849557524e-05,
+      "loss": 0.5717,
+      "step": 11
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1.9734513274336283e-05,
+      "loss": 0.6073,
+      "step": 12
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1.9646017699115046e-05,
+      "loss": 0.6069,
+      "step": 13
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1.9557522123893806e-05,
+      "loss": 0.5996,
+      "step": 14
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1.946902654867257e-05,
+      "loss": 0.5885,
+      "step": 15
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1.9380530973451328e-05,
+      "loss": 0.5838,
+      "step": 16
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1.929203539823009e-05,
+      "loss": 0.6094,
+      "step": 17
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1.9203539823008853e-05,
+      "loss": 0.5957,
+      "step": 18
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1.9115044247787613e-05,
+      "loss": 0.625,
+      "step": 19
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1.9026548672566376e-05,
+      "loss": 0.5876,
+      "step": 20
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1.8938053097345135e-05,
+      "loss": 0.5664,
+      "step": 21
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1.8849557522123894e-05,
+      "loss": 0.5474,
+      "step": 22
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1.8761061946902657e-05,
+      "loss": 0.5727,
+      "step": 23
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1.8672566371681417e-05,
+      "loss": 0.5708,
+      "step": 24
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1.858407079646018e-05,
+      "loss": 0.5623,
+      "step": 25
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1.849557522123894e-05,
+      "loss": 0.5544,
+      "step": 26
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1.8407079646017702e-05,
+      "loss": 0.5877,
+      "step": 27
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1.831858407079646e-05,
+      "loss": 0.572,
+      "step": 28
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1.823008849557522e-05,
+      "loss": 0.5372,
+      "step": 29
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1.8141592920353983e-05,
+      "loss": 0.5291,
+      "step": 30
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1.8053097345132743e-05,
+      "loss": 0.5574,
+      "step": 31
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 1.7964601769911506e-05,
+      "loss": 0.5402,
+      "step": 32
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 1.7876106194690265e-05,
+      "loss": 0.5336,
+      "step": 33
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 1.7787610619469028e-05,
+      "loss": 0.5763,
+      "step": 34
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 1.769911504424779e-05,
+      "loss": 0.558,
+      "step": 35
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 1.761061946902655e-05,
+      "loss": 0.5398,
+      "step": 36
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 1.7522123893805313e-05,
+      "loss": 0.5782,
+      "step": 37
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 1.7433628318584072e-05,
+      "loss": 0.5383,
+      "step": 38
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 1.7345132743362835e-05,
+      "loss": 0.5754,
+      "step": 39
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 1.7256637168141594e-05,
+      "loss": 0.5317,
+      "step": 40
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 1.7168141592920354e-05,
+      "loss": 0.5415,
+      "step": 41
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 1.7079646017699117e-05,
+      "loss": 0.5137,
+      "step": 42
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 1.6991150442477876e-05,
+      "loss": 0.5404,
+      "step": 43
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 1.690265486725664e-05,
+      "loss": 0.5129,
+      "step": 44
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 1.68141592920354e-05,
+      "loss": 0.5379,
+      "step": 45
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 1.672566371681416e-05,
+      "loss": 0.5112,
+      "step": 46
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1.663716814159292e-05,
+      "loss": 0.4948,
+      "step": 47
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 1.6548672566371683e-05,
+      "loss": 0.5621,
+      "step": 48
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 1.6460176991150443e-05,
+      "loss": 0.503,
+      "step": 49
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 1.6371681415929206e-05,
+      "loss": 0.538,
+      "step": 50
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1.628318584070797e-05,
+      "loss": 0.5087,
+      "step": 51
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1.6194690265486728e-05,
+      "loss": 0.5171,
+      "step": 52
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 1.6106194690265487e-05,
+      "loss": 0.5458,
+      "step": 53
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1.601769911504425e-05,
+      "loss": 0.5558,
+      "step": 54
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 1.592920353982301e-05,
+      "loss": 0.5022,
+      "step": 55
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1.5840707964601772e-05,
+      "loss": 0.5301,
+      "step": 56
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 1.5752212389380532e-05,
+      "loss": 0.533,
+      "step": 57
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1.5663716814159295e-05,
+      "loss": 0.5295,
+      "step": 58
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1.5575221238938054e-05,
+      "loss": 0.5146,
+      "step": 59
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1.5486725663716813e-05,
+      "loss": 0.4989,
+      "step": 60
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1.5398230088495576e-05,
+      "loss": 0.5032,
+      "step": 61
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1.5309734513274336e-05,
+      "loss": 0.5173,
+      "step": 62
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 1.5221238938053098e-05,
+      "loss": 0.506,
+      "step": 63
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 1.513274336283186e-05,
+      "loss": 0.4967,
+      "step": 64
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1.5044247787610619e-05,
+      "loss": 0.5163,
+      "step": 65
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 1.4955752212389383e-05,
+      "loss": 0.5079,
+      "step": 66
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1.4867256637168143e-05,
+      "loss": 0.5342,
+      "step": 67
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1.4778761061946904e-05,
+      "loss": 0.4903,
+      "step": 68
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 1.4690265486725665e-05,
+      "loss": 0.5185,
+      "step": 69
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 1.4601769911504426e-05,
+      "loss": 0.5127,
+      "step": 70
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 1.4513274336283187e-05,
+      "loss": 0.5051,
+      "step": 71
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1.4424778761061948e-05,
+      "loss": 0.5188,
+      "step": 72
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 1.433628318584071e-05,
+      "loss": 0.487,
+      "step": 73
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 1.424778761061947e-05,
+      "loss": 0.4731,
+      "step": 74
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1.4159292035398232e-05,
+      "loss": 0.4858,
+      "step": 75
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 1.4070796460176991e-05,
+      "loss": 0.4963,
+      "step": 76
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 1.3982300884955752e-05,
+      "loss": 0.4749,
+      "step": 77
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 1.3893805309734513e-05,
+      "loss": 0.4761,
+      "step": 78
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 1.3805309734513275e-05,
+      "loss": 0.4184,
+      "step": 79
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 1.3716814159292036e-05,
+      "loss": 0.3704,
+      "step": 80
+    },
+    {
+      "epoch": 1.04,
+      "learning_rate": 1.3628318584070797e-05,
+      "loss": 0.3434,
+      "step": 81
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 1.353982300884956e-05,
+      "loss": 0.369,
+      "step": 82
+    },
+    {
+      "epoch": 1.06,
+      "learning_rate": 1.345132743362832e-05,
+      "loss": 0.363,
+      "step": 83
+    },
+    {
+      "epoch": 1.08,
+      "learning_rate": 1.3362831858407082e-05,
+      "loss": 0.3729,
+      "step": 84
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 1.3274336283185843e-05,
+      "loss": 0.3468,
+      "step": 85
+    },
+    {
+      "epoch": 1.1,
+      "learning_rate": 1.3185840707964604e-05,
+      "loss": 0.326,
+      "step": 86
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 1.3097345132743363e-05,
+      "loss": 0.3541,
+      "step": 87
+    },
+    {
+      "epoch": 1.13,
+      "learning_rate": 1.3008849557522125e-05,
+      "loss": 0.3624,
+      "step": 88
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 1.2920353982300886e-05,
+      "loss": 0.3518,
+      "step": 89
+    },
+    {
+      "epoch": 1.15,
+      "learning_rate": 1.2831858407079647e-05,
+      "loss": 0.3352,
+      "step": 90
+    },
+    {
+      "epoch": 1.17,
+      "learning_rate": 1.2743362831858408e-05,
+      "loss": 0.3412,
+      "step": 91
+    },
+    {
+      "epoch": 1.18,
+      "learning_rate": 1.2654867256637169e-05,
+      "loss": 0.3465,
+      "step": 92
+    },
+    {
+      "epoch": 1.19,
+      "learning_rate": 1.256637168141593e-05,
+      "loss": 0.3776,
+      "step": 93
+    },
+    {
+      "epoch": 1.2,
+      "learning_rate": 1.2477876106194691e-05,
+      "loss": 0.3375,
+      "step": 94
+    },
+    {
+      "epoch": 1.22,
+      "learning_rate": 1.2389380530973452e-05,
+      "loss": 0.3535,
+      "step": 95
+    },
+    {
+      "epoch": 1.23,
+      "learning_rate": 1.2300884955752212e-05,
+      "loss": 0.3318,
+      "step": 96
+    },
+    {
+      "epoch": 1.24,
+      "learning_rate": 1.2212389380530973e-05,
+      "loss": 0.3501,
+      "step": 97
+    },
+    {
+      "epoch": 1.26,
+      "learning_rate": 1.2123893805309736e-05,
+      "loss": 0.3417,
+      "step": 98
+    },
+    {
+      "epoch": 1.27,
+      "learning_rate": 1.2035398230088497e-05,
+      "loss": 0.3368,
+      "step": 99
+    },
+    {
+      "epoch": 1.28,
+      "learning_rate": 1.1946902654867258e-05,
+      "loss": 0.3467,
+      "step": 100
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 1.1858407079646019e-05,
+      "loss": 0.3315,
+      "step": 101
+    },
+    {
+      "epoch": 1.31,
+      "learning_rate": 1.176991150442478e-05,
+      "loss": 0.3402,
+      "step": 102
+    },
+    {
+      "epoch": 1.32,
+      "learning_rate": 1.1681415929203541e-05,
+      "loss": 0.3269,
+      "step": 103
+    },
+    {
+      "epoch": 1.33,
+      "learning_rate": 1.1592920353982302e-05,
+      "loss": 0.3422,
+      "step": 104
+    },
+    {
+      "epoch": 1.35,
+      "learning_rate": 1.1504424778761064e-05,
+      "loss": 0.3521,
+      "step": 105
+    },
+    {
+      "epoch": 1.36,
+      "learning_rate": 1.1415929203539825e-05,
+      "loss": 0.3424,
+      "step": 106
+    },
+    {
+      "epoch": 1.37,
+      "learning_rate": 1.1327433628318584e-05,
+      "loss": 0.3503,
+      "step": 107
+    },
+    {
+      "epoch": 1.38,
+      "learning_rate": 1.1238938053097345e-05,
+      "loss": 0.3461,
+      "step": 108
+    },
+    {
+      "epoch": 1.4,
+      "learning_rate": 1.1150442477876106e-05,
+      "loss": 0.3353,
+      "step": 109
+    },
+    {
+      "epoch": 1.41,
+      "learning_rate": 1.1061946902654867e-05,
+      "loss": 0.3391,
+      "step": 110
+    },
+    {
+      "epoch": 1.42,
+      "learning_rate": 1.0973451327433629e-05,
+      "loss": 0.3136,
+      "step": 111
+    },
+    {
+      "epoch": 1.44,
+      "learning_rate": 1.088495575221239e-05,
+      "loss": 0.3493,
+      "step": 112
+    },
+    {
+      "epoch": 1.45,
+      "learning_rate": 1.079646017699115e-05,
+      "loss": 0.3405,
+      "step": 113
+    },
+    {
+      "epoch": 1.46,
+      "learning_rate": 1.0707964601769914e-05,
+      "loss": 0.3495,
+      "step": 114
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 1.0619469026548675e-05,
+      "loss": 0.328,
+      "step": 115
+    },
+    {
+      "epoch": 1.49,
+      "learning_rate": 1.0530973451327436e-05,
+      "loss": 0.3232,
+      "step": 116
+    },
+    {
+      "epoch": 1.5,
+      "learning_rate": 1.0442477876106197e-05,
+      "loss": 0.3452,
+      "step": 117
+    },
+    {
+      "epoch": 1.51,
+      "learning_rate": 1.0353982300884956e-05,
+      "loss": 0.3263,
+      "step": 118
+    },
+    {
+      "epoch": 1.52,
+      "learning_rate": 1.0265486725663717e-05,
+      "loss": 0.3339,
+      "step": 119
+    },
+    {
+      "epoch": 1.54,
+      "learning_rate": 1.0176991150442479e-05,
+      "loss": 0.3299,
+      "step": 120
+    },
+    {
+      "epoch": 1.55,
+      "learning_rate": 1.008849557522124e-05,
+      "loss": 0.3195,
+      "step": 121
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 1e-05,
+      "loss": 0.3475,
+      "step": 122
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 9.911504424778762e-06,
+      "loss": 0.3392,
+      "step": 123
+    },
+    {
+      "epoch": 1.59,
+      "learning_rate": 9.823008849557523e-06,
+      "loss": 0.3495,
+      "step": 124
+    },
+    {
+      "epoch": 1.6,
+      "learning_rate": 9.734513274336284e-06,
+      "loss": 0.3285,
+      "step": 125
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 9.646017699115045e-06,
+      "loss": 0.3412,
+      "step": 126
+    },
+    {
+      "epoch": 1.63,
+      "learning_rate": 9.557522123893806e-06,
+      "loss": 0.3351,
+      "step": 127
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 9.469026548672568e-06,
+      "loss": 0.3498,
+      "step": 128
+    },
+    {
+      "epoch": 1.65,
+      "learning_rate": 9.380530973451329e-06,
+      "loss": 0.3475,
+      "step": 129
+    },
+    {
+      "epoch": 1.67,
+      "learning_rate": 9.29203539823009e-06,
+      "loss": 0.3495,
+      "step": 130
+    },
+    {
+      "epoch": 1.68,
+      "learning_rate": 9.203539823008851e-06,
+      "loss": 0.3332,
+      "step": 131
+    },
+    {
+      "epoch": 1.69,
+      "learning_rate": 9.11504424778761e-06,
+      "loss": 0.3136,
+      "step": 132
+    },
+    {
+      "epoch": 1.7,
+      "learning_rate": 9.026548672566371e-06,
+      "loss": 0.3526,
+      "step": 133
+    },
+    {
+      "epoch": 1.72,
+      "learning_rate": 8.938053097345133e-06,
+      "loss": 0.3336,
+      "step": 134
+    },
+    {
+      "epoch": 1.73,
+      "learning_rate": 8.849557522123895e-06,
+      "loss": 0.3476,
+      "step": 135
+    },
+    {
+      "epoch": 1.74,
+      "learning_rate": 8.761061946902656e-06,
+      "loss": 0.3286,
+      "step": 136
+    },
+    {
+      "epoch": 1.76,
+      "learning_rate": 8.672566371681418e-06,
+      "loss": 0.3579,
+      "step": 137
+    },
+    {
+      "epoch": 1.77,
+      "learning_rate": 8.584070796460177e-06,
+      "loss": 0.3202,
+      "step": 138
+    },
+    {
+      "epoch": 1.78,
+      "learning_rate": 8.495575221238938e-06,
+      "loss": 0.3376,
+      "step": 139
+    },
+    {
+      "epoch": 1.79,
+      "learning_rate": 8.4070796460177e-06,
+      "loss": 0.3399,
+      "step": 140
+    },
+    {
+      "epoch": 1.81,
+      "learning_rate": 8.31858407079646e-06,
+      "loss": 0.3296,
+      "step": 141
+    },
+    {
+      "epoch": 1.82,
+      "learning_rate": 8.230088495575221e-06,
+      "loss": 0.3342,
+      "step": 142
+    },
+    {
+      "epoch": 1.83,
+      "learning_rate": 8.141592920353984e-06,
+      "loss": 0.3272,
+      "step": 143
+    },
+    {
+      "epoch": 1.84,
+      "learning_rate": 8.053097345132744e-06,
+      "loss": 0.3242,
+      "step": 144
+    },
+    {
+      "epoch": 1.86,
+      "learning_rate": 7.964601769911505e-06,
+      "loss": 0.3207,
+      "step": 145
+    },
+    {
+      "epoch": 1.87,
+      "learning_rate": 7.876106194690266e-06,
+      "loss": 0.3148,
+      "step": 146
+    },
+    {
+      "epoch": 1.88,
+      "learning_rate": 7.787610619469027e-06,
+      "loss": 0.3344,
+      "step": 147
+    },
+    {
+      "epoch": 1.9,
+      "learning_rate": 7.699115044247788e-06,
+      "loss": 0.3228,
+      "step": 148
+    },
+    {
+      "epoch": 1.91,
+      "learning_rate": 7.610619469026549e-06,
+      "loss": 0.3498,
+      "step": 149
+    },
+    {
+      "epoch": 1.92,
+      "learning_rate": 7.5221238938053095e-06,
+      "loss": 0.3427,
+      "step": 150
+    },
+    {
+      "epoch": 1.93,
+      "learning_rate": 7.4336283185840714e-06,
+      "loss": 0.346,
+      "step": 151
+    },
+    {
+      "epoch": 1.95,
+      "learning_rate": 7.3451327433628326e-06,
+      "loss": 0.3053,
+      "step": 152
+    },
+    {
+      "epoch": 1.96,
+      "learning_rate": 7.256637168141594e-06,
+      "loss": 0.3306,
+      "step": 153
+    },
+    {
+      "epoch": 1.97,
+      "learning_rate": 7.168141592920355e-06,
+      "loss": 0.3232,
+      "step": 154
+    },
+    {
+      "epoch": 1.99,
+      "learning_rate": 7.079646017699116e-06,
+      "loss": 0.3138,
+      "step": 155
+    },
+    {
+      "epoch": 2.0,
+      "learning_rate": 6.991150442477876e-06,
+      "loss": 0.3067,
+      "step": 156
+    },
+    {
+      "epoch": 2.01,
+      "learning_rate": 6.902654867256637e-06,
+      "loss": 0.2733,
+      "step": 157
+    },
+    {
+      "epoch": 2.03,
+      "learning_rate": 6.814159292035398e-06,
+      "loss": 0.2254,
+      "step": 158
+    },
+    {
+      "epoch": 2.04,
+      "learning_rate": 6.72566371681416e-06,
+      "loss": 0.2198,
+      "step": 159
+    },
+    {
+      "epoch": 2.05,
+      "learning_rate": 6.6371681415929215e-06,
+      "loss": 0.2098,
+      "step": 160
+    },
+    {
+      "epoch": 2.06,
+      "learning_rate": 6.548672566371682e-06,
+      "loss": 0.2004,
+      "step": 161
+    },
+    {
+      "epoch": 2.08,
+      "learning_rate": 6.460176991150443e-06,
+      "loss": 0.1913,
+      "step": 162
+    },
+    {
+      "epoch": 2.09,
+      "learning_rate": 6.371681415929204e-06,
+      "loss": 0.188,
+      "step": 163
+    },
+    {
+      "epoch": 2.1,
+      "learning_rate": 6.283185840707965e-06,
+      "loss": 0.1956,
+      "step": 164
+    },
+    {
+      "epoch": 2.12,
+      "learning_rate": 6.194690265486726e-06,
+      "loss": 0.1968,
+      "step": 165
+    },
+    {
+      "epoch": 2.13,
+      "learning_rate": 6.1061946902654865e-06,
+      "loss": 0.1919,
+      "step": 166
+    },
+    {
+      "epoch": 2.14,
+      "learning_rate": 6.0176991150442484e-06,
+      "loss": 0.1946,
+      "step": 167
+    },
+    {
+      "epoch": 2.15,
+      "learning_rate": 5.9292035398230096e-06,
+      "loss": 0.1955,
+      "step": 168
+    },
+    {
+      "epoch": 2.17,
+      "learning_rate": 5.840707964601771e-06,
+      "loss": 0.209,
+      "step": 169
+    },
+    {
+      "epoch": 2.18,
+      "learning_rate": 5.752212389380532e-06,
+      "loss": 0.199,
+      "step": 170
+    },
+    {
+      "epoch": 2.19,
+      "learning_rate": 5.663716814159292e-06,
+      "loss": 0.2018,
+      "step": 171
+    },
+    {
+      "epoch": 2.2,
+      "learning_rate": 5.575221238938053e-06,
+      "loss": 0.2105,
+      "step": 172
+    },
+    {
+      "epoch": 2.22,
+      "learning_rate": 5.486725663716814e-06,
+      "loss": 0.1831,
+      "step": 173
+    },
+    {
+      "epoch": 2.23,
+      "learning_rate": 5.398230088495575e-06,
+      "loss": 0.1989,
+      "step": 174
+    },
+    {
+      "epoch": 2.24,
+      "learning_rate": 5.309734513274337e-06,
+      "loss": 0.1882,
+      "step": 175
+    },
+    {
+      "epoch": 2.26,
+      "learning_rate": 5.2212389380530985e-06,
+      "loss": 0.1949,
+      "step": 176
+    },
+    {
+      "epoch": 2.27,
+      "learning_rate": 5.132743362831859e-06,
+      "loss": 0.1932,
+      "step": 177
+    },
+    {
+      "epoch": 2.28,
+      "learning_rate": 5.04424778761062e-06,
+      "loss": 0.1874,
+      "step": 178
+    },
+    {
+      "epoch": 2.29,
+      "learning_rate": 4.955752212389381e-06,
+      "loss": 0.1855,
+      "step": 179
+    },
+    {
+      "epoch": 2.31,
+      "learning_rate": 4.867256637168142e-06,
+      "loss": 0.1982,
+      "step": 180
+    },
+    {
+      "epoch": 2.32,
+      "learning_rate": 4.778761061946903e-06,
+      "loss": 0.1927,
+      "step": 181
+    },
+    {
+      "epoch": 2.33,
+      "learning_rate": 4.690265486725664e-06,
+      "loss": 0.2008,
+      "step": 182
+    },
+    {
+      "epoch": 2.35,
+      "learning_rate": 4.6017699115044254e-06,
+      "loss": 0.2016,
+      "step": 183
+    },
+    {
+      "epoch": 2.36,
+      "learning_rate": 4.513274336283186e-06,
+      "loss": 0.2026,
+      "step": 184
+    },
+    {
+      "epoch": 2.37,
+      "learning_rate": 4.424778761061948e-06,
+      "loss": 0.1817,
+      "step": 185
+    },
+    {
+      "epoch": 2.38,
+      "learning_rate": 4.336283185840709e-06,
+      "loss": 0.1809,
+      "step": 186
+    },
+    {
+      "epoch": 2.4,
+      "learning_rate": 4.247787610619469e-06,
+      "loss": 0.1942,
+      "step": 187
+    },
+    {
+      "epoch": 2.41,
+      "learning_rate": 4.15929203539823e-06,
+      "loss": 0.1849,
+      "step": 188
+    },
+    {
+      "epoch": 2.42,
+      "learning_rate": 4.070796460176992e-06,
+      "loss": 0.1933,
+      "step": 189
+    },
+    {
+      "epoch": 2.44,
+      "learning_rate": 3.982300884955752e-06,
+      "loss": 0.189,
+      "step": 190
+    },
+    {
+      "epoch": 2.45,
+      "learning_rate": 3.8938053097345135e-06,
+      "loss": 0.1914,
+      "step": 191
+    },
+    {
+      "epoch": 2.46,
+      "learning_rate": 3.8053097345132746e-06,
+      "loss": 0.2014,
+      "step": 192
+    },
+    {
+      "epoch": 2.47,
+      "learning_rate": 3.7168141592920357e-06,
+      "loss": 0.1784,
+      "step": 193
+    },
+    {
+      "epoch": 2.49,
+      "learning_rate": 3.628318584070797e-06,
+      "loss": 0.1835,
+      "step": 194
+    },
+    {
+      "epoch": 2.5,
+      "learning_rate": 3.539823008849558e-06,
+      "loss": 0.1961,
+      "step": 195
+    },
+    {
+      "epoch": 2.51,
+      "learning_rate": 3.4513274336283186e-06,
+      "loss": 0.1843,
+      "step": 196
+    },
+    {
+      "epoch": 2.52,
+      "learning_rate": 3.36283185840708e-06,
+      "loss": 0.1688,
+      "step": 197
+    },
+    {
+      "epoch": 2.54,
+      "learning_rate": 3.274336283185841e-06,
+      "loss": 0.1861,
+      "step": 198
+    },
+    {
+      "epoch": 2.55,
+      "learning_rate": 3.185840707964602e-06,
+      "loss": 0.1757,
+      "step": 199
+    },
+    {
+      "epoch": 2.56,
+      "learning_rate": 3.097345132743363e-06,
+      "loss": 0.1934,
+      "step": 200
+    },
+    {
+      "epoch": 2.58,
+      "learning_rate": 3.0088495575221242e-06,
+      "loss": 0.1946,
+      "step": 201
+    },
+    {
+      "epoch": 2.59,
+      "learning_rate": 2.9203539823008853e-06,
+      "loss": 0.1836,
+      "step": 202
+    },
+    {
+      "epoch": 2.6,
+      "learning_rate": 2.831858407079646e-06,
+      "loss": 0.1891,
+      "step": 203
+    },
+    {
+      "epoch": 2.61,
+      "learning_rate": 2.743362831858407e-06,
+      "loss": 0.1825,
+      "step": 204
+    },
+    {
+      "epoch": 2.63,
+      "learning_rate": 2.6548672566371687e-06,
+      "loss": 0.1862,
+      "step": 205
+    },
+    {
+      "epoch": 2.64,
+      "learning_rate": 2.5663716814159294e-06,
+      "loss": 0.1921,
+      "step": 206
+    },
+    {
+      "epoch": 2.65,
+      "learning_rate": 2.4778761061946905e-06,
+      "loss": 0.192,
+      "step": 207
+    },
+    {
+      "epoch": 2.67,
+      "learning_rate": 2.3893805309734516e-06,
+      "loss": 0.1883,
+      "step": 208
+    },
+    {
+      "epoch": 2.68,
+      "learning_rate": 2.3008849557522127e-06,
+      "loss": 0.1726,
+      "step": 209
+    },
+    {
+      "epoch": 2.69,
+      "learning_rate": 2.212389380530974e-06,
+      "loss": 0.189,
+      "step": 210
+    },
+    {
+      "epoch": 2.7,
+      "learning_rate": 2.1238938053097345e-06,
+      "loss": 0.1882,
+      "step": 211
+    },
+    {
+      "epoch": 2.72,
+      "learning_rate": 2.035398230088496e-06,
+      "loss": 0.1927,
+      "step": 212
+    },
+    {
+      "epoch": 2.73,
+      "learning_rate": 1.9469026548672567e-06,
+      "loss": 0.193,
+      "step": 213
+    },
+    {
+      "epoch": 2.74,
+      "learning_rate": 1.8584070796460179e-06,
+      "loss": 0.1785,
+      "step": 214
+    },
+    {
+      "epoch": 2.76,
+      "learning_rate": 1.769911504424779e-06,
+      "loss": 0.1867,
+      "step": 215
+    },
+    {
+      "epoch": 2.77,
+      "learning_rate": 1.68141592920354e-06,
+      "loss": 0.1893,
+      "step": 216
+    },
+    {
+      "epoch": 2.78,
+      "learning_rate": 1.592920353982301e-06,
+      "loss": 0.1937,
+      "step": 217
+    },
+    {
+      "epoch": 2.79,
+      "learning_rate": 1.5044247787610621e-06,
+      "loss": 0.1797,
+      "step": 218
+    },
+    {
+      "epoch": 2.81,
+      "learning_rate": 1.415929203539823e-06,
+      "loss": 0.187,
+      "step": 219
+    },
+    {
+      "epoch": 2.82,
+      "learning_rate": 1.3274336283185843e-06,
+      "loss": 0.1785,
+      "step": 220
+    },
+    {
+      "epoch": 2.83,
+      "learning_rate": 1.2389380530973452e-06,
+      "loss": 0.1759,
+      "step": 221
+    },
+    {
+      "epoch": 2.84,
+      "learning_rate": 1.1504424778761064e-06,
+      "loss": 0.1886,
+      "step": 222
+    },
+    {
+      "epoch": 2.86,
+      "learning_rate": 1.0619469026548673e-06,
+      "loss": 0.1737,
+      "step": 223
+    },
+    {
+      "epoch": 2.87,
+      "learning_rate": 9.734513274336284e-07,
+      "loss": 0.1817,
+      "step": 224
+    },
+    {
+      "epoch": 2.88,
+      "learning_rate": 8.849557522123895e-07,
+      "loss": 0.1955,
+      "step": 225
+    },
+    {
+      "epoch": 2.9,
+      "learning_rate": 7.964601769911505e-07,
+      "loss": 0.1836,
+      "step": 226
+    },
+    {
+      "epoch": 2.91,
+      "learning_rate": 7.079646017699115e-07,
+      "loss": 0.1815,
+      "step": 227
+    },
+    {
+      "epoch": 2.92,
+      "learning_rate": 6.194690265486726e-07,
+      "loss": 0.1889,
+      "step": 228
+    },
+    {
+      "epoch": 2.93,
+      "learning_rate": 5.309734513274336e-07,
+      "loss": 0.1777,
+      "step": 229
+    },
+    {
+      "epoch": 2.95,
+      "learning_rate": 4.4247787610619474e-07,
+      "loss": 0.1763,
+      "step": 230
+    },
+    {
+      "epoch": 2.96,
+      "learning_rate": 3.5398230088495575e-07,
+      "loss": 0.1764,
+      "step": 231
+    },
+    {
+      "epoch": 2.97,
+      "learning_rate": 2.654867256637168e-07,
+      "loss": 0.1984,
+      "step": 232
+    },
+    {
+      "epoch": 2.99,
+      "learning_rate": 1.7699115044247788e-07,
+      "loss": 0.1937,
+      "step": 233
+    },
+    {
+      "epoch": 3.0,
+      "learning_rate": 8.849557522123894e-08,
+      "loss": 0.1798,
+      "step": 234
+    },
+    {
+      "epoch": 3.0,
+      "step": 234,
+      "total_flos": 13371387494400.0,
+      "train_loss": 0.3646234936184353,
+      "train_runtime": 3952.8658,
+      "train_samples_per_second": 7.589,
+      "train_steps_per_second": 0.059
+    }
+  ],
+  "max_steps": 234,
+  "num_train_epochs": 3,
+  "total_flos": 13371387494400.0,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfb5e3e1be86383d751db93e294c38999cb0a4be334fcf0d79e0af3c5a7de986
+size 5051

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)