First training attempt

README.md

@@ -0,0 +1,84 @@
+---
+base_model: /src/tanner/chess-roberta-code/model_configs/chess_roberta.json
+tags:
+- generated_from_trainer
+datasets:
+- TannerGladson/chess-roberta-pretraining
+metrics:
+- accuracy
+model-index:
+- name: 2024.09.24-01.27
+  results:
+  - task:
+      name: Masked Language Modeling
+      type: fill-mask
+    dataset:
+      name: TannerGladson/chess-roberta-pretraining
+      type: TannerGladson/chess-roberta-pretraining
+    metrics:
+    - name: Accuracy
+      type: accuracy
+      value: 0.5445640378503339
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="200" height="32"/>](https://wandb.ai/tanner-gladson/huggingface/runs/cs121ioj)
+# 2024.09.24-01.27
+
+This model is a fine-tuned version of [/src/tanner/chess-roberta-code/model_configs/chess_roberta.json](https://huggingface.co//src/tanner/chess-roberta-code/model_configs/chess_roberta.json) on the TannerGladson/chess-roberta-pretraining dataset.
+It achieves the following results on the evaluation set:
+- Loss: 2.1202
+- Accuracy: 0.5446
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0005
+- train_batch_size: 8
+- eval_batch_size: 8
+- seed: 42
+- distributed_type: multi-GPU
+- gradient_accumulation_steps: 16
+- total_train_batch_size: 128
+- optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-06
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 200
+- training_steps: 9000
+
+### Training results
+
+| Training Loss | Epoch  | Step | Validation Loss | Accuracy |
+|:-------------:|:------:|:----:|:---------------:|:--------:|
+| 2.1201        | 0.0120 | 1000 | 2.1195          | 0.5445   |
+| 2.1222        | 0.0239 | 2000 | 2.1199          | 0.5446   |
+| 2.1179        | 0.0359 | 3000 | 2.1199          | 0.5444   |
+| 2.1215        | 0.0479 | 4000 | 2.1186          | 0.5446   |
+| 2.1176        | 0.0598 | 5000 | 2.1194          | 0.5445   |
+| 2.1189        | 0.0718 | 6000 | 2.1193          | 0.5445   |
+| 2.1184        | 0.0838 | 7000 | 2.1194          | 0.5446   |
+| 2.119         | 0.0957 | 8000 | 2.1190          | 0.5446   |
+| 2.1237        | 0.1077 | 9000 | 2.1201          | 0.5446   |
+
+
+### Framework versions
+
+- Transformers 4.42.4
+- Pytorch 2.0.1+cu117
+- Datasets 2.17.1
+- Tokenizers 0.19.1

all_results.json

@@ -0,0 +1,16 @@
+{
+    "epoch": 0.10768528974073271,
+    "eval_accuracy": 0.5445640378503339,
+    "eval_loss": 2.1201696395874023,
+    "eval_runtime": 8310.7479,
+    "eval_samples": 562811,
+    "eval_samples_per_second": 67.721,
+    "eval_steps_per_second": 8.465,
+    "perplexity": 8.332550898295501,
+    "total_flos": 1158274171797504.0,
+    "train_loss": 2.120608181423611,
+    "train_runtime": 152416.4466,
+    "train_samples": 10697834,
+    "train_samples_per_second": 7.558,
+    "train_steps_per_second": 0.059
+}

checkpoint-9000/config.json

@@ -0,0 +1,27 @@
+{
+  "_name_or_path": "/src/tanner/chess-roberta-code/model_configs/chess_roberta.json",
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "eos_token_id": 1,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 1030,
+  "model_type": "roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 16,
+  "pad_token_id": 2,
+  "position_embedding_type": "absolute",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.42.4",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 64
+}

checkpoint-9000/global_step9000/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c279dc883c0d65945b83146c6a1901c3265ee01b398b33b0e234c652ee2d446
+size 1378093111

checkpoint-9000/global_step9000/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78aa760df334a408b1ff1cad6785692d84af6c1c49652716d7cfd0b815914075
+size 140897

checkpoint-9000/latest

@@ -0,0 +1 @@
+global_step9000

checkpoint-9000/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76bca5d4fb610fbdcd81ca77a69064a3272c24f56e6e5af9d5f760e817a866fe
+size 14575

checkpoint-9000/special_tokens_map.json

@@ -0,0 +1,37 @@
+{
+  "cls_token": {
+    "content": "[CLS]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "[SEP]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-9000/tokenizer.json

@@ -0,0 +1,162 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "[CLS]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "[SEP]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "[PAD]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "[UNK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 4,
+      "content": "[MASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 5,
+      "content": "[ILLEGAL]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 6,
+      "content": "~",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": {
+    "type": "Sequence",
+    "normalizers": [
+      {
+        "type": "NFD"
+      },
+      {
+        "type": "StripAccents"
+      }
+    ]
+  },
+  "pre_tokenizer": {
+    "type": "Split",
+    "pattern": {
+      "String": ""
+    },
+    "behavior": "Isolated",
+    "invert": false
+  },
+  "post_processor": {
+    "type": "RobertaProcessing",
+    "sep": [
+      "[SEP]",
+      1
+    ],
+    "cls": [
+      "[CLS]",
+      0
+    ],
+    "trim_offsets": false,
+    "add_prefix_space": false
+  },
+  "decoder": {
+    "type": "WordPiece",
+    "prefix": "##",
+    "cleanup": true
+  },
+  "model": {
+    "type": "WordPiece",
+    "unk_token": "[UNK]",
+    "continuing_subword_prefix": "##",
+    "max_input_chars_per_word": 100,
+    "vocab": {
+      "[CLS]": 0,
+      "[SEP]": 1,
+      "[PAD]": 2,
+      "[UNK]": 3,
+      "[MASK]": 4,
+      "[ILLEGAL]": 5,
+      "~": 6,
+      ">": 7,
+      " ": 8,
+      "#": 9,
+      "+": 10,
+      "-": 11,
+      "/": 12,
+      "0": 13,
+      "1": 14,
+      "2": 15,
+      "3": 16,
+      "4": 17,
+      "5": 18,
+      "6": 19,
+      "7": 20,
+      "8": 21,
+      "9": 22,
+      "=": 23,
+      "B": 24,
+      "K": 25,
+      "N": 26,
+      "O": 27,
+      "P": 28,
+      "Q": 29,
+      "R": 30,
+      "a": 31,
+      "b": 32,
+      "c": 33,
+      "d": 34,
+      "e": 35,
+      "f": 36,
+      "g": 37,
+      "h": 38,
+      "k": 39,
+      "n": 40,
+      "p": 41,
+      "q": 42,
+      "r": 43,
+      "w": 44,
+      "x": 45,
+      "_": 46
+    }
+  }
+}

checkpoint-9000/tokenizer_config.json

@@ -0,0 +1,68 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "[ILLEGAL]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "~",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "[UNK]"
+}

checkpoint-9000/trainer_state.json

@@ -0,0 +1,744 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.10768528974073271,
+  "eval_steps": 1000,
+  "global_step": 9000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0011965032193414745,
+      "grad_norm": 0.4959045114979248,
+      "learning_rate": 0.0002475,
+      "loss": 2.2052,
+      "step": 100
+    },
+    {
+      "epoch": 0.002393006438682949,
+      "grad_norm": 0.5957897298098443,
+      "learning_rate": 0.0004975,
+      "loss": 2.1236,
+      "step": 200
+    },
+    {
+      "epoch": 0.0035895096580244234,
+      "grad_norm": 0.5462649355396835,
+      "learning_rate": 0.000494375,
+      "loss": 2.123,
+      "step": 300
+    },
+    {
+      "epoch": 0.004786012877365898,
+      "grad_norm": 0.479307539013867,
+      "learning_rate": 0.0004886931818181818,
+      "loss": 2.1201,
+      "step": 400
+    },
+    {
+      "epoch": 0.005982516096707372,
+      "grad_norm": 0.4702932661295846,
+      "learning_rate": 0.0004830113636363637,
+      "loss": 2.1218,
+      "step": 500
+    },
+    {
+      "epoch": 0.007179019316048847,
+      "grad_norm": 0.4536330377549505,
+      "learning_rate": 0.0004773295454545455,
+      "loss": 2.1199,
+      "step": 600
+    },
+    {
+      "epoch": 0.008375522535390322,
+      "grad_norm": 0.30397230580627244,
+      "learning_rate": 0.00047164772727272724,
+      "loss": 2.122,
+      "step": 700
+    },
+    {
+      "epoch": 0.009572025754731796,
+      "grad_norm": 0.5497084490461644,
+      "learning_rate": 0.0004659659090909091,
+      "loss": 2.1195,
+      "step": 800
+    },
+    {
+      "epoch": 0.010768528974073271,
+      "grad_norm": 0.31265244740901976,
+      "learning_rate": 0.0004602840909090909,
+      "loss": 2.1214,
+      "step": 900
+    },
+    {
+      "epoch": 0.011965032193414745,
+      "grad_norm": 0.38764126785839226,
+      "learning_rate": 0.0004546022727272727,
+      "loss": 2.1201,
+      "step": 1000
+    },
+    {
+      "epoch": 0.011965032193414745,
+      "eval_accuracy": 0.544493419943561,
+      "eval_loss": 2.119503974914551,
+      "eval_runtime": 14498.4824,
+      "eval_samples_per_second": 38.819,
+      "eval_steps_per_second": 4.852,
+      "step": 1000
+    },
+    {
+      "epoch": 0.01316153541275622,
+      "grad_norm": 0.37861216335098746,
+      "learning_rate": 0.00044892045454545456,
+      "loss": 2.1179,
+      "step": 1100
+    },
+    {
+      "epoch": 0.014358038632097694,
+      "grad_norm": 0.3244816871915935,
+      "learning_rate": 0.00044323863636363636,
+      "loss": 2.123,
+      "step": 1200
+    },
+    {
+      "epoch": 0.015554541851439169,
+      "grad_norm": 0.39134094432184025,
+      "learning_rate": 0.0004375568181818182,
+      "loss": 2.1192,
+      "step": 1300
+    },
+    {
+      "epoch": 0.016751045070780644,
+      "grad_norm": 0.31850239786818274,
+      "learning_rate": 0.000431875,
+      "loss": 2.1195,
+      "step": 1400
+    },
+    {
+      "epoch": 0.017947548290122118,
+      "grad_norm": 0.3353874978413676,
+      "learning_rate": 0.0004261931818181818,
+      "loss": 2.1187,
+      "step": 1500
+    },
+    {
+      "epoch": 0.01914405150946359,
+      "grad_norm": 0.3666941040198008,
+      "learning_rate": 0.0004205113636363637,
+      "loss": 2.1208,
+      "step": 1600
+    },
+    {
+      "epoch": 0.02034055472880507,
+      "grad_norm": 0.2874688345676318,
+      "learning_rate": 0.0004148295454545455,
+      "loss": 2.1182,
+      "step": 1700
+    },
+    {
+      "epoch": 0.021537057948146542,
+      "grad_norm": 0.28829546768632713,
+      "learning_rate": 0.00040914772727272723,
+      "loss": 2.1202,
+      "step": 1800
+    },
+    {
+      "epoch": 0.022733561167488016,
+      "grad_norm": 0.3462480959698083,
+      "learning_rate": 0.0004034659090909091,
+      "loss": 2.1185,
+      "step": 1900
+    },
+    {
+      "epoch": 0.02393006438682949,
+      "grad_norm": 0.32146933893861385,
+      "learning_rate": 0.0003977840909090909,
+      "loss": 2.1222,
+      "step": 2000
+    },
+    {
+      "epoch": 0.02393006438682949,
+      "eval_accuracy": 0.5445787464391751,
+      "eval_loss": 2.1198692321777344,
+      "eval_runtime": 8324.5002,
+      "eval_samples_per_second": 67.609,
+      "eval_steps_per_second": 8.451,
+      "step": 2000
+    },
+    {
+      "epoch": 0.025126567606170967,
+      "grad_norm": 0.3383598719071586,
+      "learning_rate": 0.0003921022727272727,
+      "loss": 2.1181,
+      "step": 2100
+    },
+    {
+      "epoch": 0.02632307082551244,
+      "grad_norm": 0.23870777428521156,
+      "learning_rate": 0.00038642045454545456,
+      "loss": 2.1201,
+      "step": 2200
+    },
+    {
+      "epoch": 0.027519574044853914,
+      "grad_norm": 0.26512869204988454,
+      "learning_rate": 0.00038073863636363636,
+      "loss": 2.1181,
+      "step": 2300
+    },
+    {
+      "epoch": 0.028716077264195387,
+      "grad_norm": 0.38291085278916004,
+      "learning_rate": 0.0003750568181818182,
+      "loss": 2.1214,
+      "step": 2400
+    },
+    {
+      "epoch": 0.029912580483536864,
+      "grad_norm": 0.2922237601337328,
+      "learning_rate": 0.000369375,
+      "loss": 2.1178,
+      "step": 2500
+    },
+    {
+      "epoch": 0.031109083702878338,
+      "grad_norm": 0.28007930758064187,
+      "learning_rate": 0.0003636931818181818,
+      "loss": 2.1174,
+      "step": 2600
+    },
+    {
+      "epoch": 0.03230558692221981,
+      "grad_norm": 0.37650924862735796,
+      "learning_rate": 0.0003580113636363637,
+      "loss": 2.1202,
+      "step": 2700
+    },
+    {
+      "epoch": 0.03350209014156129,
+      "grad_norm": 0.32087380641998864,
+      "learning_rate": 0.0003523295454545455,
+      "loss": 2.1205,
+      "step": 2800
+    },
+    {
+      "epoch": 0.03469859336090276,
+      "grad_norm": 0.3065591070785156,
+      "learning_rate": 0.00034664772727272723,
+      "loss": 2.1219,
+      "step": 2900
+    },
+    {
+      "epoch": 0.035895096580244236,
+      "grad_norm": 0.4973396205464692,
+      "learning_rate": 0.0003409659090909091,
+      "loss": 2.1179,
+      "step": 3000
+    },
+    {
+      "epoch": 0.035895096580244236,
+      "eval_accuracy": 0.5444474125165799,
+      "eval_loss": 2.119871139526367,
+      "eval_runtime": 14003.3046,
+      "eval_samples_per_second": 40.191,
+      "eval_steps_per_second": 5.024,
+      "step": 3000
+    },
+    {
+      "epoch": 0.03709159979958571,
+      "grad_norm": 0.36054926026192535,
+      "learning_rate": 0.0003352840909090909,
+      "loss": 2.1207,
+      "step": 3100
+    },
+    {
+      "epoch": 0.03828810301892718,
+      "grad_norm": 0.2609329016792503,
+      "learning_rate": 0.0003296022727272727,
+      "loss": 2.1187,
+      "step": 3200
+    },
+    {
+      "epoch": 0.03948460623826866,
+      "grad_norm": 0.4377934640382202,
+      "learning_rate": 0.00032392045454545455,
+      "loss": 2.1183,
+      "step": 3300
+    },
+    {
+      "epoch": 0.04068110945761014,
+      "grad_norm": 0.30040663333811507,
+      "learning_rate": 0.00031823863636363636,
+      "loss": 2.1207,
+      "step": 3400
+    },
+    {
+      "epoch": 0.04187761267695161,
+      "grad_norm": 0.4058031327319446,
+      "learning_rate": 0.0003125568181818182,
+      "loss": 2.1175,
+      "step": 3500
+    },
+    {
+      "epoch": 0.043074115896293085,
+      "grad_norm": 0.2221276357895751,
+      "learning_rate": 0.000306875,
+      "loss": 2.1191,
+      "step": 3600
+    },
+    {
+      "epoch": 0.044270619115634555,
+      "grad_norm": 0.2171065800667589,
+      "learning_rate": 0.0003011931818181818,
+      "loss": 2.1222,
+      "step": 3700
+    },
+    {
+      "epoch": 0.04546712233497603,
+      "grad_norm": 0.24009215814108797,
+      "learning_rate": 0.0002955113636363637,
+      "loss": 2.1206,
+      "step": 3800
+    },
+    {
+      "epoch": 0.04666362555431751,
+      "grad_norm": 0.23325080103491874,
+      "learning_rate": 0.0002898295454545455,
+      "loss": 2.1219,
+      "step": 3900
+    },
+    {
+      "epoch": 0.04786012877365898,
+      "grad_norm": 0.38496953527087835,
+      "learning_rate": 0.00028414772727272723,
+      "loss": 2.1215,
+      "step": 4000
+    },
+    {
+      "epoch": 0.04786012877365898,
+      "eval_accuracy": 0.5445841718354034,
+      "eval_loss": 2.1186254024505615,
+      "eval_runtime": 14020.6018,
+      "eval_samples_per_second": 40.142,
+      "eval_steps_per_second": 5.018,
+      "step": 4000
+    },
+    {
+      "epoch": 0.049056631993000456,
+      "grad_norm": 0.2049052054281515,
+      "learning_rate": 0.0002784659090909091,
+      "loss": 2.1184,
+      "step": 4100
+    },
+    {
+      "epoch": 0.05025313521234193,
+      "grad_norm": 0.2458322734082573,
+      "learning_rate": 0.0002727840909090909,
+      "loss": 2.1196,
+      "step": 4200
+    },
+    {
+      "epoch": 0.0514496384316834,
+      "grad_norm": 0.1995524240389471,
+      "learning_rate": 0.0002671022727272727,
+      "loss": 2.1188,
+      "step": 4300
+    },
+    {
+      "epoch": 0.05264614165102488,
+      "grad_norm": 0.25790141869897587,
+      "learning_rate": 0.00026142045454545455,
+      "loss": 2.1184,
+      "step": 4400
+    },
+    {
+      "epoch": 0.05384264487036636,
+      "grad_norm": 0.28730537897175445,
+      "learning_rate": 0.00025573863636363636,
+      "loss": 2.1202,
+      "step": 4500
+    },
+    {
+      "epoch": 0.05503914808970783,
+      "grad_norm": 0.32493426349642957,
+      "learning_rate": 0.0002500568181818182,
+      "loss": 2.118,
+      "step": 4600
+    },
+    {
+      "epoch": 0.056235651309049305,
+      "grad_norm": 0.29535898171610375,
+      "learning_rate": 0.000244375,
+      "loss": 2.1206,
+      "step": 4700
+    },
+    {
+      "epoch": 0.057432154528390775,
+      "grad_norm": 0.22683681078434978,
+      "learning_rate": 0.00023869318181818182,
+      "loss": 2.117,
+      "step": 4800
+    },
+    {
+      "epoch": 0.05862865774773225,
+      "grad_norm": 0.3053349551242291,
+      "learning_rate": 0.00023301136363636362,
+      "loss": 2.1179,
+      "step": 4900
+    },
+    {
+      "epoch": 0.05982516096707373,
+      "grad_norm": 0.1948263427424577,
+      "learning_rate": 0.00022732954545454545,
+      "loss": 2.1176,
+      "step": 5000
+    },
+    {
+      "epoch": 0.05982516096707373,
+      "eval_accuracy": 0.5445371572840895,
+      "eval_loss": 2.119377613067627,
+      "eval_runtime": 12119.2615,
+      "eval_samples_per_second": 46.439,
+      "eval_steps_per_second": 5.805,
+      "step": 5000
+    },
+    {
+      "epoch": 0.0610216641864152,
+      "grad_norm": 0.29200116272624654,
+      "learning_rate": 0.00022164772727272728,
+      "loss": 2.1192,
+      "step": 5100
+    },
+    {
+      "epoch": 0.062218167405756676,
+      "grad_norm": 0.25843746999947237,
+      "learning_rate": 0.00021596590909090912,
+      "loss": 2.1193,
+      "step": 5200
+    },
+    {
+      "epoch": 0.06341467062509815,
+      "grad_norm": 0.23460770073478823,
+      "learning_rate": 0.00021028409090909092,
+      "loss": 2.1196,
+      "step": 5300
+    },
+    {
+      "epoch": 0.06461117384443962,
+      "grad_norm": 0.2143373312023261,
+      "learning_rate": 0.00020460227272727272,
+      "loss": 2.1194,
+      "step": 5400
+    },
+    {
+      "epoch": 0.0658076770637811,
+      "grad_norm": 0.2738612197646279,
+      "learning_rate": 0.00019892045454545455,
+      "loss": 2.1215,
+      "step": 5500
+    },
+    {
+      "epoch": 0.06700418028312258,
+      "grad_norm": 0.22066659063762534,
+      "learning_rate": 0.00019323863636363636,
+      "loss": 2.1185,
+      "step": 5600
+    },
+    {
+      "epoch": 0.06820068350246405,
+      "grad_norm": 0.2649683554628019,
+      "learning_rate": 0.00018755681818181819,
+      "loss": 2.117,
+      "step": 5700
+    },
+    {
+      "epoch": 0.06939718672180552,
+      "grad_norm": 0.36660584937802804,
+      "learning_rate": 0.00018187500000000002,
+      "loss": 2.1179,
+      "step": 5800
+    },
+    {
+      "epoch": 0.070593689941147,
+      "grad_norm": 0.2683451066571477,
+      "learning_rate": 0.00017619318181818182,
+      "loss": 2.1201,
+      "step": 5900
+    },
+    {
+      "epoch": 0.07179019316048847,
+      "grad_norm": 0.1943497599849036,
+      "learning_rate": 0.00017051136363636362,
+      "loss": 2.1189,
+      "step": 6000
+    },
+    {
+      "epoch": 0.07179019316048847,
+      "eval_accuracy": 0.5445187284838173,
+      "eval_loss": 2.1193103790283203,
+      "eval_runtime": 8292.7883,
+      "eval_samples_per_second": 67.868,
+      "eval_steps_per_second": 8.484,
+      "step": 6000
+    },
+    {
+      "epoch": 0.07298669637982995,
+      "grad_norm": 0.2698036045598882,
+      "learning_rate": 0.00016482954545454545,
+      "loss": 2.1196,
+      "step": 6100
+    },
+    {
+      "epoch": 0.07418319959917143,
+      "grad_norm": 0.22553506284296373,
+      "learning_rate": 0.00015914772727272728,
+      "loss": 2.1195,
+      "step": 6200
+    },
+    {
+      "epoch": 0.07537970281851289,
+      "grad_norm": 0.21994704816511382,
+      "learning_rate": 0.00015346590909090911,
+      "loss": 2.119,
+      "step": 6300
+    },
+    {
+      "epoch": 0.07657620603785437,
+      "grad_norm": 0.20519656139775935,
+      "learning_rate": 0.00014778409090909092,
+      "loss": 2.1184,
+      "step": 6400
+    },
+    {
+      "epoch": 0.07777270925719584,
+      "grad_norm": 0.3246545352385696,
+      "learning_rate": 0.00014210227272727272,
+      "loss": 2.1181,
+      "step": 6500
+    },
+    {
+      "epoch": 0.07896921247653732,
+      "grad_norm": 0.25066878379158264,
+      "learning_rate": 0.00013642045454545455,
+      "loss": 2.1196,
+      "step": 6600
+    },
+    {
+      "epoch": 0.0801657156958788,
+      "grad_norm": 0.24038943677884456,
+      "learning_rate": 0.00013073863636363635,
+      "loss": 2.118,
+      "step": 6700
+    },
+    {
+      "epoch": 0.08136221891522027,
+      "grad_norm": 0.21920956719889126,
+      "learning_rate": 0.00012505681818181818,
+      "loss": 2.1188,
+      "step": 6800
+    },
+    {
+      "epoch": 0.08255872213456174,
+      "grad_norm": 0.2013303206012115,
+      "learning_rate": 0.000119375,
+      "loss": 2.1201,
+      "step": 6900
+    },
+    {
+      "epoch": 0.08375522535390321,
+      "grad_norm": 0.17865191614703033,
+      "learning_rate": 0.00011369318181818182,
+      "loss": 2.1184,
+      "step": 7000
+    },
+    {
+      "epoch": 0.08375522535390321,
+      "eval_accuracy": 0.5445597073721878,
+      "eval_loss": 2.119361400604248,
+      "eval_runtime": 8286.5111,
+      "eval_samples_per_second": 67.919,
+      "eval_steps_per_second": 8.49,
+      "step": 7000
+    },
+    {
+      "epoch": 0.08495172857324469,
+      "grad_norm": 0.2242220420058521,
+      "learning_rate": 0.00010801136363636365,
+      "loss": 2.1206,
+      "step": 7100
+    },
+    {
+      "epoch": 0.08614823179258617,
+      "grad_norm": 0.2361418777976826,
+      "learning_rate": 0.00010232954545454545,
+      "loss": 2.1223,
+      "step": 7200
+    },
+    {
+      "epoch": 0.08734473501192765,
+      "grad_norm": 0.2251269866404104,
+      "learning_rate": 9.664772727272728e-05,
+      "loss": 2.1188,
+      "step": 7300
+    },
+    {
+      "epoch": 0.08854123823126911,
+      "grad_norm": 0.2650506829716411,
+      "learning_rate": 9.09659090909091e-05,
+      "loss": 2.1204,
+      "step": 7400
+    },
+    {
+      "epoch": 0.08973774145061059,
+      "grad_norm": 0.2960065490838976,
+      "learning_rate": 8.52840909090909e-05,
+      "loss": 2.1166,
+      "step": 7500
+    },
+    {
+      "epoch": 0.09093424466995206,
+      "grad_norm": 0.2513130989479583,
+      "learning_rate": 7.960227272727273e-05,
+      "loss": 2.1191,
+      "step": 7600
+    },
+    {
+      "epoch": 0.09213074788929354,
+      "grad_norm": 0.23550646079468862,
+      "learning_rate": 7.392045454545454e-05,
+      "loss": 2.1207,
+      "step": 7700
+    },
+    {
+      "epoch": 0.09332725110863502,
+      "grad_norm": 0.20425354689020786,
+      "learning_rate": 6.823863636363637e-05,
+      "loss": 2.1196,
+      "step": 7800
+    },
+    {
+      "epoch": 0.0945237543279765,
+      "grad_norm": 0.27000259463730636,
+      "learning_rate": 6.255681818181818e-05,
+      "loss": 2.1208,
+      "step": 7900
+    },
+    {
+      "epoch": 0.09572025754731796,
+      "grad_norm": 0.23680396479288449,
+      "learning_rate": 5.6875e-05,
+      "loss": 2.119,
+      "step": 8000
+    },
+    {
+      "epoch": 0.09572025754731796,
+      "eval_accuracy": 0.54459064011865,
+      "eval_loss": 2.1189937591552734,
+      "eval_runtime": 8311.026,
+      "eval_samples_per_second": 67.719,
+      "eval_steps_per_second": 8.465,
+      "step": 8000
+    },
+    {
+      "epoch": 0.09691676076665943,
+      "grad_norm": 0.2356498019575465,
+      "learning_rate": 5.119318181818182e-05,
+      "loss": 2.117,
+      "step": 8100
+    },
+    {
+      "epoch": 0.09811326398600091,
+      "grad_norm": 0.2452139474957504,
+      "learning_rate": 4.5511363636363634e-05,
+      "loss": 2.1191,
+      "step": 8200
+    },
+    {
+      "epoch": 0.09930976720534239,
+      "grad_norm": 0.2570095681654715,
+      "learning_rate": 3.982954545454546e-05,
+      "loss": 2.1182,
+      "step": 8300
+    },
+    {
+      "epoch": 0.10050627042468387,
+      "grad_norm": 0.2928491115350356,
+      "learning_rate": 3.4147727272727274e-05,
+      "loss": 2.119,
+      "step": 8400
+    },
+    {
+      "epoch": 0.10170277364402533,
+      "grad_norm": 0.2534123898084123,
+      "learning_rate": 2.846590909090909e-05,
+      "loss": 2.1227,
+      "step": 8500
+    },
+    {
+      "epoch": 0.1028992768633668,
+      "grad_norm": 0.29376855634271687,
+      "learning_rate": 2.278409090909091e-05,
+      "loss": 2.1198,
+      "step": 8600
+    },
+    {
+      "epoch": 0.10409578008270828,
+      "grad_norm": 0.3837678078363553,
+      "learning_rate": 1.7102272727272728e-05,
+      "loss": 2.1206,
+      "step": 8700
+    },
+    {
+      "epoch": 0.10529228330204976,
+      "grad_norm": 0.24342629903402227,
+      "learning_rate": 1.1420454545454547e-05,
+      "loss": 2.12,
+      "step": 8800
+    },
+    {
+      "epoch": 0.10648878652139124,
+      "grad_norm": 0.18367698561178294,
+      "learning_rate": 5.7386363636363634e-06,
+      "loss": 2.1197,
+      "step": 8900
+    },
+    {
+      "epoch": 0.10768528974073271,
+      "grad_norm": 0.27943165556655064,
+      "learning_rate": 5.681818181818182e-08,
+      "loss": 2.1237,
+      "step": 9000
+    },
+    {
+      "epoch": 0.10768528974073271,
+      "eval_accuracy": 0.544624927857023,
+      "eval_loss": 2.12009596824646,
+      "eval_runtime": 8298.0507,
+      "eval_samples_per_second": 67.824,
+      "eval_steps_per_second": 8.478,
+      "step": 9000
+    }
+  ],
+  "logging_steps": 100,
+  "max_steps": 9000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 450,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1158274171797504.0,
+  "train_batch_size": 8,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-9000/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cef530fb4e1d2e9334e368a9556d9eb4c37e254304636d9458c4b52855ca4d03
+size 6459

checkpoint-9000/zero_to_fp32.py

@@ -0,0 +1,592 @@
+#!/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 . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+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)
+
+        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 files:
+        state_dict = torch.load(f, map_location=device)
+        # 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}")
+
+    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_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)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+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):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _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")
+
+
+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 = fp32_flat_groups[0].numel() * 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
+    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"Trainable params: {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 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):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _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 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("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    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, tag=args.tag)

config.json

@@ -0,0 +1,27 @@
+{
+  "_name_or_path": "/src/tanner/chess-roberta-code/model_configs/chess_roberta.json",
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "eos_token_id": 1,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 1030,
+  "model_type": "roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 16,
+  "pad_token_id": 2,
+  "position_embedding_type": "absolute",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.42.4",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 64
+}

eval_results.json

@@ -0,0 +1,10 @@
+{
+    "epoch": 0.10768528974073271,
+    "eval_accuracy": 0.5445640378503339,
+    "eval_loss": 2.1201696395874023,
+    "eval_runtime": 8310.7479,
+    "eval_samples": 562811,
+    "eval_samples_per_second": 67.721,
+    "eval_steps_per_second": 8.465,
+    "perplexity": 8.332550898295501
+}

global_step9000/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c279dc883c0d65945b83146c6a1901c3265ee01b398b33b0e234c652ee2d446
+size 1378093111

global_step9000/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78aa760df334a408b1ff1cad6785692d84af6c1c49652716d7cfd0b815914075
+size 140897

latest

@@ -0,0 +1 @@
+global_step9000

special_tokens_map.json

@@ -0,0 +1,37 @@
+{
+  "cls_token": {
+    "content": "[CLS]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "[SEP]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,162 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "[CLS]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "[SEP]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "[PAD]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "[UNK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 4,
+      "content": "[MASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 5,
+      "content": "[ILLEGAL]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 6,
+      "content": "~",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": {
+    "type": "Sequence",
+    "normalizers": [
+      {
+        "type": "NFD"
+      },
+      {
+        "type": "StripAccents"
+      }
+    ]
+  },
+  "pre_tokenizer": {
+    "type": "Split",
+    "pattern": {
+      "String": ""
+    },
+    "behavior": "Isolated",
+    "invert": false
+  },
+  "post_processor": {
+    "type": "RobertaProcessing",
+    "sep": [
+      "[SEP]",
+      1
+    ],
+    "cls": [
+      "[CLS]",
+      0
+    ],
+    "trim_offsets": false,
+    "add_prefix_space": false
+  },
+  "decoder": {
+    "type": "WordPiece",
+    "prefix": "##",
+    "cleanup": true
+  },
+  "model": {
+    "type": "WordPiece",
+    "unk_token": "[UNK]",
+    "continuing_subword_prefix": "##",
+    "max_input_chars_per_word": 100,
+    "vocab": {
+      "[CLS]": 0,
+      "[SEP]": 1,
+      "[PAD]": 2,
+      "[UNK]": 3,
+      "[MASK]": 4,
+      "[ILLEGAL]": 5,
+      "~": 6,
+      ">": 7,
+      " ": 8,
+      "#": 9,
+      "+": 10,
+      "-": 11,
+      "/": 12,
+      "0": 13,
+      "1": 14,
+      "2": 15,
+      "3": 16,
+      "4": 17,
+      "5": 18,
+      "6": 19,
+      "7": 20,
+      "8": 21,
+      "9": 22,
+      "=": 23,
+      "B": 24,
+      "K": 25,
+      "N": 26,
+      "O": 27,
+      "P": 28,
+      "Q": 29,
+      "R": 30,
+      "a": 31,
+      "b": 32,
+      "c": 33,
+      "d": 34,
+      "e": 35,
+      "f": 36,
+      "g": 37,
+      "h": 38,
+      "k": 39,
+      "n": 40,
+      "p": 41,
+      "q": 42,
+      "r": 43,
+      "w": 44,
+      "x": 45,
+      "_": 46
+    }
+  }
+}

tokenizer_config.json

@@ -0,0 +1,68 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "[ILLEGAL]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "~",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "[UNK]"
+}

train_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 0.10768528974073271,
+    "total_flos": 1158274171797504.0,
+    "train_loss": 2.120608181423611,
+    "train_runtime": 152416.4466,
+    "train_samples": 10697834,
+    "train_samples_per_second": 7.558,
+    "train_steps_per_second": 0.059
+}

trainer_state.json

@@ -0,0 +1,753 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.10768528974073271,
+  "eval_steps": 1000,
+  "global_step": 9000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0011965032193414745,
+      "grad_norm": 0.4959045114979248,
+      "learning_rate": 0.0002475,
+      "loss": 2.2052,
+      "step": 100
+    },
+    {
+      "epoch": 0.002393006438682949,
+      "grad_norm": 0.5957897298098443,
+      "learning_rate": 0.0004975,
+      "loss": 2.1236,
+      "step": 200
+    },
+    {
+      "epoch": 0.0035895096580244234,
+      "grad_norm": 0.5462649355396835,
+      "learning_rate": 0.000494375,
+      "loss": 2.123,
+      "step": 300
+    },
+    {
+      "epoch": 0.004786012877365898,
+      "grad_norm": 0.479307539013867,
+      "learning_rate": 0.0004886931818181818,
+      "loss": 2.1201,
+      "step": 400
+    },
+    {
+      "epoch": 0.005982516096707372,
+      "grad_norm": 0.4702932661295846,
+      "learning_rate": 0.0004830113636363637,
+      "loss": 2.1218,
+      "step": 500
+    },
+    {
+      "epoch": 0.007179019316048847,
+      "grad_norm": 0.4536330377549505,
+      "learning_rate": 0.0004773295454545455,
+      "loss": 2.1199,
+      "step": 600
+    },
+    {
+      "epoch": 0.008375522535390322,
+      "grad_norm": 0.30397230580627244,
+      "learning_rate": 0.00047164772727272724,
+      "loss": 2.122,
+      "step": 700
+    },
+    {
+      "epoch": 0.009572025754731796,
+      "grad_norm": 0.5497084490461644,
+      "learning_rate": 0.0004659659090909091,
+      "loss": 2.1195,
+      "step": 800
+    },
+    {
+      "epoch": 0.010768528974073271,
+      "grad_norm": 0.31265244740901976,
+      "learning_rate": 0.0004602840909090909,
+      "loss": 2.1214,
+      "step": 900
+    },
+    {
+      "epoch": 0.011965032193414745,
+      "grad_norm": 0.38764126785839226,
+      "learning_rate": 0.0004546022727272727,
+      "loss": 2.1201,
+      "step": 1000
+    },
+    {
+      "epoch": 0.011965032193414745,
+      "eval_accuracy": 0.544493419943561,
+      "eval_loss": 2.119503974914551,
+      "eval_runtime": 14498.4824,
+      "eval_samples_per_second": 38.819,
+      "eval_steps_per_second": 4.852,
+      "step": 1000
+    },
+    {
+      "epoch": 0.01316153541275622,
+      "grad_norm": 0.37861216335098746,
+      "learning_rate": 0.00044892045454545456,
+      "loss": 2.1179,
+      "step": 1100
+    },
+    {
+      "epoch": 0.014358038632097694,
+      "grad_norm": 0.3244816871915935,
+      "learning_rate": 0.00044323863636363636,
+      "loss": 2.123,
+      "step": 1200
+    },
+    {
+      "epoch": 0.015554541851439169,
+      "grad_norm": 0.39134094432184025,
+      "learning_rate": 0.0004375568181818182,
+      "loss": 2.1192,
+      "step": 1300
+    },
+    {
+      "epoch": 0.016751045070780644,
+      "grad_norm": 0.31850239786818274,
+      "learning_rate": 0.000431875,
+      "loss": 2.1195,
+      "step": 1400
+    },
+    {
+      "epoch": 0.017947548290122118,
+      "grad_norm": 0.3353874978413676,
+      "learning_rate": 0.0004261931818181818,
+      "loss": 2.1187,
+      "step": 1500
+    },
+    {
+      "epoch": 0.01914405150946359,
+      "grad_norm": 0.3666941040198008,
+      "learning_rate": 0.0004205113636363637,
+      "loss": 2.1208,
+      "step": 1600
+    },
+    {
+      "epoch": 0.02034055472880507,
+      "grad_norm": 0.2874688345676318,
+      "learning_rate": 0.0004148295454545455,
+      "loss": 2.1182,
+      "step": 1700
+    },
+    {
+      "epoch": 0.021537057948146542,
+      "grad_norm": 0.28829546768632713,
+      "learning_rate": 0.00040914772727272723,
+      "loss": 2.1202,
+      "step": 1800
+    },
+    {
+      "epoch": 0.022733561167488016,
+      "grad_norm": 0.3462480959698083,
+      "learning_rate": 0.0004034659090909091,
+      "loss": 2.1185,
+      "step": 1900
+    },
+    {
+      "epoch": 0.02393006438682949,
+      "grad_norm": 0.32146933893861385,
+      "learning_rate": 0.0003977840909090909,
+      "loss": 2.1222,
+      "step": 2000
+    },
+    {
+      "epoch": 0.02393006438682949,
+      "eval_accuracy": 0.5445787464391751,
+      "eval_loss": 2.1198692321777344,
+      "eval_runtime": 8324.5002,
+      "eval_samples_per_second": 67.609,
+      "eval_steps_per_second": 8.451,
+      "step": 2000
+    },
+    {
+      "epoch": 0.025126567606170967,
+      "grad_norm": 0.3383598719071586,
+      "learning_rate": 0.0003921022727272727,
+      "loss": 2.1181,
+      "step": 2100
+    },
+    {
+      "epoch": 0.02632307082551244,
+      "grad_norm": 0.23870777428521156,
+      "learning_rate": 0.00038642045454545456,
+      "loss": 2.1201,
+      "step": 2200
+    },
+    {
+      "epoch": 0.027519574044853914,
+      "grad_norm": 0.26512869204988454,
+      "learning_rate": 0.00038073863636363636,
+      "loss": 2.1181,
+      "step": 2300
+    },
+    {
+      "epoch": 0.028716077264195387,
+      "grad_norm": 0.38291085278916004,
+      "learning_rate": 0.0003750568181818182,
+      "loss": 2.1214,
+      "step": 2400
+    },
+    {
+      "epoch": 0.029912580483536864,
+      "grad_norm": 0.2922237601337328,
+      "learning_rate": 0.000369375,
+      "loss": 2.1178,
+      "step": 2500
+    },
+    {
+      "epoch": 0.031109083702878338,
+      "grad_norm": 0.28007930758064187,
+      "learning_rate": 0.0003636931818181818,
+      "loss": 2.1174,
+      "step": 2600
+    },
+    {
+      "epoch": 0.03230558692221981,
+      "grad_norm": 0.37650924862735796,
+      "learning_rate": 0.0003580113636363637,
+      "loss": 2.1202,
+      "step": 2700
+    },
+    {
+      "epoch": 0.03350209014156129,
+      "grad_norm": 0.32087380641998864,
+      "learning_rate": 0.0003523295454545455,
+      "loss": 2.1205,
+      "step": 2800
+    },
+    {
+      "epoch": 0.03469859336090276,
+      "grad_norm": 0.3065591070785156,
+      "learning_rate": 0.00034664772727272723,
+      "loss": 2.1219,
+      "step": 2900
+    },
+    {
+      "epoch": 0.035895096580244236,
+      "grad_norm": 0.4973396205464692,
+      "learning_rate": 0.0003409659090909091,
+      "loss": 2.1179,
+      "step": 3000
+    },
+    {
+      "epoch": 0.035895096580244236,
+      "eval_accuracy": 0.5444474125165799,
+      "eval_loss": 2.119871139526367,
+      "eval_runtime": 14003.3046,
+      "eval_samples_per_second": 40.191,
+      "eval_steps_per_second": 5.024,
+      "step": 3000
+    },
+    {
+      "epoch": 0.03709159979958571,
+      "grad_norm": 0.36054926026192535,
+      "learning_rate": 0.0003352840909090909,
+      "loss": 2.1207,
+      "step": 3100
+    },
+    {
+      "epoch": 0.03828810301892718,
+      "grad_norm": 0.2609329016792503,
+      "learning_rate": 0.0003296022727272727,
+      "loss": 2.1187,
+      "step": 3200
+    },
+    {
+      "epoch": 0.03948460623826866,
+      "grad_norm": 0.4377934640382202,
+      "learning_rate": 0.00032392045454545455,
+      "loss": 2.1183,
+      "step": 3300
+    },
+    {
+      "epoch": 0.04068110945761014,
+      "grad_norm": 0.30040663333811507,
+      "learning_rate": 0.00031823863636363636,
+      "loss": 2.1207,
+      "step": 3400
+    },
+    {
+      "epoch": 0.04187761267695161,
+      "grad_norm": 0.4058031327319446,
+      "learning_rate": 0.0003125568181818182,
+      "loss": 2.1175,
+      "step": 3500
+    },
+    {
+      "epoch": 0.043074115896293085,
+      "grad_norm": 0.2221276357895751,
+      "learning_rate": 0.000306875,
+      "loss": 2.1191,
+      "step": 3600
+    },
+    {
+      "epoch": 0.044270619115634555,
+      "grad_norm": 0.2171065800667589,
+      "learning_rate": 0.0003011931818181818,
+      "loss": 2.1222,
+      "step": 3700
+    },
+    {
+      "epoch": 0.04546712233497603,
+      "grad_norm": 0.24009215814108797,
+      "learning_rate": 0.0002955113636363637,
+      "loss": 2.1206,
+      "step": 3800
+    },
+    {
+      "epoch": 0.04666362555431751,
+      "grad_norm": 0.23325080103491874,
+      "learning_rate": 0.0002898295454545455,
+      "loss": 2.1219,
+      "step": 3900
+    },
+    {
+      "epoch": 0.04786012877365898,
+      "grad_norm": 0.38496953527087835,
+      "learning_rate": 0.00028414772727272723,
+      "loss": 2.1215,
+      "step": 4000
+    },
+    {
+      "epoch": 0.04786012877365898,
+      "eval_accuracy": 0.5445841718354034,
+      "eval_loss": 2.1186254024505615,
+      "eval_runtime": 14020.6018,
+      "eval_samples_per_second": 40.142,
+      "eval_steps_per_second": 5.018,
+      "step": 4000
+    },
+    {
+      "epoch": 0.049056631993000456,
+      "grad_norm": 0.2049052054281515,
+      "learning_rate": 0.0002784659090909091,
+      "loss": 2.1184,
+      "step": 4100
+    },
+    {
+      "epoch": 0.05025313521234193,
+      "grad_norm": 0.2458322734082573,
+      "learning_rate": 0.0002727840909090909,
+      "loss": 2.1196,
+      "step": 4200
+    },
+    {
+      "epoch": 0.0514496384316834,
+      "grad_norm": 0.1995524240389471,
+      "learning_rate": 0.0002671022727272727,
+      "loss": 2.1188,
+      "step": 4300
+    },
+    {
+      "epoch": 0.05264614165102488,
+      "grad_norm": 0.25790141869897587,
+      "learning_rate": 0.00026142045454545455,
+      "loss": 2.1184,
+      "step": 4400
+    },
+    {
+      "epoch": 0.05384264487036636,
+      "grad_norm": 0.28730537897175445,
+      "learning_rate": 0.00025573863636363636,
+      "loss": 2.1202,
+      "step": 4500
+    },
+    {
+      "epoch": 0.05503914808970783,
+      "grad_norm": 0.32493426349642957,
+      "learning_rate": 0.0002500568181818182,
+      "loss": 2.118,
+      "step": 4600
+    },
+    {
+      "epoch": 0.056235651309049305,
+      "grad_norm": 0.29535898171610375,
+      "learning_rate": 0.000244375,
+      "loss": 2.1206,
+      "step": 4700
+    },
+    {
+      "epoch": 0.057432154528390775,
+      "grad_norm": 0.22683681078434978,
+      "learning_rate": 0.00023869318181818182,
+      "loss": 2.117,
+      "step": 4800
+    },
+    {
+      "epoch": 0.05862865774773225,
+      "grad_norm": 0.3053349551242291,
+      "learning_rate": 0.00023301136363636362,
+      "loss": 2.1179,
+      "step": 4900
+    },
+    {
+      "epoch": 0.05982516096707373,
+      "grad_norm": 0.1948263427424577,
+      "learning_rate": 0.00022732954545454545,
+      "loss": 2.1176,
+      "step": 5000
+    },
+    {
+      "epoch": 0.05982516096707373,
+      "eval_accuracy": 0.5445371572840895,
+      "eval_loss": 2.119377613067627,
+      "eval_runtime": 12119.2615,
+      "eval_samples_per_second": 46.439,
+      "eval_steps_per_second": 5.805,
+      "step": 5000
+    },
+    {
+      "epoch": 0.0610216641864152,
+      "grad_norm": 0.29200116272624654,
+      "learning_rate": 0.00022164772727272728,
+      "loss": 2.1192,
+      "step": 5100
+    },
+    {
+      "epoch": 0.062218167405756676,
+      "grad_norm": 0.25843746999947237,
+      "learning_rate": 0.00021596590909090912,
+      "loss": 2.1193,
+      "step": 5200
+    },
+    {
+      "epoch": 0.06341467062509815,
+      "grad_norm": 0.23460770073478823,
+      "learning_rate": 0.00021028409090909092,
+      "loss": 2.1196,
+      "step": 5300
+    },
+    {
+      "epoch": 0.06461117384443962,
+      "grad_norm": 0.2143373312023261,
+      "learning_rate": 0.00020460227272727272,
+      "loss": 2.1194,
+      "step": 5400
+    },
+    {
+      "epoch": 0.0658076770637811,
+      "grad_norm": 0.2738612197646279,
+      "learning_rate": 0.00019892045454545455,
+      "loss": 2.1215,
+      "step": 5500
+    },
+    {
+      "epoch": 0.06700418028312258,
+      "grad_norm": 0.22066659063762534,
+      "learning_rate": 0.00019323863636363636,
+      "loss": 2.1185,
+      "step": 5600
+    },
+    {
+      "epoch": 0.06820068350246405,
+      "grad_norm": 0.2649683554628019,
+      "learning_rate": 0.00018755681818181819,
+      "loss": 2.117,
+      "step": 5700
+    },
+    {
+      "epoch": 0.06939718672180552,
+      "grad_norm": 0.36660584937802804,
+      "learning_rate": 0.00018187500000000002,
+      "loss": 2.1179,
+      "step": 5800
+    },
+    {
+      "epoch": 0.070593689941147,
+      "grad_norm": 0.2683451066571477,
+      "learning_rate": 0.00017619318181818182,
+      "loss": 2.1201,
+      "step": 5900
+    },
+    {
+      "epoch": 0.07179019316048847,
+      "grad_norm": 0.1943497599849036,
+      "learning_rate": 0.00017051136363636362,
+      "loss": 2.1189,
+      "step": 6000
+    },
+    {
+      "epoch": 0.07179019316048847,
+      "eval_accuracy": 0.5445187284838173,
+      "eval_loss": 2.1193103790283203,
+      "eval_runtime": 8292.7883,
+      "eval_samples_per_second": 67.868,
+      "eval_steps_per_second": 8.484,
+      "step": 6000
+    },
+    {
+      "epoch": 0.07298669637982995,
+      "grad_norm": 0.2698036045598882,
+      "learning_rate": 0.00016482954545454545,
+      "loss": 2.1196,
+      "step": 6100
+    },
+    {
+      "epoch": 0.07418319959917143,
+      "grad_norm": 0.22553506284296373,
+      "learning_rate": 0.00015914772727272728,
+      "loss": 2.1195,
+      "step": 6200
+    },
+    {
+      "epoch": 0.07537970281851289,
+      "grad_norm": 0.21994704816511382,
+      "learning_rate": 0.00015346590909090911,
+      "loss": 2.119,
+      "step": 6300
+    },
+    {
+      "epoch": 0.07657620603785437,
+      "grad_norm": 0.20519656139775935,
+      "learning_rate": 0.00014778409090909092,
+      "loss": 2.1184,
+      "step": 6400
+    },
+    {
+      "epoch": 0.07777270925719584,
+      "grad_norm": 0.3246545352385696,
+      "learning_rate": 0.00014210227272727272,
+      "loss": 2.1181,
+      "step": 6500
+    },
+    {
+      "epoch": 0.07896921247653732,
+      "grad_norm": 0.25066878379158264,
+      "learning_rate": 0.00013642045454545455,
+      "loss": 2.1196,
+      "step": 6600
+    },
+    {
+      "epoch": 0.0801657156958788,
+      "grad_norm": 0.24038943677884456,
+      "learning_rate": 0.00013073863636363635,
+      "loss": 2.118,
+      "step": 6700
+    },
+    {
+      "epoch": 0.08136221891522027,
+      "grad_norm": 0.21920956719889126,
+      "learning_rate": 0.00012505681818181818,
+      "loss": 2.1188,
+      "step": 6800
+    },
+    {
+      "epoch": 0.08255872213456174,
+      "grad_norm": 0.2013303206012115,
+      "learning_rate": 0.000119375,
+      "loss": 2.1201,
+      "step": 6900
+    },
+    {
+      "epoch": 0.08375522535390321,
+      "grad_norm": 0.17865191614703033,
+      "learning_rate": 0.00011369318181818182,
+      "loss": 2.1184,
+      "step": 7000
+    },
+    {
+      "epoch": 0.08375522535390321,
+      "eval_accuracy": 0.5445597073721878,
+      "eval_loss": 2.119361400604248,
+      "eval_runtime": 8286.5111,
+      "eval_samples_per_second": 67.919,
+      "eval_steps_per_second": 8.49,
+      "step": 7000
+    },
+    {
+      "epoch": 0.08495172857324469,
+      "grad_norm": 0.2242220420058521,
+      "learning_rate": 0.00010801136363636365,
+      "loss": 2.1206,
+      "step": 7100
+    },
+    {
+      "epoch": 0.08614823179258617,
+      "grad_norm": 0.2361418777976826,
+      "learning_rate": 0.00010232954545454545,
+      "loss": 2.1223,
+      "step": 7200
+    },
+    {
+      "epoch": 0.08734473501192765,
+      "grad_norm": 0.2251269866404104,
+      "learning_rate": 9.664772727272728e-05,
+      "loss": 2.1188,
+      "step": 7300
+    },
+    {
+      "epoch": 0.08854123823126911,
+      "grad_norm": 0.2650506829716411,
+      "learning_rate": 9.09659090909091e-05,
+      "loss": 2.1204,
+      "step": 7400
+    },
+    {
+      "epoch": 0.08973774145061059,
+      "grad_norm": 0.2960065490838976,
+      "learning_rate": 8.52840909090909e-05,
+      "loss": 2.1166,
+      "step": 7500
+    },
+    {
+      "epoch": 0.09093424466995206,
+      "grad_norm": 0.2513130989479583,
+      "learning_rate": 7.960227272727273e-05,
+      "loss": 2.1191,
+      "step": 7600
+    },
+    {
+      "epoch": 0.09213074788929354,
+      "grad_norm": 0.23550646079468862,
+      "learning_rate": 7.392045454545454e-05,
+      "loss": 2.1207,
+      "step": 7700
+    },
+    {
+      "epoch": 0.09332725110863502,
+      "grad_norm": 0.20425354689020786,
+      "learning_rate": 6.823863636363637e-05,
+      "loss": 2.1196,
+      "step": 7800
+    },
+    {
+      "epoch": 0.0945237543279765,
+      "grad_norm": 0.27000259463730636,
+      "learning_rate": 6.255681818181818e-05,
+      "loss": 2.1208,
+      "step": 7900
+    },
+    {
+      "epoch": 0.09572025754731796,
+      "grad_norm": 0.23680396479288449,
+      "learning_rate": 5.6875e-05,
+      "loss": 2.119,
+      "step": 8000
+    },
+    {
+      "epoch": 0.09572025754731796,
+      "eval_accuracy": 0.54459064011865,
+      "eval_loss": 2.1189937591552734,
+      "eval_runtime": 8311.026,
+      "eval_samples_per_second": 67.719,
+      "eval_steps_per_second": 8.465,
+      "step": 8000
+    },
+    {
+      "epoch": 0.09691676076665943,
+      "grad_norm": 0.2356498019575465,
+      "learning_rate": 5.119318181818182e-05,
+      "loss": 2.117,
+      "step": 8100
+    },
+    {
+      "epoch": 0.09811326398600091,
+      "grad_norm": 0.2452139474957504,
+      "learning_rate": 4.5511363636363634e-05,
+      "loss": 2.1191,
+      "step": 8200
+    },
+    {
+      "epoch": 0.09930976720534239,
+      "grad_norm": 0.2570095681654715,
+      "learning_rate": 3.982954545454546e-05,
+      "loss": 2.1182,
+      "step": 8300
+    },
+    {
+      "epoch": 0.10050627042468387,
+      "grad_norm": 0.2928491115350356,
+      "learning_rate": 3.4147727272727274e-05,
+      "loss": 2.119,
+      "step": 8400
+    },
+    {
+      "epoch": 0.10170277364402533,
+      "grad_norm": 0.2534123898084123,
+      "learning_rate": 2.846590909090909e-05,
+      "loss": 2.1227,
+      "step": 8500
+    },
+    {
+      "epoch": 0.1028992768633668,
+      "grad_norm": 0.29376855634271687,
+      "learning_rate": 2.278409090909091e-05,
+      "loss": 2.1198,
+      "step": 8600
+    },
+    {
+      "epoch": 0.10409578008270828,
+      "grad_norm": 0.3837678078363553,
+      "learning_rate": 1.7102272727272728e-05,
+      "loss": 2.1206,
+      "step": 8700
+    },
+    {
+      "epoch": 0.10529228330204976,
+      "grad_norm": 0.24342629903402227,
+      "learning_rate": 1.1420454545454547e-05,
+      "loss": 2.12,
+      "step": 8800
+    },
+    {
+      "epoch": 0.10648878652139124,
+      "grad_norm": 0.18367698561178294,
+      "learning_rate": 5.7386363636363634e-06,
+      "loss": 2.1197,
+      "step": 8900
+    },
+    {
+      "epoch": 0.10768528974073271,
+      "grad_norm": 0.27943165556655064,
+      "learning_rate": 5.681818181818182e-08,
+      "loss": 2.1237,
+      "step": 9000
+    },
+    {
+      "epoch": 0.10768528974073271,
+      "eval_accuracy": 0.544624927857023,
+      "eval_loss": 2.12009596824646,
+      "eval_runtime": 8298.0507,
+      "eval_samples_per_second": 67.824,
+      "eval_steps_per_second": 8.478,
+      "step": 9000
+    },
+    {
+      "epoch": 0.10768528974073271,
+      "step": 9000,
+      "total_flos": 1158274171797504.0,
+      "train_loss": 2.120608181423611,
+      "train_runtime": 152416.4466,
+      "train_samples_per_second": 7.558,
+      "train_steps_per_second": 0.059
+    }
+  ],
+  "logging_steps": 100,
+  "max_steps": 9000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 450,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1158274171797504.0,
+  "train_batch_size": 8,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cef530fb4e1d2e9334e368a9556d9eb4c37e254304636d9458c4b52855ca4d03
+size 6459

zero_to_fp32.py

@@ -0,0 +1,592 @@
+#!/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 . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+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)
+
+        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 files:
+        state_dict = torch.load(f, map_location=device)
+        # 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}")
+
+    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_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)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+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):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _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")
+
+
+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 = fp32_flat_groups[0].numel() * 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
+    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"Trainable params: {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 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):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _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 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("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    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, tag=args.tag)