Add files using upload-large-folder tool

added_tokens.json

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

config.json

@@ -0,0 +1,28 @@
+{
+  "_name_or_path": "../models/TinyLlama_v1.1/",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 5632,
+  "max_position_embeddings": 2048,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 22,
+  "num_key_value_heads": 4,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.37.0",
+  "use_cache": true,
+  "vocab_size": 32001
+}

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "max_length": 2048,
+  "pad_token_id": 0,
+  "transformers_version": "4.37.0"
+}

global_step400/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7520e90678489f5d58a03fda7681a35f23f8e09b973a822ce895be50f07a2ee1
+size 13200649212

global_step400/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99743bf44afcaaa741db23f923e003456ce26a6a6ad6ff1db250d97f45ee8d88
+size 2200176684

latest

@@ -0,0 +1 @@
+global_step400

metrics.json

@@ -0,0 +1,32 @@
+{
+    "PPL": 11.989105224609375,
+    "arc_challenge": {
+        "acc": 0.2431740614334471,
+        "acc_stderr": 0.012536554144587092,
+        "acc_norm": 0.28071672354948807,
+        "acc_norm_stderr": 0.013131238126975583
+    },
+    "arc_easy": {
+        "acc": 0.4196127946127946,
+        "acc_stderr": 0.010126315840891551,
+        "acc_norm": 0.38846801346801346,
+        "acc_norm_stderr": 0.010001276044485233
+    },
+    "hellaswag": {
+        "acc": 0.3899621589324836,
+        "acc_stderr": 0.004867445945277156,
+        "acc_norm": 0.49790878311093406,
+        "acc_norm_stderr": 0.004989737768749952
+    },
+    "piqa": {
+        "acc": 0.6621327529923831,
+        "acc_stderr": 0.011035474307853836,
+        "acc_norm": 0.6692056583242655,
+        "acc_norm_stderr": 0.01097752058471443
+    },
+    "winogrande": {
+        "acc": 0.5027624309392266,
+        "acc_stderr": 0.014052271211616441
+    },
+    "QA Avg": 0.443528839782067
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f421b717964007fbbacd13e4ddd844f7c8571327de608341c6cc7250b1f2581
+size 2200128056

rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cc0337928392aec080563235f3789477dca72e8d5517982c85ed52eb0f6124b
+size 14244

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

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

tokenizer_config.json

@@ -0,0 +1,50 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32000": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": false,
+  "model_max_length": 1024,
+  "pad_token": "[PAD]",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

trainer_state.json

@@ -0,0 +1,2453 @@
+{
+  "best_metric": 101.20526885986328,
+  "best_model_checkpoint": "./ckpts/tinyllama_v1.1/int3-g128/checkpoint-400",
+  "epoch": 4.0,
+  "eval_steps": 100,
+  "global_step": 400,
+  "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": 256.4583,
+      "step": 1
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 2e-05,
+      "loss": 355.0836,
+      "step": 2
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 2e-05,
+      "loss": 252.4541,
+      "step": 3
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 2e-05,
+      "loss": 761.1344,
+      "step": 4
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 2e-05,
+      "loss": 465.4658,
+      "step": 5
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 2e-05,
+      "loss": 841.6429,
+      "step": 6
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 2e-05,
+      "loss": 753.775,
+      "step": 7
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 2e-05,
+      "loss": 379.0151,
+      "step": 8
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 2e-05,
+      "loss": 402.6746,
+      "step": 9
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 2e-05,
+      "loss": 216.2126,
+      "step": 10
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 2e-05,
+      "loss": 227.9789,
+      "step": 11
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 2e-05,
+      "loss": 237.3062,
+      "step": 12
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 2e-05,
+      "loss": 299.5192,
+      "step": 13
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 2e-05,
+      "loss": 186.9158,
+      "step": 14
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 2e-05,
+      "loss": 209.0085,
+      "step": 15
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 2e-05,
+      "loss": 151.9208,
+      "step": 16
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 2e-05,
+      "loss": 177.2832,
+      "step": 17
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 2e-05,
+      "loss": 180.6637,
+      "step": 18
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 2e-05,
+      "loss": 153.5752,
+      "step": 19
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 2e-05,
+      "loss": 241.4604,
+      "step": 20
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 2e-05,
+      "loss": 154.7056,
+      "step": 21
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 2e-05,
+      "loss": 152.9164,
+      "step": 22
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 2e-05,
+      "loss": 117.8652,
+      "step": 23
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 2e-05,
+      "loss": 139.5359,
+      "step": 24
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 2e-05,
+      "loss": 125.2291,
+      "step": 25
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 2e-05,
+      "loss": 153.2976,
+      "step": 26
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 2e-05,
+      "loss": 138.6838,
+      "step": 27
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 2e-05,
+      "loss": 147.1582,
+      "step": 28
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 2e-05,
+      "loss": 120.231,
+      "step": 29
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 2e-05,
+      "loss": 136.0073,
+      "step": 30
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 2e-05,
+      "loss": 125.3789,
+      "step": 31
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 2e-05,
+      "loss": 135.6559,
+      "step": 32
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 2e-05,
+      "loss": 151.9706,
+      "step": 33
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 2e-05,
+      "loss": 126.2539,
+      "step": 34
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 2e-05,
+      "loss": 110.8474,
+      "step": 35
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 2e-05,
+      "loss": 150.4755,
+      "step": 36
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 2e-05,
+      "loss": 133.6725,
+      "step": 37
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 2e-05,
+      "loss": 166.5977,
+      "step": 38
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 2e-05,
+      "loss": 167.9467,
+      "step": 39
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 2e-05,
+      "loss": 125.6363,
+      "step": 40
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 2e-05,
+      "loss": 129.4,
+      "step": 41
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 2e-05,
+      "loss": 109.1935,
+      "step": 42
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 2e-05,
+      "loss": 164.8529,
+      "step": 43
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 2e-05,
+      "loss": 127.5447,
+      "step": 44
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 2e-05,
+      "loss": 126.8883,
+      "step": 45
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 2e-05,
+      "loss": 156.4396,
+      "step": 46
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 2e-05,
+      "loss": 90.8995,
+      "step": 47
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 2e-05,
+      "loss": 131.5748,
+      "step": 48
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 2e-05,
+      "loss": 137.7466,
+      "step": 49
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 2e-05,
+      "loss": 105.3827,
+      "step": 50
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 2e-05,
+      "loss": 131.8921,
+      "step": 51
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 2e-05,
+      "loss": 112.9179,
+      "step": 52
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 2e-05,
+      "loss": 143.8214,
+      "step": 53
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 2e-05,
+      "loss": 168.2706,
+      "step": 54
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 2e-05,
+      "loss": 153.9795,
+      "step": 55
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 2e-05,
+      "loss": 115.3438,
+      "step": 56
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 2e-05,
+      "loss": 144.034,
+      "step": 57
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 2e-05,
+      "loss": 90.717,
+      "step": 58
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 2e-05,
+      "loss": 117.4731,
+      "step": 59
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 2e-05,
+      "loss": 143.9666,
+      "step": 60
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 2e-05,
+      "loss": 127.3566,
+      "step": 61
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 2e-05,
+      "loss": 107.2432,
+      "step": 62
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 2e-05,
+      "loss": 117.6042,
+      "step": 63
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 2e-05,
+      "loss": 92.6121,
+      "step": 64
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 2e-05,
+      "loss": 107.5126,
+      "step": 65
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 2e-05,
+      "loss": 133.897,
+      "step": 66
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 2e-05,
+      "loss": 101.0651,
+      "step": 67
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 2e-05,
+      "loss": 99.0262,
+      "step": 68
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 2e-05,
+      "loss": 156.479,
+      "step": 69
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 2e-05,
+      "loss": 172.746,
+      "step": 70
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 2e-05,
+      "loss": 104.9099,
+      "step": 71
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 2e-05,
+      "loss": 94.8904,
+      "step": 72
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 2e-05,
+      "loss": 105.7694,
+      "step": 73
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 2e-05,
+      "loss": 92.2346,
+      "step": 74
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 2e-05,
+      "loss": 118.597,
+      "step": 75
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 2e-05,
+      "loss": 139.2074,
+      "step": 76
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 2e-05,
+      "loss": 91.171,
+      "step": 77
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 2e-05,
+      "loss": 121.2573,
+      "step": 78
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 2e-05,
+      "loss": 91.8224,
+      "step": 79
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 2e-05,
+      "loss": 107.0462,
+      "step": 80
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 2e-05,
+      "loss": 98.3316,
+      "step": 81
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 2e-05,
+      "loss": 111.1321,
+      "step": 82
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 2e-05,
+      "loss": 77.9181,
+      "step": 83
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 2e-05,
+      "loss": 91.1592,
+      "step": 84
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 2e-05,
+      "loss": 125.3041,
+      "step": 85
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 2e-05,
+      "loss": 165.3387,
+      "step": 86
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 2e-05,
+      "loss": 84.1901,
+      "step": 87
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 2e-05,
+      "loss": 120.4,
+      "step": 88
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 2e-05,
+      "loss": 116.1603,
+      "step": 89
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 2e-05,
+      "loss": 102.2226,
+      "step": 90
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 2e-05,
+      "loss": 102.1463,
+      "step": 91
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 2e-05,
+      "loss": 156.9727,
+      "step": 92
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 2e-05,
+      "loss": 98.1492,
+      "step": 93
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 2e-05,
+      "loss": 89.7699,
+      "step": 94
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 2e-05,
+      "loss": 106.7468,
+      "step": 95
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 2e-05,
+      "loss": 87.7702,
+      "step": 96
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 2e-05,
+      "loss": 149.5999,
+      "step": 97
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 2e-05,
+      "loss": 105.2152,
+      "step": 98
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 2e-05,
+      "loss": 110.4924,
+      "step": 99
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 2e-05,
+      "loss": 159.247,
+      "step": 100
+    },
+    {
+      "epoch": 1.0,
+      "eval_loss": 127.25352478027344,
+      "eval_runtime": 80.8206,
+      "eval_samples_per_second": 19.747,
+      "eval_steps_per_second": 1.237,
+      "step": 100
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 2e-05,
+      "loss": 117.8373,
+      "step": 101
+    },
+    {
+      "epoch": 1.02,
+      "learning_rate": 2e-05,
+      "loss": 81.0658,
+      "step": 102
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 2e-05,
+      "loss": 79.7859,
+      "step": 103
+    },
+    {
+      "epoch": 1.04,
+      "learning_rate": 2e-05,
+      "loss": 168.3286,
+      "step": 104
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 2e-05,
+      "loss": 172.5758,
+      "step": 105
+    },
+    {
+      "epoch": 1.06,
+      "learning_rate": 2e-05,
+      "loss": 83.6487,
+      "step": 106
+    },
+    {
+      "epoch": 1.07,
+      "learning_rate": 2e-05,
+      "loss": 140.8672,
+      "step": 107
+    },
+    {
+      "epoch": 1.08,
+      "learning_rate": 2e-05,
+      "loss": 118.9715,
+      "step": 108
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 2e-05,
+      "loss": 102.4753,
+      "step": 109
+    },
+    {
+      "epoch": 1.1,
+      "learning_rate": 2e-05,
+      "loss": 90.8443,
+      "step": 110
+    },
+    {
+      "epoch": 1.11,
+      "learning_rate": 2e-05,
+      "loss": 108.1487,
+      "step": 111
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 2e-05,
+      "loss": 80.9609,
+      "step": 112
+    },
+    {
+      "epoch": 1.13,
+      "learning_rate": 2e-05,
+      "loss": 99.0464,
+      "step": 113
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 2e-05,
+      "loss": 100.1197,
+      "step": 114
+    },
+    {
+      "epoch": 1.15,
+      "learning_rate": 2e-05,
+      "loss": 110.2522,
+      "step": 115
+    },
+    {
+      "epoch": 1.16,
+      "learning_rate": 2e-05,
+      "loss": 75.1395,
+      "step": 116
+    },
+    {
+      "epoch": 1.17,
+      "learning_rate": 2e-05,
+      "loss": 170.2633,
+      "step": 117
+    },
+    {
+      "epoch": 1.18,
+      "learning_rate": 2e-05,
+      "loss": 96.691,
+      "step": 118
+    },
+    {
+      "epoch": 1.19,
+      "learning_rate": 2e-05,
+      "loss": 126.8882,
+      "step": 119
+    },
+    {
+      "epoch": 1.2,
+      "learning_rate": 2e-05,
+      "loss": 127.7979,
+      "step": 120
+    },
+    {
+      "epoch": 1.21,
+      "learning_rate": 2e-05,
+      "loss": 163.4338,
+      "step": 121
+    },
+    {
+      "epoch": 1.22,
+      "learning_rate": 2e-05,
+      "loss": 175.9516,
+      "step": 122
+    },
+    {
+      "epoch": 1.23,
+      "learning_rate": 2e-05,
+      "loss": 88.2406,
+      "step": 123
+    },
+    {
+      "epoch": 1.24,
+      "learning_rate": 2e-05,
+      "loss": 146.5514,
+      "step": 124
+    },
+    {
+      "epoch": 1.25,
+      "learning_rate": 2e-05,
+      "loss": 98.5611,
+      "step": 125
+    },
+    {
+      "epoch": 1.26,
+      "learning_rate": 2e-05,
+      "loss": 101.5959,
+      "step": 126
+    },
+    {
+      "epoch": 1.27,
+      "learning_rate": 2e-05,
+      "loss": 105.368,
+      "step": 127
+    },
+    {
+      "epoch": 1.28,
+      "learning_rate": 2e-05,
+      "loss": 130.4881,
+      "step": 128
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 2e-05,
+      "loss": 111.5857,
+      "step": 129
+    },
+    {
+      "epoch": 1.3,
+      "learning_rate": 2e-05,
+      "loss": 102.0928,
+      "step": 130
+    },
+    {
+      "epoch": 1.31,
+      "learning_rate": 2e-05,
+      "loss": 99.7146,
+      "step": 131
+    },
+    {
+      "epoch": 1.32,
+      "learning_rate": 2e-05,
+      "loss": 100.4634,
+      "step": 132
+    },
+    {
+      "epoch": 1.33,
+      "learning_rate": 2e-05,
+      "loss": 95.0521,
+      "step": 133
+    },
+    {
+      "epoch": 1.34,
+      "learning_rate": 2e-05,
+      "loss": 163.5093,
+      "step": 134
+    },
+    {
+      "epoch": 1.35,
+      "learning_rate": 2e-05,
+      "loss": 164.4761,
+      "step": 135
+    },
+    {
+      "epoch": 1.36,
+      "learning_rate": 2e-05,
+      "loss": 108.3719,
+      "step": 136
+    },
+    {
+      "epoch": 1.37,
+      "learning_rate": 2e-05,
+      "loss": 164.0315,
+      "step": 137
+    },
+    {
+      "epoch": 1.38,
+      "learning_rate": 2e-05,
+      "loss": 80.1498,
+      "step": 138
+    },
+    {
+      "epoch": 1.39,
+      "learning_rate": 2e-05,
+      "loss": 82.825,
+      "step": 139
+    },
+    {
+      "epoch": 1.4,
+      "learning_rate": 2e-05,
+      "loss": 119.8586,
+      "step": 140
+    },
+    {
+      "epoch": 1.41,
+      "learning_rate": 2e-05,
+      "loss": 118.1383,
+      "step": 141
+    },
+    {
+      "epoch": 1.42,
+      "learning_rate": 2e-05,
+      "loss": 145.6537,
+      "step": 142
+    },
+    {
+      "epoch": 1.43,
+      "learning_rate": 2e-05,
+      "loss": 131.2617,
+      "step": 143
+    },
+    {
+      "epoch": 1.44,
+      "learning_rate": 2e-05,
+      "loss": 122.246,
+      "step": 144
+    },
+    {
+      "epoch": 1.45,
+      "learning_rate": 2e-05,
+      "loss": 98.2258,
+      "step": 145
+    },
+    {
+      "epoch": 1.46,
+      "learning_rate": 2e-05,
+      "loss": 166.0939,
+      "step": 146
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 2e-05,
+      "loss": 112.9639,
+      "step": 147
+    },
+    {
+      "epoch": 1.48,
+      "learning_rate": 2e-05,
+      "loss": 122.0515,
+      "step": 148
+    },
+    {
+      "epoch": 1.49,
+      "learning_rate": 2e-05,
+      "loss": 140.3237,
+      "step": 149
+    },
+    {
+      "epoch": 1.5,
+      "learning_rate": 2e-05,
+      "loss": 101.4403,
+      "step": 150
+    },
+    {
+      "epoch": 1.51,
+      "learning_rate": 2e-05,
+      "loss": 95.5035,
+      "step": 151
+    },
+    {
+      "epoch": 1.52,
+      "learning_rate": 2e-05,
+      "loss": 145.9391,
+      "step": 152
+    },
+    {
+      "epoch": 1.53,
+      "learning_rate": 2e-05,
+      "loss": 105.1173,
+      "step": 153
+    },
+    {
+      "epoch": 1.54,
+      "learning_rate": 2e-05,
+      "loss": 105.8428,
+      "step": 154
+    },
+    {
+      "epoch": 1.55,
+      "learning_rate": 2e-05,
+      "loss": 85.6933,
+      "step": 155
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 2e-05,
+      "loss": 118.8751,
+      "step": 156
+    },
+    {
+      "epoch": 1.57,
+      "learning_rate": 2e-05,
+      "loss": 101.1875,
+      "step": 157
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 2e-05,
+      "loss": 105.5553,
+      "step": 158
+    },
+    {
+      "epoch": 1.59,
+      "learning_rate": 2e-05,
+      "loss": 112.3205,
+      "step": 159
+    },
+    {
+      "epoch": 1.6,
+      "learning_rate": 2e-05,
+      "loss": 143.0757,
+      "step": 160
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 2e-05,
+      "loss": 112.1303,
+      "step": 161
+    },
+    {
+      "epoch": 1.62,
+      "learning_rate": 2e-05,
+      "loss": 95.4685,
+      "step": 162
+    },
+    {
+      "epoch": 1.63,
+      "learning_rate": 2e-05,
+      "loss": 111.5353,
+      "step": 163
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 2e-05,
+      "loss": 121.3872,
+      "step": 164
+    },
+    {
+      "epoch": 1.65,
+      "learning_rate": 2e-05,
+      "loss": 90.6964,
+      "step": 165
+    },
+    {
+      "epoch": 1.66,
+      "learning_rate": 2e-05,
+      "loss": 96.5425,
+      "step": 166
+    },
+    {
+      "epoch": 1.67,
+      "learning_rate": 2e-05,
+      "loss": 112.4037,
+      "step": 167
+    },
+    {
+      "epoch": 1.68,
+      "learning_rate": 2e-05,
+      "loss": 143.2002,
+      "step": 168
+    },
+    {
+      "epoch": 1.69,
+      "learning_rate": 2e-05,
+      "loss": 114.4325,
+      "step": 169
+    },
+    {
+      "epoch": 1.7,
+      "learning_rate": 2e-05,
+      "loss": 116.9987,
+      "step": 170
+    },
+    {
+      "epoch": 1.71,
+      "learning_rate": 2e-05,
+      "loss": 148.152,
+      "step": 171
+    },
+    {
+      "epoch": 1.72,
+      "learning_rate": 2e-05,
+      "loss": 82.8326,
+      "step": 172
+    },
+    {
+      "epoch": 1.73,
+      "learning_rate": 2e-05,
+      "loss": 86.1253,
+      "step": 173
+    },
+    {
+      "epoch": 1.74,
+      "learning_rate": 2e-05,
+      "loss": 96.1883,
+      "step": 174
+    },
+    {
+      "epoch": 1.75,
+      "learning_rate": 2e-05,
+      "loss": 192.9743,
+      "step": 175
+    },
+    {
+      "epoch": 1.76,
+      "learning_rate": 2e-05,
+      "loss": 98.351,
+      "step": 176
+    },
+    {
+      "epoch": 1.77,
+      "learning_rate": 2e-05,
+      "loss": 361.8705,
+      "step": 177
+    },
+    {
+      "epoch": 1.78,
+      "learning_rate": 2e-05,
+      "loss": 102.5058,
+      "step": 178
+    },
+    {
+      "epoch": 1.79,
+      "learning_rate": 2e-05,
+      "loss": 154.9456,
+      "step": 179
+    },
+    {
+      "epoch": 1.8,
+      "learning_rate": 2e-05,
+      "loss": 86.3787,
+      "step": 180
+    },
+    {
+      "epoch": 1.81,
+      "learning_rate": 2e-05,
+      "loss": 139.9809,
+      "step": 181
+    },
+    {
+      "epoch": 1.82,
+      "learning_rate": 2e-05,
+      "loss": 112.6604,
+      "step": 182
+    },
+    {
+      "epoch": 1.83,
+      "learning_rate": 2e-05,
+      "loss": 238.6009,
+      "step": 183
+    },
+    {
+      "epoch": 1.84,
+      "learning_rate": 2e-05,
+      "loss": 102.7639,
+      "step": 184
+    },
+    {
+      "epoch": 1.85,
+      "learning_rate": 2e-05,
+      "loss": 146.489,
+      "step": 185
+    },
+    {
+      "epoch": 1.86,
+      "learning_rate": 2e-05,
+      "loss": 76.5674,
+      "step": 186
+    },
+    {
+      "epoch": 1.87,
+      "learning_rate": 2e-05,
+      "loss": 105.2609,
+      "step": 187
+    },
+    {
+      "epoch": 1.88,
+      "learning_rate": 2e-05,
+      "loss": 92.5275,
+      "step": 188
+    },
+    {
+      "epoch": 1.89,
+      "learning_rate": 2e-05,
+      "loss": 120.9601,
+      "step": 189
+    },
+    {
+      "epoch": 1.9,
+      "learning_rate": 2e-05,
+      "loss": 139.8529,
+      "step": 190
+    },
+    {
+      "epoch": 1.91,
+      "learning_rate": 2e-05,
+      "loss": 86.0494,
+      "step": 191
+    },
+    {
+      "epoch": 1.92,
+      "learning_rate": 2e-05,
+      "loss": 104.6433,
+      "step": 192
+    },
+    {
+      "epoch": 1.93,
+      "learning_rate": 2e-05,
+      "loss": 82.0844,
+      "step": 193
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 2e-05,
+      "loss": 98.7083,
+      "step": 194
+    },
+    {
+      "epoch": 1.95,
+      "learning_rate": 2e-05,
+      "loss": 103.9359,
+      "step": 195
+    },
+    {
+      "epoch": 1.96,
+      "learning_rate": 2e-05,
+      "loss": 138.1158,
+      "step": 196
+    },
+    {
+      "epoch": 1.97,
+      "learning_rate": 2e-05,
+      "loss": 122.3703,
+      "step": 197
+    },
+    {
+      "epoch": 1.98,
+      "learning_rate": 2e-05,
+      "loss": 92.0145,
+      "step": 198
+    },
+    {
+      "epoch": 1.99,
+      "learning_rate": 2e-05,
+      "loss": 86.6408,
+      "step": 199
+    },
+    {
+      "epoch": 2.0,
+      "learning_rate": 2e-05,
+      "loss": 118.3698,
+      "step": 200
+    },
+    {
+      "epoch": 2.0,
+      "eval_loss": 123.61127471923828,
+      "eval_runtime": 80.8471,
+      "eval_samples_per_second": 19.741,
+      "eval_steps_per_second": 1.237,
+      "step": 200
+    },
+    {
+      "epoch": 2.01,
+      "learning_rate": 2e-05,
+      "loss": 92.5948,
+      "step": 201
+    },
+    {
+      "epoch": 2.02,
+      "learning_rate": 2e-05,
+      "loss": 106.3745,
+      "step": 202
+    },
+    {
+      "epoch": 2.03,
+      "learning_rate": 2e-05,
+      "loss": 90.8155,
+      "step": 203
+    },
+    {
+      "epoch": 2.04,
+      "learning_rate": 2e-05,
+      "loss": 121.7541,
+      "step": 204
+    },
+    {
+      "epoch": 2.05,
+      "learning_rate": 2e-05,
+      "loss": 69.567,
+      "step": 205
+    },
+    {
+      "epoch": 2.06,
+      "learning_rate": 2e-05,
+      "loss": 112.1705,
+      "step": 206
+    },
+    {
+      "epoch": 2.07,
+      "learning_rate": 2e-05,
+      "loss": 99.2473,
+      "step": 207
+    },
+    {
+      "epoch": 2.08,
+      "learning_rate": 2e-05,
+      "loss": 109.7618,
+      "step": 208
+    },
+    {
+      "epoch": 2.09,
+      "learning_rate": 2e-05,
+      "loss": 89.1843,
+      "step": 209
+    },
+    {
+      "epoch": 2.1,
+      "learning_rate": 2e-05,
+      "loss": 133.0722,
+      "step": 210
+    },
+    {
+      "epoch": 2.11,
+      "learning_rate": 2e-05,
+      "loss": 70.0389,
+      "step": 211
+    },
+    {
+      "epoch": 2.12,
+      "learning_rate": 2e-05,
+      "loss": 127.7498,
+      "step": 212
+    },
+    {
+      "epoch": 2.13,
+      "learning_rate": 2e-05,
+      "loss": 175.2826,
+      "step": 213
+    },
+    {
+      "epoch": 2.14,
+      "learning_rate": 2e-05,
+      "loss": 98.1167,
+      "step": 214
+    },
+    {
+      "epoch": 2.15,
+      "learning_rate": 2e-05,
+      "loss": 131.4668,
+      "step": 215
+    },
+    {
+      "epoch": 2.16,
+      "learning_rate": 2e-05,
+      "loss": 163.0928,
+      "step": 216
+    },
+    {
+      "epoch": 2.17,
+      "learning_rate": 2e-05,
+      "loss": 139.924,
+      "step": 217
+    },
+    {
+      "epoch": 2.18,
+      "learning_rate": 2e-05,
+      "loss": 100.2071,
+      "step": 218
+    },
+    {
+      "epoch": 2.19,
+      "learning_rate": 2e-05,
+      "loss": 100.4305,
+      "step": 219
+    },
+    {
+      "epoch": 2.2,
+      "learning_rate": 2e-05,
+      "loss": 112.9075,
+      "step": 220
+    },
+    {
+      "epoch": 2.21,
+      "learning_rate": 2e-05,
+      "loss": 119.6214,
+      "step": 221
+    },
+    {
+      "epoch": 2.22,
+      "learning_rate": 2e-05,
+      "loss": 88.9599,
+      "step": 222
+    },
+    {
+      "epoch": 2.23,
+      "learning_rate": 2e-05,
+      "loss": 95.6708,
+      "step": 223
+    },
+    {
+      "epoch": 2.24,
+      "learning_rate": 2e-05,
+      "loss": 88.6898,
+      "step": 224
+    },
+    {
+      "epoch": 2.25,
+      "learning_rate": 2e-05,
+      "loss": 93.868,
+      "step": 225
+    },
+    {
+      "epoch": 2.26,
+      "learning_rate": 2e-05,
+      "loss": 108.3749,
+      "step": 226
+    },
+    {
+      "epoch": 2.27,
+      "learning_rate": 2e-05,
+      "loss": 94.327,
+      "step": 227
+    },
+    {
+      "epoch": 2.28,
+      "learning_rate": 2e-05,
+      "loss": 81.9183,
+      "step": 228
+    },
+    {
+      "epoch": 2.29,
+      "learning_rate": 2e-05,
+      "loss": 134.5717,
+      "step": 229
+    },
+    {
+      "epoch": 2.3,
+      "learning_rate": 2e-05,
+      "loss": 92.9158,
+      "step": 230
+    },
+    {
+      "epoch": 2.31,
+      "learning_rate": 2e-05,
+      "loss": 118.63,
+      "step": 231
+    },
+    {
+      "epoch": 2.32,
+      "learning_rate": 2e-05,
+      "loss": 90.0151,
+      "step": 232
+    },
+    {
+      "epoch": 2.33,
+      "learning_rate": 2e-05,
+      "loss": 128.4179,
+      "step": 233
+    },
+    {
+      "epoch": 2.34,
+      "learning_rate": 2e-05,
+      "loss": 132.9601,
+      "step": 234
+    },
+    {
+      "epoch": 2.35,
+      "learning_rate": 2e-05,
+      "loss": 110.689,
+      "step": 235
+    },
+    {
+      "epoch": 2.36,
+      "learning_rate": 2e-05,
+      "loss": 302.4031,
+      "step": 236
+    },
+    {
+      "epoch": 2.37,
+      "learning_rate": 2e-05,
+      "loss": 128.6338,
+      "step": 237
+    },
+    {
+      "epoch": 2.38,
+      "learning_rate": 2e-05,
+      "loss": 357.8953,
+      "step": 238
+    },
+    {
+      "epoch": 2.39,
+      "learning_rate": 2e-05,
+      "loss": 73.8348,
+      "step": 239
+    },
+    {
+      "epoch": 2.4,
+      "learning_rate": 2e-05,
+      "loss": 99.6502,
+      "step": 240
+    },
+    {
+      "epoch": 2.41,
+      "learning_rate": 2e-05,
+      "loss": 91.3185,
+      "step": 241
+    },
+    {
+      "epoch": 2.42,
+      "learning_rate": 2e-05,
+      "loss": 83.6264,
+      "step": 242
+    },
+    {
+      "epoch": 2.43,
+      "learning_rate": 2e-05,
+      "loss": 99.7315,
+      "step": 243
+    },
+    {
+      "epoch": 2.44,
+      "learning_rate": 2e-05,
+      "loss": 74.5451,
+      "step": 244
+    },
+    {
+      "epoch": 2.45,
+      "learning_rate": 2e-05,
+      "loss": 86.7471,
+      "step": 245
+    },
+    {
+      "epoch": 2.46,
+      "learning_rate": 2e-05,
+      "loss": 92.1487,
+      "step": 246
+    },
+    {
+      "epoch": 2.47,
+      "learning_rate": 2e-05,
+      "loss": 71.9365,
+      "step": 247
+    },
+    {
+      "epoch": 2.48,
+      "learning_rate": 2e-05,
+      "loss": 85.449,
+      "step": 248
+    },
+    {
+      "epoch": 2.49,
+      "learning_rate": 2e-05,
+      "loss": 74.9259,
+      "step": 249
+    },
+    {
+      "epoch": 2.5,
+      "learning_rate": 2e-05,
+      "loss": 212.1755,
+      "step": 250
+    },
+    {
+      "epoch": 2.51,
+      "learning_rate": 2e-05,
+      "loss": 141.1738,
+      "step": 251
+    },
+    {
+      "epoch": 2.52,
+      "learning_rate": 2e-05,
+      "loss": 68.8899,
+      "step": 252
+    },
+    {
+      "epoch": 2.53,
+      "learning_rate": 2e-05,
+      "loss": 89.9649,
+      "step": 253
+    },
+    {
+      "epoch": 2.54,
+      "learning_rate": 2e-05,
+      "loss": 85.3244,
+      "step": 254
+    },
+    {
+      "epoch": 2.55,
+      "learning_rate": 2e-05,
+      "loss": 102.01,
+      "step": 255
+    },
+    {
+      "epoch": 2.56,
+      "learning_rate": 2e-05,
+      "loss": 87.7782,
+      "step": 256
+    },
+    {
+      "epoch": 2.57,
+      "learning_rate": 2e-05,
+      "loss": 98.1192,
+      "step": 257
+    },
+    {
+      "epoch": 2.58,
+      "learning_rate": 2e-05,
+      "loss": 96.4169,
+      "step": 258
+    },
+    {
+      "epoch": 2.59,
+      "learning_rate": 2e-05,
+      "loss": 79.6971,
+      "step": 259
+    },
+    {
+      "epoch": 2.6,
+      "learning_rate": 2e-05,
+      "loss": 102.9386,
+      "step": 260
+    },
+    {
+      "epoch": 2.61,
+      "learning_rate": 2e-05,
+      "loss": 140.3832,
+      "step": 261
+    },
+    {
+      "epoch": 2.62,
+      "learning_rate": 2e-05,
+      "loss": 85.5271,
+      "step": 262
+    },
+    {
+      "epoch": 2.63,
+      "learning_rate": 2e-05,
+      "loss": 95.7376,
+      "step": 263
+    },
+    {
+      "epoch": 2.64,
+      "learning_rate": 2e-05,
+      "loss": 75.8506,
+      "step": 264
+    },
+    {
+      "epoch": 2.65,
+      "learning_rate": 2e-05,
+      "loss": 72.1343,
+      "step": 265
+    },
+    {
+      "epoch": 2.66,
+      "learning_rate": 2e-05,
+      "loss": 86.7606,
+      "step": 266
+    },
+    {
+      "epoch": 2.67,
+      "learning_rate": 2e-05,
+      "loss": 219.3665,
+      "step": 267
+    },
+    {
+      "epoch": 2.68,
+      "learning_rate": 2e-05,
+      "loss": 72.1039,
+      "step": 268
+    },
+    {
+      "epoch": 2.69,
+      "learning_rate": 2e-05,
+      "loss": 93.7625,
+      "step": 269
+    },
+    {
+      "epoch": 2.7,
+      "learning_rate": 2e-05,
+      "loss": 72.9779,
+      "step": 270
+    },
+    {
+      "epoch": 2.71,
+      "learning_rate": 2e-05,
+      "loss": 76.5516,
+      "step": 271
+    },
+    {
+      "epoch": 2.72,
+      "learning_rate": 2e-05,
+      "loss": 96.9595,
+      "step": 272
+    },
+    {
+      "epoch": 2.73,
+      "learning_rate": 2e-05,
+      "loss": 86.1967,
+      "step": 273
+    },
+    {
+      "epoch": 2.74,
+      "learning_rate": 2e-05,
+      "loss": 151.0705,
+      "step": 274
+    },
+    {
+      "epoch": 2.75,
+      "learning_rate": 2e-05,
+      "loss": 110.2601,
+      "step": 275
+    },
+    {
+      "epoch": 2.76,
+      "learning_rate": 2e-05,
+      "loss": 76.5183,
+      "step": 276
+    },
+    {
+      "epoch": 2.77,
+      "learning_rate": 2e-05,
+      "loss": 82.5286,
+      "step": 277
+    },
+    {
+      "epoch": 2.78,
+      "learning_rate": 2e-05,
+      "loss": 77.1184,
+      "step": 278
+    },
+    {
+      "epoch": 2.79,
+      "learning_rate": 2e-05,
+      "loss": 71.968,
+      "step": 279
+    },
+    {
+      "epoch": 2.8,
+      "learning_rate": 2e-05,
+      "loss": 84.9687,
+      "step": 280
+    },
+    {
+      "epoch": 2.81,
+      "learning_rate": 2e-05,
+      "loss": 82.9308,
+      "step": 281
+    },
+    {
+      "epoch": 2.82,
+      "learning_rate": 2e-05,
+      "loss": 75.31,
+      "step": 282
+    },
+    {
+      "epoch": 2.83,
+      "learning_rate": 2e-05,
+      "loss": 92.4162,
+      "step": 283
+    },
+    {
+      "epoch": 2.84,
+      "learning_rate": 2e-05,
+      "loss": 147.4037,
+      "step": 284
+    },
+    {
+      "epoch": 2.85,
+      "learning_rate": 2e-05,
+      "loss": 158.5738,
+      "step": 285
+    },
+    {
+      "epoch": 2.86,
+      "learning_rate": 2e-05,
+      "loss": 74.0794,
+      "step": 286
+    },
+    {
+      "epoch": 2.87,
+      "learning_rate": 2e-05,
+      "loss": 76.1675,
+      "step": 287
+    },
+    {
+      "epoch": 2.88,
+      "learning_rate": 2e-05,
+      "loss": 119.9409,
+      "step": 288
+    },
+    {
+      "epoch": 2.89,
+      "learning_rate": 2e-05,
+      "loss": 83.3971,
+      "step": 289
+    },
+    {
+      "epoch": 2.9,
+      "learning_rate": 2e-05,
+      "loss": 123.4943,
+      "step": 290
+    },
+    {
+      "epoch": 2.91,
+      "learning_rate": 2e-05,
+      "loss": 163.733,
+      "step": 291
+    },
+    {
+      "epoch": 2.92,
+      "learning_rate": 2e-05,
+      "loss": 84.0254,
+      "step": 292
+    },
+    {
+      "epoch": 2.93,
+      "learning_rate": 2e-05,
+      "loss": 100.9542,
+      "step": 293
+    },
+    {
+      "epoch": 2.94,
+      "learning_rate": 2e-05,
+      "loss": 79.933,
+      "step": 294
+    },
+    {
+      "epoch": 2.95,
+      "learning_rate": 2e-05,
+      "loss": 88.4858,
+      "step": 295
+    },
+    {
+      "epoch": 2.96,
+      "learning_rate": 2e-05,
+      "loss": 94.7497,
+      "step": 296
+    },
+    {
+      "epoch": 2.97,
+      "learning_rate": 2e-05,
+      "loss": 119.0287,
+      "step": 297
+    },
+    {
+      "epoch": 2.98,
+      "learning_rate": 2e-05,
+      "loss": 127.9053,
+      "step": 298
+    },
+    {
+      "epoch": 2.99,
+      "learning_rate": 2e-05,
+      "loss": 88.5058,
+      "step": 299
+    },
+    {
+      "epoch": 3.0,
+      "learning_rate": 2e-05,
+      "loss": 86.8415,
+      "step": 300
+    },
+    {
+      "epoch": 3.0,
+      "eval_loss": 119.90803527832031,
+      "eval_runtime": 80.9383,
+      "eval_samples_per_second": 19.719,
+      "eval_steps_per_second": 1.236,
+      "step": 300
+    },
+    {
+      "epoch": 3.01,
+      "learning_rate": 2e-05,
+      "loss": 103.3673,
+      "step": 301
+    },
+    {
+      "epoch": 3.02,
+      "learning_rate": 2e-05,
+      "loss": 107.5574,
+      "step": 302
+    },
+    {
+      "epoch": 3.03,
+      "learning_rate": 2e-05,
+      "loss": 95.606,
+      "step": 303
+    },
+    {
+      "epoch": 3.04,
+      "learning_rate": 2e-05,
+      "loss": 78.7149,
+      "step": 304
+    },
+    {
+      "epoch": 3.05,
+      "learning_rate": 2e-05,
+      "loss": 72.7483,
+      "step": 305
+    },
+    {
+      "epoch": 3.06,
+      "learning_rate": 2e-05,
+      "loss": 337.7869,
+      "step": 306
+    },
+    {
+      "epoch": 3.07,
+      "learning_rate": 2e-05,
+      "loss": 81.3133,
+      "step": 307
+    },
+    {
+      "epoch": 3.08,
+      "learning_rate": 2e-05,
+      "loss": 95.4,
+      "step": 308
+    },
+    {
+      "epoch": 3.09,
+      "learning_rate": 2e-05,
+      "loss": 117.039,
+      "step": 309
+    },
+    {
+      "epoch": 3.1,
+      "learning_rate": 2e-05,
+      "loss": 83.7925,
+      "step": 310
+    },
+    {
+      "epoch": 3.11,
+      "learning_rate": 2e-05,
+      "loss": 80.7962,
+      "step": 311
+    },
+    {
+      "epoch": 3.12,
+      "learning_rate": 2e-05,
+      "loss": 163.7045,
+      "step": 312
+    },
+    {
+      "epoch": 3.13,
+      "learning_rate": 2e-05,
+      "loss": 90.9779,
+      "step": 313
+    },
+    {
+      "epoch": 3.14,
+      "learning_rate": 2e-05,
+      "loss": 79.9667,
+      "step": 314
+    },
+    {
+      "epoch": 3.15,
+      "learning_rate": 2e-05,
+      "loss": 91.632,
+      "step": 315
+    },
+    {
+      "epoch": 3.16,
+      "learning_rate": 2e-05,
+      "loss": 69.9252,
+      "step": 316
+    },
+    {
+      "epoch": 3.17,
+      "learning_rate": 2e-05,
+      "loss": 94.818,
+      "step": 317
+    },
+    {
+      "epoch": 3.18,
+      "learning_rate": 2e-05,
+      "loss": 68.6146,
+      "step": 318
+    },
+    {
+      "epoch": 3.19,
+      "learning_rate": 2e-05,
+      "loss": 76.8548,
+      "step": 319
+    },
+    {
+      "epoch": 3.2,
+      "learning_rate": 2e-05,
+      "loss": 186.7993,
+      "step": 320
+    },
+    {
+      "epoch": 3.21,
+      "learning_rate": 2e-05,
+      "loss": 79.4541,
+      "step": 321
+    },
+    {
+      "epoch": 3.22,
+      "learning_rate": 2e-05,
+      "loss": 91.6424,
+      "step": 322
+    },
+    {
+      "epoch": 3.23,
+      "learning_rate": 2e-05,
+      "loss": 100.9761,
+      "step": 323
+    },
+    {
+      "epoch": 3.24,
+      "learning_rate": 2e-05,
+      "loss": 117.9019,
+      "step": 324
+    },
+    {
+      "epoch": 3.25,
+      "learning_rate": 2e-05,
+      "loss": 83.9062,
+      "step": 325
+    },
+    {
+      "epoch": 3.26,
+      "learning_rate": 2e-05,
+      "loss": 77.0838,
+      "step": 326
+    },
+    {
+      "epoch": 3.27,
+      "learning_rate": 2e-05,
+      "loss": 108.4668,
+      "step": 327
+    },
+    {
+      "epoch": 3.28,
+      "learning_rate": 2e-05,
+      "loss": 137.5493,
+      "step": 328
+    },
+    {
+      "epoch": 3.29,
+      "learning_rate": 2e-05,
+      "loss": 86.924,
+      "step": 329
+    },
+    {
+      "epoch": 3.3,
+      "learning_rate": 2e-05,
+      "loss": 101.1962,
+      "step": 330
+    },
+    {
+      "epoch": 3.31,
+      "learning_rate": 2e-05,
+      "loss": 81.7145,
+      "step": 331
+    },
+    {
+      "epoch": 3.32,
+      "learning_rate": 2e-05,
+      "loss": 106.9531,
+      "step": 332
+    },
+    {
+      "epoch": 3.33,
+      "learning_rate": 2e-05,
+      "loss": 89.0009,
+      "step": 333
+    },
+    {
+      "epoch": 3.34,
+      "learning_rate": 2e-05,
+      "loss": 111.5955,
+      "step": 334
+    },
+    {
+      "epoch": 3.35,
+      "learning_rate": 2e-05,
+      "loss": 108.5236,
+      "step": 335
+    },
+    {
+      "epoch": 3.36,
+      "learning_rate": 2e-05,
+      "loss": 78.6253,
+      "step": 336
+    },
+    {
+      "epoch": 3.37,
+      "learning_rate": 2e-05,
+      "loss": 92.5898,
+      "step": 337
+    },
+    {
+      "epoch": 3.38,
+      "learning_rate": 2e-05,
+      "loss": 73.9657,
+      "step": 338
+    },
+    {
+      "epoch": 3.39,
+      "learning_rate": 2e-05,
+      "loss": 76.5264,
+      "step": 339
+    },
+    {
+      "epoch": 3.4,
+      "learning_rate": 2e-05,
+      "loss": 76.9044,
+      "step": 340
+    },
+    {
+      "epoch": 3.41,
+      "learning_rate": 2e-05,
+      "loss": 73.2403,
+      "step": 341
+    },
+    {
+      "epoch": 3.42,
+      "learning_rate": 2e-05,
+      "loss": 94.6415,
+      "step": 342
+    },
+    {
+      "epoch": 3.43,
+      "learning_rate": 2e-05,
+      "loss": 76.2357,
+      "step": 343
+    },
+    {
+      "epoch": 3.44,
+      "learning_rate": 2e-05,
+      "loss": 76.6514,
+      "step": 344
+    },
+    {
+      "epoch": 3.45,
+      "learning_rate": 2e-05,
+      "loss": 65.6005,
+      "step": 345
+    },
+    {
+      "epoch": 3.46,
+      "learning_rate": 2e-05,
+      "loss": 108.3323,
+      "step": 346
+    },
+    {
+      "epoch": 3.47,
+      "learning_rate": 2e-05,
+      "loss": 93.3479,
+      "step": 347
+    },
+    {
+      "epoch": 3.48,
+      "learning_rate": 2e-05,
+      "loss": 90.2143,
+      "step": 348
+    },
+    {
+      "epoch": 3.49,
+      "learning_rate": 2e-05,
+      "loss": 127.5598,
+      "step": 349
+    },
+    {
+      "epoch": 3.5,
+      "learning_rate": 2e-05,
+      "loss": 86.6066,
+      "step": 350
+    },
+    {
+      "epoch": 3.51,
+      "learning_rate": 2e-05,
+      "loss": 73.5306,
+      "step": 351
+    },
+    {
+      "epoch": 3.52,
+      "learning_rate": 2e-05,
+      "loss": 60.8274,
+      "step": 352
+    },
+    {
+      "epoch": 3.53,
+      "learning_rate": 2e-05,
+      "loss": 75.1176,
+      "step": 353
+    },
+    {
+      "epoch": 3.54,
+      "learning_rate": 2e-05,
+      "loss": 108.5266,
+      "step": 354
+    },
+    {
+      "epoch": 3.55,
+      "learning_rate": 2e-05,
+      "loss": 69.0899,
+      "step": 355
+    },
+    {
+      "epoch": 3.56,
+      "learning_rate": 2e-05,
+      "loss": 92.8721,
+      "step": 356
+    },
+    {
+      "epoch": 3.57,
+      "learning_rate": 2e-05,
+      "loss": 90.7292,
+      "step": 357
+    },
+    {
+      "epoch": 3.58,
+      "learning_rate": 2e-05,
+      "loss": 55.9025,
+      "step": 358
+    },
+    {
+      "epoch": 3.59,
+      "learning_rate": 2e-05,
+      "loss": 80.1136,
+      "step": 359
+    },
+    {
+      "epoch": 3.6,
+      "learning_rate": 2e-05,
+      "loss": 129.3794,
+      "step": 360
+    },
+    {
+      "epoch": 3.61,
+      "learning_rate": 2e-05,
+      "loss": 65.8789,
+      "step": 361
+    },
+    {
+      "epoch": 3.62,
+      "learning_rate": 2e-05,
+      "loss": 89.3556,
+      "step": 362
+    },
+    {
+      "epoch": 3.63,
+      "learning_rate": 2e-05,
+      "loss": 104.0831,
+      "step": 363
+    },
+    {
+      "epoch": 3.64,
+      "learning_rate": 2e-05,
+      "loss": 90.5861,
+      "step": 364
+    },
+    {
+      "epoch": 3.65,
+      "learning_rate": 2e-05,
+      "loss": 70.7352,
+      "step": 365
+    },
+    {
+      "epoch": 3.66,
+      "learning_rate": 2e-05,
+      "loss": 113.6718,
+      "step": 366
+    },
+    {
+      "epoch": 3.67,
+      "learning_rate": 2e-05,
+      "loss": 81.7421,
+      "step": 367
+    },
+    {
+      "epoch": 3.68,
+      "learning_rate": 2e-05,
+      "loss": 59.3722,
+      "step": 368
+    },
+    {
+      "epoch": 3.69,
+      "learning_rate": 2e-05,
+      "loss": 74.7749,
+      "step": 369
+    },
+    {
+      "epoch": 3.7,
+      "learning_rate": 2e-05,
+      "loss": 98.0698,
+      "step": 370
+    },
+    {
+      "epoch": 3.71,
+      "learning_rate": 2e-05,
+      "loss": 81.5395,
+      "step": 371
+    },
+    {
+      "epoch": 3.72,
+      "learning_rate": 2e-05,
+      "loss": 70.3278,
+      "step": 372
+    },
+    {
+      "epoch": 3.73,
+      "learning_rate": 2e-05,
+      "loss": 77.4071,
+      "step": 373
+    },
+    {
+      "epoch": 3.74,
+      "learning_rate": 2e-05,
+      "loss": 203.0746,
+      "step": 374
+    },
+    {
+      "epoch": 3.75,
+      "learning_rate": 2e-05,
+      "loss": 248.7437,
+      "step": 375
+    },
+    {
+      "epoch": 3.76,
+      "learning_rate": 2e-05,
+      "loss": 73.0192,
+      "step": 376
+    },
+    {
+      "epoch": 3.77,
+      "learning_rate": 2e-05,
+      "loss": 72.3248,
+      "step": 377
+    },
+    {
+      "epoch": 3.78,
+      "learning_rate": 2e-05,
+      "loss": 77.9258,
+      "step": 378
+    },
+    {
+      "epoch": 3.79,
+      "learning_rate": 2e-05,
+      "loss": 72.0043,
+      "step": 379
+    },
+    {
+      "epoch": 3.8,
+      "learning_rate": 2e-05,
+      "loss": 112.4846,
+      "step": 380
+    },
+    {
+      "epoch": 3.81,
+      "learning_rate": 2e-05,
+      "loss": 69.6211,
+      "step": 381
+    },
+    {
+      "epoch": 3.82,
+      "learning_rate": 2e-05,
+      "loss": 89.8621,
+      "step": 382
+    },
+    {
+      "epoch": 3.83,
+      "learning_rate": 2e-05,
+      "loss": 75.9898,
+      "step": 383
+    },
+    {
+      "epoch": 3.84,
+      "learning_rate": 2e-05,
+      "loss": 122.9664,
+      "step": 384
+    },
+    {
+      "epoch": 3.85,
+      "learning_rate": 2e-05,
+      "loss": 76.6402,
+      "step": 385
+    },
+    {
+      "epoch": 3.86,
+      "learning_rate": 2e-05,
+      "loss": 88.9508,
+      "step": 386
+    },
+    {
+      "epoch": 3.87,
+      "learning_rate": 2e-05,
+      "loss": 89.0049,
+      "step": 387
+    },
+    {
+      "epoch": 3.88,
+      "learning_rate": 2e-05,
+      "loss": 81.3426,
+      "step": 388
+    },
+    {
+      "epoch": 3.89,
+      "learning_rate": 2e-05,
+      "loss": 70.0686,
+      "step": 389
+    },
+    {
+      "epoch": 3.9,
+      "learning_rate": 2e-05,
+      "loss": 87.9824,
+      "step": 390
+    },
+    {
+      "epoch": 3.91,
+      "learning_rate": 2e-05,
+      "loss": 82.3193,
+      "step": 391
+    },
+    {
+      "epoch": 3.92,
+      "learning_rate": 2e-05,
+      "loss": 89.7547,
+      "step": 392
+    },
+    {
+      "epoch": 3.93,
+      "learning_rate": 2e-05,
+      "loss": 61.2356,
+      "step": 393
+    },
+    {
+      "epoch": 3.94,
+      "learning_rate": 2e-05,
+      "loss": 76.2497,
+      "step": 394
+    },
+    {
+      "epoch": 3.95,
+      "learning_rate": 2e-05,
+      "loss": 90.9001,
+      "step": 395
+    },
+    {
+      "epoch": 3.96,
+      "learning_rate": 2e-05,
+      "loss": 74.1517,
+      "step": 396
+    },
+    {
+      "epoch": 3.97,
+      "learning_rate": 2e-05,
+      "loss": 74.2053,
+      "step": 397
+    },
+    {
+      "epoch": 3.98,
+      "learning_rate": 2e-05,
+      "loss": 63.4142,
+      "step": 398
+    },
+    {
+      "epoch": 3.99,
+      "learning_rate": 2e-05,
+      "loss": 99.6599,
+      "step": 399
+    },
+    {
+      "epoch": 4.0,
+      "learning_rate": 2e-05,
+      "loss": 80.2025,
+      "step": 400
+    },
+    {
+      "epoch": 4.0,
+      "eval_loss": 101.20526885986328,
+      "eval_runtime": 80.9879,
+      "eval_samples_per_second": 19.707,
+      "eval_steps_per_second": 1.235,
+      "step": 400
+    }
+  ],
+  "logging_steps": 1.0,
+  "max_steps": 400,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 20,
+  "total_flos": 1.624099600371548e+17,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be461beb564711b9b63a9bd3ce6ea9ad668099d4f862360e41ab1000eb4b384c
+size 6392

zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 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 . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# 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,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+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_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        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]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
+    """
+    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_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # 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.")
+
+    # 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() if _has_callable(shape, 'numel') else math.prod(shape)
+            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")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    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 _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # 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
+
+    # 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
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        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"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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 Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    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_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+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_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    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_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)