Duplicate from georgeyw/gpt-2-small-testing

Co-authored-by: George Wang <georgeyw@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

checkpoint-10/config.json

@@ -0,0 +1,31 @@
+{
+  "_name_or_path": "georgeyw/gpt-2-small-init-seed-5",
+  "architectures": [
+    "GPTNeoXForCausalLM"
+  ],
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "classifier_dropout": 0.1,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 1024,
+  "model_type": "gpt_neox",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "rope_scaling": null,
+  "rotary_emb_base": 10000,
+  "rotary_pct": 0.25,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.2",
+  "use_cache": true,
+  "use_parallel_residual": true,
+  "vocab_size": 50304
+}

checkpoint-10/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "transformers_version": "4.38.2"
+}

checkpoint-10/global_step10/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:ed0df787a125d81094d867c743d5ffebb34766f877f42e3b5da9ad0e51fbf81e
+size 973946896

checkpoint-10/global_step10/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f6768e7d81521ef703a230f41460e04972f79ded2de3b754bcd0fb7a96f4ff0
+size 973946832

checkpoint-10/global_step10/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7589ffd4359a842efe83dc7cc5ce4ff6a161f1fce829498ade432719a3c40d91
+size 324689964

checkpoint-10/latest

@@ -0,0 +1 @@
+global_step10

checkpoint-10/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6ba4e69094dffe72d300db52887b581d271c61d1a95702fdea83bf7d6d5697f
+size 324662984

checkpoint-10/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0bb7d2ecdd48fd7d0be1e75b0e3f29004064381052fa203ed926e88b90ef530
+size 14512

checkpoint-10/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:177d534a379bd6b276474c2cb140e318dc65db4457b6c1b6f25a1a9dd563af82
+size 14512

checkpoint-10/trainer_state.json

@@ -0,0 +1,91 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.009995002498750625,
+  "eval_steps": 500,
+  "global_step": 10,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "grad_norm": 3.3340553137590683,
+      "learning_rate": 0.0,
+      "loss": 11.0,
+      "step": 1
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.398799707355898,
+      "learning_rate": 5.9999999999999995e-05,
+      "loss": 10.125,
+      "step": 2
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.3943029297945575,
+      "learning_rate": 0.00011999999999999999,
+      "loss": 10.1172,
+      "step": 3
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.9959117709404242,
+      "learning_rate": 0.00017999999999999998,
+      "loss": 9.875,
+      "step": 4
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.8270696218303057,
+      "learning_rate": 0.00023999999999999998,
+      "loss": 9.6641,
+      "step": 5
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.7854351602113614,
+      "learning_rate": 0.0003,
+      "loss": 9.4844,
+      "step": 6
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.7194174424274788,
+      "learning_rate": 0.00035999999999999997,
+      "loss": 9.3281,
+      "step": 7
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.463772638994466,
+      "learning_rate": 0.00041999999999999996,
+      "loss": 9.2109,
+      "step": 8
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.439323678271545,
+      "learning_rate": 0.00047999999999999996,
+      "loss": 8.9453,
+      "step": 9
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.2936126396494727,
+      "learning_rate": 0.00054,
+      "loss": 8.7109,
+      "step": 10
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 1000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 10,
+  "total_flos": 0.0,
+  "train_batch_size": 32,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-10/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95cc4290cc90782d57f7376defd26743b3a36943fc93e80e2734385bc57e8b78
+size 6520

checkpoint-10/zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/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, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    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, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_file,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoint-100/config.json

@@ -0,0 +1,31 @@
+{
+  "_name_or_path": "georgeyw/gpt-2-small-init-seed-5",
+  "architectures": [
+    "GPTNeoXForCausalLM"
+  ],
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "classifier_dropout": 0.1,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 1024,
+  "model_type": "gpt_neox",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "rope_scaling": null,
+  "rotary_emb_base": 10000,
+  "rotary_pct": 0.25,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.2",
+  "use_cache": true,
+  "use_parallel_residual": true,
+  "vocab_size": 50304
+}

checkpoint-100/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "transformers_version": "4.38.2"
+}

checkpoint-100/global_step100/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:f13244aa35b694472ea5bd7fde3bcee569b6a0b7251253f818b7e5ae6a048797
+size 973946896

checkpoint-100/global_step100/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1039a6c333c1647ef2ddfc010c1b966f38d1001aff73ac8b3cdede7858acdc7
+size 973946832

checkpoint-100/global_step100/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26fe4253d323b90c0fa9e0d5da8cb64819f86baf288d72f171c73184ca877562
+size 324689964

checkpoint-100/latest

@@ -0,0 +1 @@
+global_step100

checkpoint-100/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c9a5c8eae560953cb2afb2b3bf5d1330d089af788edaa1bf1d3f77b218837bc
+size 324662984

checkpoint-100/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0bb7d2ecdd48fd7d0be1e75b0e3f29004064381052fa203ed926e88b90ef530
+size 14512

checkpoint-100/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:177d534a379bd6b276474c2cb140e318dc65db4457b6c1b6f25a1a9dd563af82
+size 14512

checkpoint-100/trainer_state.json

@@ -0,0 +1,721 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.09995002498750624,
+  "eval_steps": 500,
+  "global_step": 100,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "grad_norm": 3.3340563149001086,
+      "learning_rate": 0.0,
+      "loss": 11.0,
+      "step": 1
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.398812329952019,
+      "learning_rate": 5.9999999999999995e-05,
+      "loss": 10.125,
+      "step": 2
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.394322446895115,
+      "learning_rate": 0.00011999999999999999,
+      "loss": 10.1172,
+      "step": 3
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.9958816684399585,
+      "learning_rate": 0.00017999999999999998,
+      "loss": 9.875,
+      "step": 4
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.8270465897882062,
+      "learning_rate": 0.00023999999999999998,
+      "loss": 9.6641,
+      "step": 5
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.7854046471397795,
+      "learning_rate": 0.0003,
+      "loss": 9.4844,
+      "step": 6
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.719416749115252,
+      "learning_rate": 0.00035999999999999997,
+      "loss": 9.3281,
+      "step": 7
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.4637825746112274,
+      "learning_rate": 0.00041999999999999996,
+      "loss": 9.2109,
+      "step": 8
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.4393631015406718,
+      "learning_rate": 0.00047999999999999996,
+      "loss": 8.9453,
+      "step": 9
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.2936734586915988,
+      "learning_rate": 0.00054,
+      "loss": 8.7109,
+      "step": 10
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.0756922378227356,
+      "learning_rate": 0.0005999986405514987,
+      "loss": 8.4609,
+      "step": 11
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.9277829127413892,
+      "learning_rate": 0.0005999945622196846,
+      "loss": 8.2344,
+      "step": 12
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.8084581786682467,
+      "learning_rate": 0.0005999877650456265,
+      "loss": 8.125,
+      "step": 13
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.7635084596900947,
+      "learning_rate": 0.000599978249097772,
+      "loss": 7.9766,
+      "step": 14
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.9186699644247788,
+      "learning_rate": 0.0005999660144719463,
+      "loss": 7.8555,
+      "step": 15
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6609504256551479,
+      "learning_rate": 0.0005999510612913519,
+      "loss": 7.7734,
+      "step": 16
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.7086232844782971,
+      "learning_rate": 0.0005999333897065673,
+      "loss": 7.7148,
+      "step": 17
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 16.38048851691348,
+      "learning_rate": 0.0005999129998955453,
+      "loss": 8.4844,
+      "step": 18
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 1.3057527590449889,
+      "learning_rate": 0.0005998898920636111,
+      "loss": 7.7539,
+      "step": 19
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6966048242948986,
+      "learning_rate": 0.00059986406644346,
+      "loss": 7.75,
+      "step": 20
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6348089115348993,
+      "learning_rate": 0.0005998355232951559,
+      "loss": 7.7031,
+      "step": 21
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.7829163518610293,
+      "learning_rate": 0.0005998042629061279,
+      "loss": 7.6992,
+      "step": 22
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.5900591778980369,
+      "learning_rate": 0.0005997702855911678,
+      "loss": 7.6016,
+      "step": 23
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.4655170213064256,
+      "learning_rate": 0.0005997335916924268,
+      "loss": 7.5977,
+      "step": 24
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6287348258915756,
+      "learning_rate": 0.0005996941815794121,
+      "loss": 7.5586,
+      "step": 25
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.6137321903884564,
+      "learning_rate": 0.0005996520556489831,
+      "loss": 7.5898,
+      "step": 26
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.44962562710631065,
+      "learning_rate": 0.0005996072143253473,
+      "loss": 7.4336,
+      "step": 27
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.46130046454703316,
+      "learning_rate": 0.0005995596580600566,
+      "loss": 7.4023,
+      "step": 28
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.4686712675731326,
+      "learning_rate": 0.0005995093873320018,
+      "loss": 7.3789,
+      "step": 29
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.4672147564288997,
+      "learning_rate": 0.0005994564026474087,
+      "loss": 7.3711,
+      "step": 30
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.40408354581233474,
+      "learning_rate": 0.0005994007045398324,
+      "loss": 7.3672,
+      "step": 31
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.46032146732584733,
+      "learning_rate": 0.0005993422935701524,
+      "loss": 7.3477,
+      "step": 32
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.4765534634593268,
+      "learning_rate": 0.0005992811703265664,
+      "loss": 7.3555,
+      "step": 33
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.46208489386235113,
+      "learning_rate": 0.0005992173354245849,
+      "loss": 7.3047,
+      "step": 34
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.2956144524964961,
+      "learning_rate": 0.0005991507895070244,
+      "loss": 7.3125,
+      "step": 35
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.4834645389868856,
+      "learning_rate": 0.0005990815332440017,
+      "loss": 7.207,
+      "step": 36
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.4411831350968505,
+      "learning_rate": 0.0005990095673329266,
+      "loss": 7.1758,
+      "step": 37
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.24809297748968667,
+      "learning_rate": 0.0005989348924984951,
+      "loss": 7.2188,
+      "step": 38
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.39402988416840584,
+      "learning_rate": 0.0005988575094926817,
+      "loss": 7.1953,
+      "step": 39
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.3868345222189167,
+      "learning_rate": 0.0005987774190947328,
+      "loss": 7.1641,
+      "step": 40
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.3777261230135448,
+      "learning_rate": 0.0005986946221111575,
+      "loss": 7.1328,
+      "step": 41
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.4687511444077827,
+      "learning_rate": 0.0005986091193757206,
+      "loss": 7.0898,
+      "step": 42
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.34935796211612463,
+      "learning_rate": 0.0005985209117494337,
+      "loss": 7.1367,
+      "step": 43
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.38764476686849886,
+      "learning_rate": 0.0005984300001205466,
+      "loss": 7.125,
+      "step": 44
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.3956487898882936,
+      "learning_rate": 0.0005983363854045386,
+      "loss": 7.1094,
+      "step": 45
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.31140257544677513,
+      "learning_rate": 0.0005982400685441084,
+      "loss": 7.0898,
+      "step": 46
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.3664476570531787,
+      "learning_rate": 0.0005981410505091662,
+      "loss": 7.0664,
+      "step": 47
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.31891741142945207,
+      "learning_rate": 0.0005980393322968223,
+      "loss": 7.0273,
+      "step": 48
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.4533529037337155,
+      "learning_rate": 0.0005979349149313778,
+      "loss": 7.0586,
+      "step": 49
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.30532331638835586,
+      "learning_rate": 0.0005978277994643147,
+      "loss": 7.0195,
+      "step": 50
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.6501991746260075,
+      "learning_rate": 0.0005977179869742844,
+      "loss": 6.9648,
+      "step": 51
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.43904455901717926,
+      "learning_rate": 0.0005976054785670975,
+      "loss": 6.9805,
+      "step": 52
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.4826001598483571,
+      "learning_rate": 0.0005974902753757124,
+      "loss": 6.9297,
+      "step": 53
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.2924998027034648,
+      "learning_rate": 0.000597372378560224,
+      "loss": 6.8984,
+      "step": 54
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.4439033666380787,
+      "learning_rate": 0.0005972517893078517,
+      "loss": 6.8945,
+      "step": 55
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.6135914255073411,
+      "learning_rate": 0.0005971285088329284,
+      "loss": 6.9727,
+      "step": 56
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.5575686565598483,
+      "learning_rate": 0.0005970025383768866,
+      "loss": 6.9219,
+      "step": 57
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.4820951675994578,
+      "learning_rate": 0.0005968738792082478,
+      "loss": 6.8516,
+      "step": 58
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.40164190019465584,
+      "learning_rate": 0.0005967425326226082,
+      "loss": 6.7734,
+      "step": 59
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.46129863945181293,
+      "learning_rate": 0.0005966084999426265,
+      "loss": 6.8125,
+      "step": 60
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.33322355827118677,
+      "learning_rate": 0.0005964717825180101,
+      "loss": 6.7891,
+      "step": 61
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.3847525153855558,
+      "learning_rate": 0.0005963323817255024,
+      "loss": 6.8242,
+      "step": 62
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.3384433591375982,
+      "learning_rate": 0.0005961902989688674,
+      "loss": 6.707,
+      "step": 63
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.3937003195165685,
+      "learning_rate": 0.000596045535678877,
+      "loss": 6.8203,
+      "step": 64
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.35423488053528107,
+      "learning_rate": 0.0005958980933132962,
+      "loss": 6.7383,
+      "step": 65
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.36005939745315396,
+      "learning_rate": 0.0005957479733568675,
+      "loss": 6.7109,
+      "step": 66
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3499278317706933,
+      "learning_rate": 0.0005955951773212976,
+      "loss": 6.7266,
+      "step": 67
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3708385192137018,
+      "learning_rate": 0.0005954397067452407,
+      "loss": 6.7617,
+      "step": 68
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3775657656205869,
+      "learning_rate": 0.0005952815631942839,
+      "loss": 6.7148,
+      "step": 69
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3040083750375816,
+      "learning_rate": 0.0005951207482609307,
+      "loss": 6.5938,
+      "step": 70
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3443020808841468,
+      "learning_rate": 0.0005949572635645861,
+      "loss": 6.6523,
+      "step": 71
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3520066316939,
+      "learning_rate": 0.0005947911107515389,
+      "loss": 6.6211,
+      "step": 72
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3739040572679613,
+      "learning_rate": 0.0005946222914949462,
+      "loss": 6.5547,
+      "step": 73
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.34890731989025553,
+      "learning_rate": 0.000594450807494816,
+      "loss": 6.5859,
+      "step": 74
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.40910932350136514,
+      "learning_rate": 0.0005942766604779903,
+      "loss": 6.5547,
+      "step": 75
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5698342865852906,
+      "learning_rate": 0.0005940998521981274,
+      "loss": 6.457,
+      "step": 76
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5179452709555474,
+      "learning_rate": 0.0005939203844356852,
+      "loss": 6.5547,
+      "step": 77
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5222512938673792,
+      "learning_rate": 0.0005937382589979016,
+      "loss": 6.5039,
+      "step": 78
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5682332793686307,
+      "learning_rate": 0.0005935534777187781,
+      "loss": 6.5547,
+      "step": 79
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3869287710460676,
+      "learning_rate": 0.0005933660424590598,
+      "loss": 6.5156,
+      "step": 80
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3078211032807607,
+      "learning_rate": 0.000593175955106218,
+      "loss": 6.4258,
+      "step": 81
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3611357511872241,
+      "learning_rate": 0.00059298321757443,
+      "loss": 6.4727,
+      "step": 82
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.29633467844266953,
+      "learning_rate": 0.0005927878318045608,
+      "loss": 6.3281,
+      "step": 83
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3257574200776832,
+      "learning_rate": 0.0005925897997641426,
+      "loss": 6.3203,
+      "step": 84
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.2824054533852328,
+      "learning_rate": 0.0005923891234473562,
+      "loss": 6.4062,
+      "step": 85
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.3056199770204573,
+      "learning_rate": 0.0005921858048750097,
+      "loss": 6.3984,
+      "step": 86
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.2966438824341908,
+      "learning_rate": 0.000591979846094519,
+      "loss": 6.3555,
+      "step": 87
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.32782438676663733,
+      "learning_rate": 0.0005917712491798866,
+      "loss": 6.4023,
+      "step": 88
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.3538316399620157,
+      "learning_rate": 0.0005915600162316811,
+      "loss": 6.2812,
+      "step": 89
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.375858298192913,
+      "learning_rate": 0.0005913461493770162,
+      "loss": 6.3086,
+      "step": 90
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.5189251339815161,
+      "learning_rate": 0.0005911296507695284,
+      "loss": 6.2812,
+      "step": 91
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.6304909542669104,
+      "learning_rate": 0.0005909105225893564,
+      "loss": 6.2969,
+      "step": 92
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.4655662819622591,
+      "learning_rate": 0.0005906887670431187,
+      "loss": 6.1953,
+      "step": 93
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.39035390983920965,
+      "learning_rate": 0.000590464386363891,
+      "loss": 6.2617,
+      "step": 94
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.4918417851770978,
+      "learning_rate": 0.0005902373828111843,
+      "loss": 6.2148,
+      "step": 95
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.35670770889552555,
+      "learning_rate": 0.0005900077586709219,
+      "loss": 6.2461,
+      "step": 96
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.4177985869939347,
+      "learning_rate": 0.0005897755162554163,
+      "loss": 6.1797,
+      "step": 97
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.3742471130708234,
+      "learning_rate": 0.000589540657903346,
+      "loss": 6.1406,
+      "step": 98
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.28627666723978284,
+      "learning_rate": 0.0005893031859797322,
+      "loss": 6.2031,
+      "step": 99
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.32238563846046103,
+      "learning_rate": 0.0005890631028759143,
+      "loss": 6.0625,
+      "step": 100
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 1000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 100,
+  "total_flos": 0.0,
+  "train_batch_size": 32,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-100/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36ce7ab48fa86ef42491eaad3583773d2b60353997a5e7b6fb4ffc1414828749
+size 6520

checkpoint-100/zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/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, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    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, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_file,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoint-20/config.json

@@ -0,0 +1,31 @@
+{
+  "_name_or_path": "georgeyw/gpt-2-small-init-seed-5",
+  "architectures": [
+    "GPTNeoXForCausalLM"
+  ],
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "classifier_dropout": 0.1,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 1024,
+  "model_type": "gpt_neox",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "rope_scaling": null,
+  "rotary_emb_base": 10000,
+  "rotary_pct": 0.25,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.2",
+  "use_cache": true,
+  "use_parallel_residual": true,
+  "vocab_size": 50304
+}

checkpoint-20/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "transformers_version": "4.38.2"
+}

checkpoint-20/global_step20/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d9619b0d3ceeed4c45a4c6fd5be8946b1be76a5801c466123be2ca841e5e337
+size 973946896

checkpoint-20/global_step20/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c51b9f289ed8d24931448d39d8478de2e97527186251a29c92850b1de7562ea
+size 973946832

checkpoint-20/global_step20/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06da85a865332b0c223e8dc427f16b0c81330388f43d5d7b854b9b349b9e2f89
+size 324689964

checkpoint-20/latest

@@ -0,0 +1 @@
+global_step20

checkpoint-20/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99984de686ddb1e07f6ea7778c91ab6e3547271af7b06526b184e6578f6bf40d
+size 324662984

checkpoint-20/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0bb7d2ecdd48fd7d0be1e75b0e3f29004064381052fa203ed926e88b90ef530
+size 14512

checkpoint-20/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:177d534a379bd6b276474c2cb140e318dc65db4457b6c1b6f25a1a9dd563af82
+size 14512

checkpoint-20/trainer_state.json

@@ -0,0 +1,161 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.01999000499750125,
+  "eval_steps": 500,
+  "global_step": 20,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "grad_norm": 3.3340553137590683,
+      "learning_rate": 0.0,
+      "loss": 11.0,
+      "step": 1
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.398799707355898,
+      "learning_rate": 5.9999999999999995e-05,
+      "loss": 10.125,
+      "step": 2
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.3943029297945575,
+      "learning_rate": 0.00011999999999999999,
+      "loss": 10.1172,
+      "step": 3
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.9959117709404242,
+      "learning_rate": 0.00017999999999999998,
+      "loss": 9.875,
+      "step": 4
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.8270696218303057,
+      "learning_rate": 0.00023999999999999998,
+      "loss": 9.6641,
+      "step": 5
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.7854351602113614,
+      "learning_rate": 0.0003,
+      "loss": 9.4844,
+      "step": 6
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.7194174424274788,
+      "learning_rate": 0.00035999999999999997,
+      "loss": 9.3281,
+      "step": 7
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.463772638994466,
+      "learning_rate": 0.00041999999999999996,
+      "loss": 9.2109,
+      "step": 8
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.439323678271545,
+      "learning_rate": 0.00047999999999999996,
+      "loss": 8.9453,
+      "step": 9
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.2936126396494727,
+      "learning_rate": 0.00054,
+      "loss": 8.7109,
+      "step": 10
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.0757761814549318,
+      "learning_rate": 0.0005999986405514987,
+      "loss": 8.4609,
+      "step": 11
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.9278570154341632,
+      "learning_rate": 0.0005999945622196846,
+      "loss": 8.2344,
+      "step": 12
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.8086775215724974,
+      "learning_rate": 0.0005999877650456265,
+      "loss": 8.125,
+      "step": 13
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.7630413213242441,
+      "learning_rate": 0.000599978249097772,
+      "loss": 7.9766,
+      "step": 14
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.9172017565891333,
+      "learning_rate": 0.0005999660144719463,
+      "loss": 7.8555,
+      "step": 15
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6610052304024877,
+      "learning_rate": 0.0005999510612913519,
+      "loss": 7.7734,
+      "step": 16
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.7091485456070775,
+      "learning_rate": 0.0005999333897065673,
+      "loss": 7.7148,
+      "step": 17
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 16.771353248766836,
+      "learning_rate": 0.0005999129998955453,
+      "loss": 8.5078,
+      "step": 18
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 1.3123969082989795,
+      "learning_rate": 0.0005998898920636111,
+      "loss": 7.7539,
+      "step": 19
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6992078172905232,
+      "learning_rate": 0.00059986406644346,
+      "loss": 7.75,
+      "step": 20
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 1000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 10,
+  "total_flos": 0.0,
+  "train_batch_size": 32,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-20/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95cc4290cc90782d57f7376defd26743b3a36943fc93e80e2734385bc57e8b78
+size 6520

checkpoint-20/zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/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, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    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, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_file,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoint-200/config.json

@@ -0,0 +1,31 @@
+{
+  "_name_or_path": "georgeyw/gpt-2-small-init-seed-5",
+  "architectures": [
+    "GPTNeoXForCausalLM"
+  ],
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "classifier_dropout": 0.1,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 1024,
+  "model_type": "gpt_neox",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "rope_scaling": null,
+  "rotary_emb_base": 10000,
+  "rotary_pct": 0.25,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.2",
+  "use_cache": true,
+  "use_parallel_residual": true,
+  "vocab_size": 50304
+}

checkpoint-200/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "transformers_version": "4.38.2"
+}

checkpoint-200/global_step200/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:cc073673977d2260d446a5288c182308ae00d243deb8aa1287eed2bdc0ca55eb
+size 973946896

checkpoint-200/global_step200/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dd88f92e1b9309a12e7505328b079538457daedef3bcf8d59f5cfcef86aca6d
+size 973946832

checkpoint-200/global_step200/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea030c07bc1914d14db0809ed9a4fc3f9c76b4d11dbdabd30acf57c7c93e685c
+size 324689964

checkpoint-200/latest

@@ -0,0 +1 @@
+global_step200

checkpoint-200/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd81f3f4d4881c9ad52c0835b8eedd34cd74ec334ef06ba64c40c85ac825a476
+size 324662984

checkpoint-200/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0bb7d2ecdd48fd7d0be1e75b0e3f29004064381052fa203ed926e88b90ef530
+size 14512

checkpoint-200/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:177d534a379bd6b276474c2cb140e318dc65db4457b6c1b6f25a1a9dd563af82
+size 14512

checkpoint-200/trainer_state.json

@@ -0,0 +1,1421 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.19990004997501248,
+  "eval_steps": 500,
+  "global_step": 200,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "grad_norm": 3.3340563149001086,
+      "learning_rate": 0.0,
+      "loss": 11.0,
+      "step": 1
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.398812329952019,
+      "learning_rate": 5.9999999999999995e-05,
+      "loss": 10.125,
+      "step": 2
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 2.394322446895115,
+      "learning_rate": 0.00011999999999999999,
+      "loss": 10.1172,
+      "step": 3
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.9958816684399585,
+      "learning_rate": 0.00017999999999999998,
+      "loss": 9.875,
+      "step": 4
+    },
+    {
+      "epoch": 0.0,
+      "grad_norm": 1.8270465897882062,
+      "learning_rate": 0.00023999999999999998,
+      "loss": 9.6641,
+      "step": 5
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.7854046471397795,
+      "learning_rate": 0.0003,
+      "loss": 9.4844,
+      "step": 6
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.719416749115252,
+      "learning_rate": 0.00035999999999999997,
+      "loss": 9.3281,
+      "step": 7
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.4637825746112274,
+      "learning_rate": 0.00041999999999999996,
+      "loss": 9.2109,
+      "step": 8
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.4393631015406718,
+      "learning_rate": 0.00047999999999999996,
+      "loss": 8.9453,
+      "step": 9
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.2936734586915988,
+      "learning_rate": 0.00054,
+      "loss": 8.7109,
+      "step": 10
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.0756922378227356,
+      "learning_rate": 0.0005999986405514987,
+      "loss": 8.4609,
+      "step": 11
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.9277829127413892,
+      "learning_rate": 0.0005999945622196846,
+      "loss": 8.2344,
+      "step": 12
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.8084581786682467,
+      "learning_rate": 0.0005999877650456265,
+      "loss": 8.125,
+      "step": 13
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.7635084596900947,
+      "learning_rate": 0.000599978249097772,
+      "loss": 7.9766,
+      "step": 14
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 0.9186699644247788,
+      "learning_rate": 0.0005999660144719463,
+      "loss": 7.8555,
+      "step": 15
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6609504256551479,
+      "learning_rate": 0.0005999510612913519,
+      "loss": 7.7734,
+      "step": 16
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.7086232844782971,
+      "learning_rate": 0.0005999333897065673,
+      "loss": 7.7148,
+      "step": 17
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 16.38048851691348,
+      "learning_rate": 0.0005999129998955453,
+      "loss": 8.4844,
+      "step": 18
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 1.3057527590449889,
+      "learning_rate": 0.0005998898920636111,
+      "loss": 7.7539,
+      "step": 19
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6966048242948986,
+      "learning_rate": 0.00059986406644346,
+      "loss": 7.75,
+      "step": 20
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6348089115348993,
+      "learning_rate": 0.0005998355232951559,
+      "loss": 7.7031,
+      "step": 21
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.7829163518610293,
+      "learning_rate": 0.0005998042629061279,
+      "loss": 7.6992,
+      "step": 22
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.5900591778980369,
+      "learning_rate": 0.0005997702855911678,
+      "loss": 7.6016,
+      "step": 23
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.4655170213064256,
+      "learning_rate": 0.0005997335916924268,
+      "loss": 7.5977,
+      "step": 24
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.6287348258915756,
+      "learning_rate": 0.0005996941815794121,
+      "loss": 7.5586,
+      "step": 25
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.6137321903884564,
+      "learning_rate": 0.0005996520556489831,
+      "loss": 7.5898,
+      "step": 26
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.44962562710631065,
+      "learning_rate": 0.0005996072143253473,
+      "loss": 7.4336,
+      "step": 27
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.46130046454703316,
+      "learning_rate": 0.0005995596580600566,
+      "loss": 7.4023,
+      "step": 28
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.4686712675731326,
+      "learning_rate": 0.0005995093873320018,
+      "loss": 7.3789,
+      "step": 29
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.4672147564288997,
+      "learning_rate": 0.0005994564026474087,
+      "loss": 7.3711,
+      "step": 30
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.40408354581233474,
+      "learning_rate": 0.0005994007045398324,
+      "loss": 7.3672,
+      "step": 31
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.46032146732584733,
+      "learning_rate": 0.0005993422935701524,
+      "loss": 7.3477,
+      "step": 32
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.4765534634593268,
+      "learning_rate": 0.0005992811703265664,
+      "loss": 7.3555,
+      "step": 33
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.46208489386235113,
+      "learning_rate": 0.0005992173354245849,
+      "loss": 7.3047,
+      "step": 34
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 0.2956144524964961,
+      "learning_rate": 0.0005991507895070244,
+      "loss": 7.3125,
+      "step": 35
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.4834645389868856,
+      "learning_rate": 0.0005990815332440017,
+      "loss": 7.207,
+      "step": 36
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.4411831350968505,
+      "learning_rate": 0.0005990095673329266,
+      "loss": 7.1758,
+      "step": 37
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.24809297748968667,
+      "learning_rate": 0.0005989348924984951,
+      "loss": 7.2188,
+      "step": 38
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.39402988416840584,
+      "learning_rate": 0.0005988575094926817,
+      "loss": 7.1953,
+      "step": 39
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.3868345222189167,
+      "learning_rate": 0.0005987774190947328,
+      "loss": 7.1641,
+      "step": 40
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.3777261230135448,
+      "learning_rate": 0.0005986946221111575,
+      "loss": 7.1328,
+      "step": 41
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.4687511444077827,
+      "learning_rate": 0.0005986091193757206,
+      "loss": 7.0898,
+      "step": 42
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.34935796211612463,
+      "learning_rate": 0.0005985209117494337,
+      "loss": 7.1367,
+      "step": 43
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.38764476686849886,
+      "learning_rate": 0.0005984300001205466,
+      "loss": 7.125,
+      "step": 44
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.3956487898882936,
+      "learning_rate": 0.0005983363854045386,
+      "loss": 7.1094,
+      "step": 45
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.31140257544677513,
+      "learning_rate": 0.0005982400685441084,
+      "loss": 7.0898,
+      "step": 46
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.3664476570531787,
+      "learning_rate": 0.0005981410505091662,
+      "loss": 7.0664,
+      "step": 47
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.31891741142945207,
+      "learning_rate": 0.0005980393322968223,
+      "loss": 7.0273,
+      "step": 48
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.4533529037337155,
+      "learning_rate": 0.0005979349149313778,
+      "loss": 7.0586,
+      "step": 49
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.30532331638835586,
+      "learning_rate": 0.0005978277994643147,
+      "loss": 7.0195,
+      "step": 50
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.6501991746260075,
+      "learning_rate": 0.0005977179869742844,
+      "loss": 6.9648,
+      "step": 51
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.43904455901717926,
+      "learning_rate": 0.0005976054785670975,
+      "loss": 6.9805,
+      "step": 52
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.4826001598483571,
+      "learning_rate": 0.0005974902753757124,
+      "loss": 6.9297,
+      "step": 53
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.2924998027034648,
+      "learning_rate": 0.000597372378560224,
+      "loss": 6.8984,
+      "step": 54
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.4439033666380787,
+      "learning_rate": 0.0005972517893078517,
+      "loss": 6.8945,
+      "step": 55
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.6135914255073411,
+      "learning_rate": 0.0005971285088329284,
+      "loss": 6.9727,
+      "step": 56
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.5575686565598483,
+      "learning_rate": 0.0005970025383768866,
+      "loss": 6.9219,
+      "step": 57
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.4820951675994578,
+      "learning_rate": 0.0005968738792082478,
+      "loss": 6.8516,
+      "step": 58
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.40164190019465584,
+      "learning_rate": 0.0005967425326226082,
+      "loss": 6.7734,
+      "step": 59
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.46129863945181293,
+      "learning_rate": 0.0005966084999426265,
+      "loss": 6.8125,
+      "step": 60
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.33322355827118677,
+      "learning_rate": 0.0005964717825180101,
+      "loss": 6.7891,
+      "step": 61
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.3847525153855558,
+      "learning_rate": 0.0005963323817255024,
+      "loss": 6.8242,
+      "step": 62
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.3384433591375982,
+      "learning_rate": 0.0005961902989688674,
+      "loss": 6.707,
+      "step": 63
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.3937003195165685,
+      "learning_rate": 0.000596045535678877,
+      "loss": 6.8203,
+      "step": 64
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.35423488053528107,
+      "learning_rate": 0.0005958980933132962,
+      "loss": 6.7383,
+      "step": 65
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.36005939745315396,
+      "learning_rate": 0.0005957479733568675,
+      "loss": 6.7109,
+      "step": 66
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3499278317706933,
+      "learning_rate": 0.0005955951773212976,
+      "loss": 6.7266,
+      "step": 67
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3708385192137018,
+      "learning_rate": 0.0005954397067452407,
+      "loss": 6.7617,
+      "step": 68
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3775657656205869,
+      "learning_rate": 0.0005952815631942839,
+      "loss": 6.7148,
+      "step": 69
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3040083750375816,
+      "learning_rate": 0.0005951207482609307,
+      "loss": 6.5938,
+      "step": 70
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3443020808841468,
+      "learning_rate": 0.0005949572635645861,
+      "loss": 6.6523,
+      "step": 71
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3520066316939,
+      "learning_rate": 0.0005947911107515389,
+      "loss": 6.6211,
+      "step": 72
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.3739040572679613,
+      "learning_rate": 0.0005946222914949462,
+      "loss": 6.5547,
+      "step": 73
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.34890731989025553,
+      "learning_rate": 0.000594450807494816,
+      "loss": 6.5859,
+      "step": 74
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.40910932350136514,
+      "learning_rate": 0.0005942766604779903,
+      "loss": 6.5547,
+      "step": 75
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5698342865852906,
+      "learning_rate": 0.0005940998521981274,
+      "loss": 6.457,
+      "step": 76
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5179452709555474,
+      "learning_rate": 0.0005939203844356852,
+      "loss": 6.5547,
+      "step": 77
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5222512938673792,
+      "learning_rate": 0.0005937382589979016,
+      "loss": 6.5039,
+      "step": 78
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.5682332793686307,
+      "learning_rate": 0.0005935534777187781,
+      "loss": 6.5547,
+      "step": 79
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3869287710460676,
+      "learning_rate": 0.0005933660424590598,
+      "loss": 6.5156,
+      "step": 80
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3078211032807607,
+      "learning_rate": 0.000593175955106218,
+      "loss": 6.4258,
+      "step": 81
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3611357511872241,
+      "learning_rate": 0.00059298321757443,
+      "loss": 6.4727,
+      "step": 82
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.29633467844266953,
+      "learning_rate": 0.0005927878318045608,
+      "loss": 6.3281,
+      "step": 83
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.3257574200776832,
+      "learning_rate": 0.0005925897997641426,
+      "loss": 6.3203,
+      "step": 84
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.2824054533852328,
+      "learning_rate": 0.0005923891234473562,
+      "loss": 6.4062,
+      "step": 85
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.3056199770204573,
+      "learning_rate": 0.0005921858048750097,
+      "loss": 6.3984,
+      "step": 86
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.2966438824341908,
+      "learning_rate": 0.000591979846094519,
+      "loss": 6.3555,
+      "step": 87
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.32782438676663733,
+      "learning_rate": 0.0005917712491798866,
+      "loss": 6.4023,
+      "step": 88
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.3538316399620157,
+      "learning_rate": 0.0005915600162316811,
+      "loss": 6.2812,
+      "step": 89
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.375858298192913,
+      "learning_rate": 0.0005913461493770162,
+      "loss": 6.3086,
+      "step": 90
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.5189251339815161,
+      "learning_rate": 0.0005911296507695284,
+      "loss": 6.2812,
+      "step": 91
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.6304909542669104,
+      "learning_rate": 0.0005909105225893564,
+      "loss": 6.2969,
+      "step": 92
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.4655662819622591,
+      "learning_rate": 0.0005906887670431187,
+      "loss": 6.1953,
+      "step": 93
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.39035390983920965,
+      "learning_rate": 0.000590464386363891,
+      "loss": 6.2617,
+      "step": 94
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.4918417851770978,
+      "learning_rate": 0.0005902373828111843,
+      "loss": 6.2148,
+      "step": 95
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.35670770889552555,
+      "learning_rate": 0.0005900077586709219,
+      "loss": 6.2461,
+      "step": 96
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.4177985869939347,
+      "learning_rate": 0.0005897755162554163,
+      "loss": 6.1797,
+      "step": 97
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.3742471130708234,
+      "learning_rate": 0.000589540657903346,
+      "loss": 6.1406,
+      "step": 98
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.28627666723978284,
+      "learning_rate": 0.0005893031859797322,
+      "loss": 6.2031,
+      "step": 99
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.32238563846046103,
+      "learning_rate": 0.0005890631028759143,
+      "loss": 6.0625,
+      "step": 100
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.2556625657587849,
+      "learning_rate": 0.0005888204110095265,
+      "loss": 6.1797,
+      "step": 101
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.35463629701710253,
+      "learning_rate": 0.0005885751128244734,
+      "loss": 6.125,
+      "step": 102
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.31975770214936095,
+      "learning_rate": 0.0005883272107909048,
+      "loss": 6.1836,
+      "step": 103
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.3464621815245048,
+      "learning_rate": 0.0005880767074051915,
+      "loss": 6.125,
+      "step": 104
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.3663428920796654,
+      "learning_rate": 0.0005878236051898998,
+      "loss": 6.0781,
+      "step": 105
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.31594460565215293,
+      "learning_rate": 0.0005875679066937664,
+      "loss": 6.082,
+      "step": 106
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.3552617109396582,
+      "learning_rate": 0.000587309614491672,
+      "loss": 6.1016,
+      "step": 107
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.307016409692456,
+      "learning_rate": 0.0005870487311846164,
+      "loss": 6.1406,
+      "step": 108
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.32188902148474213,
+      "learning_rate": 0.0005867852593996914,
+      "loss": 6.0039,
+      "step": 109
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.25501199715105083,
+      "learning_rate": 0.0005865192017900551,
+      "loss": 6.0938,
+      "step": 110
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.3416203070024056,
+      "learning_rate": 0.0005862505610349049,
+      "loss": 6.0234,
+      "step": 111
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.3562508875852537,
+      "learning_rate": 0.0005859793398394498,
+      "loss": 6.0469,
+      "step": 112
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.4443953757302568,
+      "learning_rate": 0.0005857055409348845,
+      "loss": 5.9766,
+      "step": 113
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.42023839332714596,
+      "learning_rate": 0.0005854291670783607,
+      "loss": 6.0781,
+      "step": 114
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.4618323255809241,
+      "learning_rate": 0.0005851502210529604,
+      "loss": 5.9727,
+      "step": 115
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.379195014798667,
+      "learning_rate": 0.0005848687056676668,
+      "loss": 5.9922,
+      "step": 116
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.3931552573296799,
+      "learning_rate": 0.0005845846237573366,
+      "loss": 5.9492,
+      "step": 117
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.2567080044949908,
+      "learning_rate": 0.0005842979781826717,
+      "loss": 6.0273,
+      "step": 118
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.4190305965377807,
+      "learning_rate": 0.0005840087718301895,
+      "loss": 6.0391,
+      "step": 119
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.3996803869430228,
+      "learning_rate": 0.0005837170076121951,
+      "loss": 5.9531,
+      "step": 120
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.478219248015785,
+      "learning_rate": 0.000583422688466751,
+      "loss": 6.0586,
+      "step": 121
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.40869844309811526,
+      "learning_rate": 0.0005831258173576474,
+      "loss": 6.0117,
+      "step": 122
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.3728598080697978,
+      "learning_rate": 0.0005828263972743733,
+      "loss": 5.9375,
+      "step": 123
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.3560055462882015,
+      "learning_rate": 0.0005825244312320856,
+      "loss": 5.9531,
+      "step": 124
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.40446932887864323,
+      "learning_rate": 0.0005822199222715787,
+      "loss": 5.9609,
+      "step": 125
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.38514065739946723,
+      "learning_rate": 0.000581912873459255,
+      "loss": 5.8594,
+      "step": 126
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.35367576386319416,
+      "learning_rate": 0.0005816032878870921,
+      "loss": 5.9023,
+      "step": 127
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.3341681995122829,
+      "learning_rate": 0.0005812911686726135,
+      "loss": 5.9062,
+      "step": 128
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.3387022688975784,
+      "learning_rate": 0.0005809765189588563,
+      "loss": 5.8945,
+      "step": 129
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.31638659898934757,
+      "learning_rate": 0.0005806593419143395,
+      "loss": 5.8242,
+      "step": 130
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.3229678508227436,
+      "learning_rate": 0.0005803396407330325,
+      "loss": 5.8516,
+      "step": 131
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.35499490868584455,
+      "learning_rate": 0.0005800174186343226,
+      "loss": 5.9258,
+      "step": 132
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.40753171542848754,
+      "learning_rate": 0.0005796926788629828,
+      "loss": 5.8242,
+      "step": 133
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.3625374018348824,
+      "learning_rate": 0.0005793654246891389,
+      "loss": 5.832,
+      "step": 134
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.3583489573569317,
+      "learning_rate": 0.000579035659408237,
+      "loss": 5.8398,
+      "step": 135
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.39657706318861896,
+      "learning_rate": 0.0005787033863410095,
+      "loss": 5.8633,
+      "step": 136
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.3965837889564036,
+      "learning_rate": 0.0005783686088334428,
+      "loss": 5.8633,
+      "step": 137
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.29496474301865566,
+      "learning_rate": 0.0005780313302567424,
+      "loss": 5.8203,
+      "step": 138
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.44637192639243695,
+      "learning_rate": 0.0005776915540073001,
+      "loss": 5.8477,
+      "step": 139
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.39605473508683114,
+      "learning_rate": 0.0005773492835066587,
+      "loss": 5.7383,
+      "step": 140
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.3008962634266945,
+      "learning_rate": 0.0005770045222014786,
+      "loss": 5.7617,
+      "step": 141
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.36915495506607826,
+      "learning_rate": 0.0005766572735635022,
+      "loss": 5.7695,
+      "step": 142
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.3282300349560706,
+      "learning_rate": 0.0005763075410895193,
+      "loss": 5.8281,
+      "step": 143
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.2747449814083844,
+      "learning_rate": 0.0005759553283013323,
+      "loss": 5.7812,
+      "step": 144
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.28905882704179764,
+      "learning_rate": 0.00057560063874572,
+      "loss": 5.7344,
+      "step": 145
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.280625988867192,
+      "learning_rate": 0.000575243475994402,
+      "loss": 5.7773,
+      "step": 146
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.41061863948012467,
+      "learning_rate": 0.0005748838436440035,
+      "loss": 5.7578,
+      "step": 147
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.4920152483870267,
+      "learning_rate": 0.0005745217453160183,
+      "loss": 5.7305,
+      "step": 148
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.5463207978955044,
+      "learning_rate": 0.0005741571846567725,
+      "loss": 5.7383,
+      "step": 149
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.3986359831157306,
+      "learning_rate": 0.0005737901653373878,
+      "loss": 5.668,
+      "step": 150
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.37908758170100293,
+      "learning_rate": 0.0005734206910537447,
+      "loss": 5.6875,
+      "step": 151
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.35929793070492694,
+      "learning_rate": 0.0005730487655264451,
+      "loss": 5.7188,
+      "step": 152
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.4217799574145456,
+      "learning_rate": 0.0005726743925007751,
+      "loss": 5.7305,
+      "step": 153
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.4024411981587195,
+      "learning_rate": 0.0005722975757466667,
+      "loss": 5.6289,
+      "step": 154
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.3472391905877033,
+      "learning_rate": 0.0005719183190586606,
+      "loss": 5.6523,
+      "step": 155
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.31752956812138816,
+      "learning_rate": 0.0005715366262558675,
+      "loss": 5.6172,
+      "step": 156
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.3170152384332457,
+      "learning_rate": 0.0005711525011819294,
+      "loss": 5.6172,
+      "step": 157
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.40520629326601837,
+      "learning_rate": 0.0005707659477049818,
+      "loss": 5.625,
+      "step": 158
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.3965976910198806,
+      "learning_rate": 0.0005703769697176137,
+      "loss": 5.6562,
+      "step": 159
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.40422960541801994,
+      "learning_rate": 0.0005699855711368293,
+      "loss": 5.6836,
+      "step": 160
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.3780813184050647,
+      "learning_rate": 0.0005695917559040079,
+      "loss": 5.5938,
+      "step": 161
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.36917638857736573,
+      "learning_rate": 0.0005691955279848645,
+      "loss": 5.668,
+      "step": 162
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.37769176081037814,
+      "learning_rate": 0.0005687968913694098,
+      "loss": 5.4961,
+      "step": 163
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.3255116524991148,
+      "learning_rate": 0.0005683958500719103,
+      "loss": 5.5117,
+      "step": 164
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.31897629016848805,
+      "learning_rate": 0.0005679924081308471,
+      "loss": 5.5664,
+      "step": 165
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.2869064236553046,
+      "learning_rate": 0.0005675865696088764,
+      "loss": 5.5391,
+      "step": 166
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.29226729022634845,
+      "learning_rate": 0.0005671783385927873,
+      "loss": 5.5586,
+      "step": 167
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.2534117210955766,
+      "learning_rate": 0.0005667677191934618,
+      "loss": 5.5312,
+      "step": 168
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.289828484125484,
+      "learning_rate": 0.0005663547155458326,
+      "loss": 5.6484,
+      "step": 169
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.2717242930342115,
+      "learning_rate": 0.0005659393318088419,
+      "loss": 5.5352,
+      "step": 170
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.3595538109137759,
+      "learning_rate": 0.0005655215721653993,
+      "loss": 5.5742,
+      "step": 171
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.4255054350471108,
+      "learning_rate": 0.0005651014408223398,
+      "loss": 5.5469,
+      "step": 172
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.3670561941219979,
+      "learning_rate": 0.0005646789420103814,
+      "loss": 5.5078,
+      "step": 173
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.40280130904983164,
+      "learning_rate": 0.0005642540799840822,
+      "loss": 5.5,
+      "step": 174
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.41159472035983025,
+      "learning_rate": 0.0005638268590217984,
+      "loss": 5.5039,
+      "step": 175
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.4316778037513652,
+      "learning_rate": 0.0005633972834256401,
+      "loss": 5.5352,
+      "step": 176
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.5674781128363939,
+      "learning_rate": 0.000562965357521429,
+      "loss": 5.4336,
+      "step": 177
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.41654662151365446,
+      "learning_rate": 0.0005625310856586541,
+      "loss": 5.6211,
+      "step": 178
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.5159976364107484,
+      "learning_rate": 0.0005620944722104282,
+      "loss": 5.4844,
+      "step": 179
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.34364678177014185,
+      "learning_rate": 0.0005616555215734438,
+      "loss": 5.4922,
+      "step": 180
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.3708077784459011,
+      "learning_rate": 0.0005612142381679289,
+      "loss": 5.5234,
+      "step": 181
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.3620051253453866,
+      "learning_rate": 0.0005607706264376028,
+      "loss": 5.4961,
+      "step": 182
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.34735585210929654,
+      "learning_rate": 0.0005603246908496305,
+      "loss": 5.4453,
+      "step": 183
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.37719874705792217,
+      "learning_rate": 0.0005598764358945783,
+      "loss": 5.4844,
+      "step": 184
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.3749130664831207,
+      "learning_rate": 0.0005594258660863689,
+      "loss": 5.4648,
+      "step": 185
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.40951353306235827,
+      "learning_rate": 0.0005589729859622351,
+      "loss": 5.5039,
+      "step": 186
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.40146882563949804,
+      "learning_rate": 0.0005585178000826745,
+      "loss": 5.3672,
+      "step": 187
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.4062987628428303,
+      "learning_rate": 0.0005580603130314043,
+      "loss": 5.3984,
+      "step": 188
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.35626322654799136,
+      "learning_rate": 0.0005576005294153138,
+      "loss": 5.3984,
+      "step": 189
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.3140647930801716,
+      "learning_rate": 0.0005571384538644188,
+      "loss": 5.3906,
+      "step": 190
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.2990060538353662,
+      "learning_rate": 0.0005566740910318153,
+      "loss": 5.3711,
+      "step": 191
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.3337525907515936,
+      "learning_rate": 0.0005562074455936315,
+      "loss": 5.4023,
+      "step": 192
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.3381587051014816,
+      "learning_rate": 0.000555738522248982,
+      "loss": 5.4414,
+      "step": 193
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.2954008999469894,
+      "learning_rate": 0.0005552673257199197,
+      "loss": 5.418,
+      "step": 194
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.3242310900810155,
+      "learning_rate": 0.0005547938607513882,
+      "loss": 5.418,
+      "step": 195
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.3149021804393678,
+      "learning_rate": 0.0005543181321111747,
+      "loss": 5.4375,
+      "step": 196
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.32859412218218814,
+      "learning_rate": 0.0005538401445898612,
+      "loss": 5.4492,
+      "step": 197
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.2960282598050701,
+      "learning_rate": 0.0005533599030007768,
+      "loss": 5.3867,
+      "step": 198
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.2866762878199755,
+      "learning_rate": 0.0005528774121799489,
+      "loss": 5.3789,
+      "step": 199
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.34865216327038784,
+      "learning_rate": 0.0005523926769860549,
+      "loss": 5.3711,
+      "step": 200
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 1000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 100,
+  "total_flos": 0.0,
+  "train_batch_size": 32,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-200/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbd9a6067cf818494e2505097746a1cad30533fc72eb13916de34f7671e3e456
+size 6520

checkpoint-200/zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/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, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    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, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_file,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoint-25/config.json

@@ -0,0 +1,31 @@
+{
+  "_name_or_path": "georgeyw/gpt-2-small-init-seed-5",
+  "architectures": [
+    "GPTNeoXForCausalLM"
+  ],
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "classifier_dropout": 0.1,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 1024,
+  "model_type": "gpt_neox",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "rope_scaling": null,
+  "rotary_emb_base": 10000,
+  "rotary_pct": 0.25,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.2",
+  "use_cache": true,
+  "use_parallel_residual": true,
+  "vocab_size": 50304
+}