Training in progress, step 285

.gitignore

@@ -0,0 +1 @@
+checkpoint-*/

config.json

@@ -0,0 +1,74 @@
+{
+  "_name_or_path": "EleutherAI/gpt-neo-1.3B",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPTNeoForCausalLM"
+  ],
+  "attention_dropout": 0,
+  "attention_layers": [
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local"
+  ],
+  "attention_types": [
+    [
+      [
+        "global",
+        "local"
+      ],
+      12
+    ]
+  ],
+  "bos_token_id": 50256,
+  "embed_dropout": 0,
+  "eos_token_id": 50256,
+  "gradient_checkpointing": false,
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": null,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 2048,
+  "model_type": "gpt_neo",
+  "num_heads": 16,
+  "num_layers": 24,
+  "resid_dropout": 0,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50,
+      "temperature": 0.9
+    }
+  },
+  "tokenizer_class": "GPT2Tokenizer",
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.0",
+  "use_cache": false,
+  "vocab_size": 50257,
+  "window_size": 256
+}

last-checkpoint/config.json

@@ -0,0 +1,74 @@
+{
+  "_name_or_path": "EleutherAI/gpt-neo-1.3B",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPTNeoForCausalLM"
+  ],
+  "attention_dropout": 0,
+  "attention_layers": [
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local"
+  ],
+  "attention_types": [
+    [
+      [
+        "global",
+        "local"
+      ],
+      12
+    ]
+  ],
+  "bos_token_id": 50256,
+  "embed_dropout": 0,
+  "eos_token_id": 50256,
+  "gradient_checkpointing": false,
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": null,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 2048,
+  "model_type": "gpt_neo",
+  "num_heads": 16,
+  "num_layers": 24,
+  "resid_dropout": 0,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50,
+      "temperature": 0.9
+    }
+  },
+  "tokenizer_class": "GPT2Tokenizer",
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.0",
+  "use_cache": false,
+  "vocab_size": 50257,
+  "window_size": 256
+}

last-checkpoint/generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.26.0",
+  "use_cache": false
+}

last-checkpoint/global_step285/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fa42080a65eed08b747fe605449ed3eb9a2087fddc890abc4bd5da53906bf40
+size 5363072554

last-checkpoint/global_step285/zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79389c44d9db01a221667fd743bd63df37efdfcc47afd04c7d05aee643c2ee37
+size 3946735038

last-checkpoint/global_step285/zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd66ce33b12f683058ed2c582348bdcb0a4c684b394f65177dee53ca7b980700
+size 3946736318

last-checkpoint/global_step285/zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8411bc2e8b133d6d45440dab4a4f85632c772a6c066177129ae358928f64a328
+size 3946737086

last-checkpoint/global_step285/zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc129765c2d02d524c92ff41fdfba31d787fa449387d71031923462aea98be97
+size 3946736574

last-checkpoint/latest

@@ -0,0 +1 @@
+global_step285

last-checkpoint/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da2622d1fc7fd4f286bc5969b5c41d5b9fc1c68b054d356b8fb4c1234a332a0e
+size 5363024236

last-checkpoint/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a961c7aa8293531040f379f129f1930093670f11c91d3736bf6d8518af2b1507
+size 14583

last-checkpoint/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbc4475e8615a434844642d2b6db8c82a3eb4d1532c0358142650af35982f7d4
+size 14583

last-checkpoint/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:881eb4ea22a97cf8b6f92b147ff4c6e2ae57dcc7e36f52b570b5be3eeb7887b6
+size 14583

last-checkpoint/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b494daea51340989a686bc325649dfd2d62f65af56b0d25143e33790c8a957a
+size 14583

last-checkpoint/special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>"
+}

last-checkpoint/tokenizer_config.json

@@ -0,0 +1,34 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "model_max_length": 2048,
+  "name_or_path": "EleutherAI/gpt-neo-1.3B",
+  "pad_token": null,
+  "special_tokens_map_file": null,
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

last-checkpoint/trainer_state.json

@@ -0,0 +1,364 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.2490442381212452,
+  "global_step": 285,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0,
+      "loss": 5.4178,
+      "step": 1
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 1.294882868674145e-05,
+      "loss": 0.446,
+      "step": 5
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 1.852558565662928e-05,
+      "loss": 0.3922,
+      "step": 10
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 2.1787779359648994e-05,
+      "loss": 0.3732,
+      "step": 15
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 2.41023426265171e-05,
+      "loss": 0.3591,
+      "step": 20
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 2.58976573734829e-05,
+      "loss": 0.3384,
+      "step": 25
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 2.7364536329536817e-05,
+      "loss": 0.3291,
+      "step": 30
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 2.8604764815275082e-05,
+      "loss": 0.326,
+      "step": 35
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 2.9679099596404923e-05,
+      "loss": 0.3234,
+      "step": 40
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 3.0626730032556536e-05,
+      "loss": 0.3089,
+      "step": 45
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 3.147441434337073e-05,
+      "loss": 0.2995,
+      "step": 50
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 3.224123807782732e-05,
+      "loss": 0.2933,
+      "step": 55
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 3.294129329942464e-05,
+      "loss": 0.2777,
+      "step": 60
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 3.358528167653452e-05,
+      "loss": 0.2757,
+      "step": 65
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 3.4181521785162905e-05,
+      "loss": 0.2674,
+      "step": 70
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 3.473660804639045e-05,
+      "loss": 0.258,
+      "step": 75
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 3.525585656629274e-05,
+      "loss": 0.2524,
+      "step": 80
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 3.574361557584177e-05,
+      "loss": 0.2498,
+      "step": 85
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 3.620348700244436e-05,
+      "loss": 0.2412,
+      "step": 90
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 3.6638488054916214e-05,
+      "loss": 0.231,
+      "step": 95
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 3.705117131325856e-05,
+      "loss": 0.227,
+      "step": 100
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 3.7443715488182624e-05,
+      "loss": 0.219,
+      "step": 105
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 3.781799504771514e-05,
+      "loss": 0.2189,
+      "step": 110
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 3.81756343539018e-05,
+      "loss": 0.2148,
+      "step": 115
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 3.851805026931246e-05,
+      "loss": 0.2054,
+      "step": 120
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 3.8846486060224364e-05,
+      "loss": 0.2012,
+      "step": 125
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 3.916203864642234e-05,
+      "loss": 0.1942,
+      "step": 130
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 3.946568070546408e-05,
+      "loss": 0.1882,
+      "step": 135
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 3.975827875505073e-05,
+      "loss": 0.1829,
+      "step": 140
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 4.004060806090172e-05,
+      "loss": 0.1845,
+      "step": 145
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 4.031336501627827e-05,
+      "loss": 0.1678,
+      "step": 150
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 4.0577177490884e-05,
+      "loss": 0.1798,
+      "step": 155
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 4.0832613536180565e-05,
+      "loss": 0.1735,
+      "step": 160
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 4.1080188750734856e-05,
+      "loss": 0.176,
+      "step": 165
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 4.1320372545729594e-05,
+      "loss": 0.1662,
+      "step": 170
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 4.155359350201654e-05,
+      "loss": 0.1565,
+      "step": 175
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 4.178024397233218e-05,
+      "loss": 0.1574,
+      "step": 180
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 4.200068405281827e-05,
+      "loss": 0.1549,
+      "step": 185
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 4.221524502480404e-05,
+      "loss": 0.1573,
+      "step": 190
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 4.242423234944206e-05,
+      "loss": 0.1477,
+      "step": 195
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 4.262792828314637e-05,
+      "loss": 0.1567,
+      "step": 200
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 4.282659417003183e-05,
+      "loss": 0.1423,
+      "step": 205
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 4.302047245807045e-05,
+      "loss": 0.1405,
+      "step": 210
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 4.320978847798302e-05,
+      "loss": 0.1406,
+      "step": 215
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 4.3394752017602966e-05,
+      "loss": 0.1381,
+      "step": 220
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 4.357555871929799e-05,
+      "loss": 0.131,
+      "step": 225
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 4.375239132378962e-05,
+      "loss": 0.1507,
+      "step": 230
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 4.392542078019592e-05,
+      "loss": 0.1199,
+      "step": 235
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 4.4094807239200284e-05,
+      "loss": 0.1257,
+      "step": 240
+    },
+    {
+      "epoch": 1.07,
+      "learning_rate": 4.426070094380871e-05,
+      "loss": 0.1185,
+      "step": 245
+    },
+    {
+      "epoch": 1.1,
+      "learning_rate": 4.442324303011218e-05,
+      "loss": 0.115,
+      "step": 250
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 4.458256624874931e-05,
+      "loss": 0.1188,
+      "step": 255
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 4.4738795616310163e-05,
+      "loss": 0.108,
+      "step": 260
+    },
+    {
+      "epoch": 1.16,
+      "learning_rate": 4.48920490046898e-05,
+      "loss": 0.1125,
+      "step": 265
+    },
+    {
+      "epoch": 1.18,
+      "learning_rate": 4.50424376753519e-05,
+      "loss": 0.1094,
+      "step": 270
+    },
+    {
+      "epoch": 1.21,
+      "learning_rate": 4.5190066764568774e-05,
+      "loss": 0.1033,
+      "step": 275
+    },
+    {
+      "epoch": 1.23,
+      "learning_rate": 4.533503572493855e-05,
+      "loss": 0.1037,
+      "step": 280
+    },
+    {
+      "epoch": 1.25,
+      "learning_rate": 4.547743872782376e-05,
+      "loss": 0.1038,
+      "step": 285
+    }
+  ],
+  "max_steps": 1140,
+  "num_train_epochs": 5,
+  "total_flos": 2.1734911911007355e+18,
+  "trial_name": null,
+  "trial_params": null
+}

last-checkpoint/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35e1a0fcbf25720ed542bd10424c03d13a02ac8f5eff85f51e6389fe1f87d97f
+size 4603

last-checkpoint/zero_to_fp32.py

@@ -0,0 +1,482 @@
+#!/usr/bin/env python
+
+# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION,
+                                            OPTIMIZER_STATE_DICT,
+                                            SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS,
+                                            ZERO_STAGE,
+                                            PARTITION_COUNT,
+                                            PARAM_SHAPES,
+                                            BUFFER_NAMES)
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage == 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_optim_files(checkpoint_dir):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
+                                                "*_optim_states.pt")),
+                         key=natural_keys)
+
+    if len(optim_files) == 0:
+        raise FileNotFoundError(
+            f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
+
+    return optim_files
+
+
+def parse_model_state(file):
+    state_dict = torch.load(file, map_location=device)
+
+    if BUFFER_NAMES not in state_dict:
+        raise ValueError(f"{file} is not a model state checkpoint")
+    buffer_names = state_dict[BUFFER_NAMES]
+    if debug:
+        print("Found buffers:", buffer_names)
+
+    # recover just the buffers while restoring them to fp32 if they were saved in fp16
+    buffers = {
+        k: v.float()
+        for k,
+        v in state_dict["module"].items() if k in buffer_names
+    }
+    param_shapes = state_dict[PARAM_SHAPES]
+
+    ds_version = state_dict.get(DS_VERSION, None)
+
+    return buffers, param_shapes, ds_version
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage == 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage == 2:
+        fp32_flat_groups = [
+            state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
+            for i in range(len(state_dicts))
+        ]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
+                      0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(
+        f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
+    buffers, param_shapes, ds_version = parse_model_state(model_file)
+    print(f'Parsing checkpoint created by deepspeed=={ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(
+                    f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum([
+        full_single_fp32_vector.numel()
+        for full_single_fp32_vector in merged_single_partition_of_fp32_groups
+    ])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum(
+            [sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(
+                    f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
+                )
+            state_dict[name] = full_single_fp32_vector.narrow(
+                0,
+                offset,
+                unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(
+                f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(
+        f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0,
+                                             offset,
+                                             partitioned_numel)
+                  for i in range(world_size)),
+            0).narrow(0,
+                      0,
+                      unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(
+            f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(
+        f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "checkpoint_dir",
+        type=str,
+        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help=
+        "path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
+    )
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da2622d1fc7fd4f286bc5969b5c41d5b9fc1c68b054d356b8fb4c1234a332a0e
+size 5363024236

special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>"
+}

tokenizer_config.json

@@ -0,0 +1,34 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "model_max_length": 2048,
+  "name_or_path": "EleutherAI/gpt-neo-1.3B",
+  "pad_token": null,
+  "special_tokens_map_file": null,
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35e1a0fcbf25720ed542bd10424c03d13a02ac8f5eff85f51e6389fe1f87d97f
+size 4603