Add files using upload-large-folder tool

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1000/added_tokens.json

@@ -0,0 +1,4 @@
+{
+  "[PAD]": 50257,
+  "🅁": 50258
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1000/latest

@@ -0,0 +1 @@
+global_step1000

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1000/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)

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1800/added_tokens.json

@@ -0,0 +1,4 @@
+{
+  "[PAD]": 50257,
+  "🅁": 50258
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1800/config.json

@@ -0,0 +1,39 @@
+{
+  "_name_or_path": "gpt2",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "torch_dtype": "float16",
+  "transformers_version": "4.45.1",
+  "use_cache": true,
+  "vocab_size": 50257
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1800/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.45.1"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1800/latest

@@ -0,0 +1 @@
+global_step1800

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1800/special_tokens_map.json

@@ -0,0 +1,12 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1800/tokenizer_config.json

@@ -0,0 +1,36 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "50256": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50257": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50258": {
+      "content": "🅁",
+      "lstrip": true,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1800/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)

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1900/config.json

@@ -0,0 +1,39 @@
+{
+  "_name_or_path": "gpt2",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "torch_dtype": "float16",
+  "transformers_version": "4.45.1",
+  "use_cache": true,
+  "vocab_size": 50257
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1900/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.45.1"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1900/special_tokens_map.json

@@ -0,0 +1,12 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-1900/tokenizer_config.json

@@ -0,0 +1,36 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "50256": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50257": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50258": {
+      "content": "🅁",
+      "lstrip": true,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-200/added_tokens.json

@@ -0,0 +1,4 @@
+{
+  "[PAD]": 50257,
+  "🅁": 50258
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-200/config.json

@@ -0,0 +1,39 @@
+{
+  "_name_or_path": "gpt2",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "torch_dtype": "float16",
+  "transformers_version": "4.45.1",
+  "use_cache": true,
+  "vocab_size": 50257
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-200/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.45.1"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-200/latest

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

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-200/special_tokens_map.json

@@ -0,0 +1,12 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-200/tokenizer_config.json

@@ -0,0 +1,36 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "50256": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50257": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50258": {
+      "content": "🅁",
+      "lstrip": true,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-200/trainer_state.json

@@ -0,0 +1,1593 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.2054442732408834,
+  "eval_steps": 10,
+  "global_step": 200,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.001027221366204417,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8613,
+      "step": 1
+    },
+    {
+      "epoch": 0.002054442732408834,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8757,
+      "step": 2
+    },
+    {
+      "epoch": 0.0030816640986132513,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9408,
+      "step": 3
+    },
+    {
+      "epoch": 0.004108885464817668,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9219,
+      "step": 4
+    },
+    {
+      "epoch": 0.005136106831022085,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9043,
+      "step": 5
+    },
+    {
+      "epoch": 0.0061633281972265025,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8965,
+      "step": 6
+    },
+    {
+      "epoch": 0.007190549563430919,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8307,
+      "step": 7
+    },
+    {
+      "epoch": 0.008217770929635337,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9173,
+      "step": 8
+    },
+    {
+      "epoch": 0.009244992295839754,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9069,
+      "step": 9
+    },
+    {
+      "epoch": 0.01027221366204417,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9076,
+      "step": 10
+    },
+    {
+      "epoch": 0.01027221366204417,
+      "eval_loss": 4.63671875,
+      "eval_runtime": 8.1008,
+      "eval_samples_per_second": 123.444,
+      "eval_steps_per_second": 5.185,
+      "step": 10
+    },
+    {
+      "epoch": 0.011299435028248588,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9036,
+      "step": 11
+    },
+    {
+      "epoch": 0.012326656394453005,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8652,
+      "step": 12
+    },
+    {
+      "epoch": 0.01335387776065742,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9336,
+      "step": 13
+    },
+    {
+      "epoch": 0.014381099126861838,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9245,
+      "step": 14
+    },
+    {
+      "epoch": 0.015408320493066256,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9336,
+      "step": 15
+    },
+    {
+      "epoch": 0.016435541859270673,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8516,
+      "step": 16
+    },
+    {
+      "epoch": 0.01746276322547509,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8704,
+      "step": 17
+    },
+    {
+      "epoch": 0.01848998459167951,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.9004,
+      "step": 18
+    },
+    {
+      "epoch": 0.019517205957883924,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8112,
+      "step": 19
+    },
+    {
+      "epoch": 0.02054442732408834,
+      "grad_norm": 0.0,
+      "learning_rate": 0.0,
+      "loss": 4.8672,
+      "step": 20
+    },
+    {
+      "epoch": 0.02054442732408834,
+      "eval_loss": 4.63671875,
+      "eval_runtime": 8.0523,
+      "eval_samples_per_second": 124.188,
+      "eval_steps_per_second": 5.216,
+      "step": 20
+    },
+    {
+      "epoch": 0.02157164869029276,
+      "grad_norm": 28.116819381713867,
+      "learning_rate": 1.7123287671232876e-08,
+      "loss": 4.9049,
+      "step": 21
+    },
+    {
+      "epoch": 0.022598870056497175,
+      "grad_norm": 27.03044319152832,
+      "learning_rate": 3.424657534246575e-08,
+      "loss": 4.918,
+      "step": 22
+    },
+    {
+      "epoch": 0.02362609142270159,
+      "grad_norm": 27.387975692749023,
+      "learning_rate": 5.136986301369863e-08,
+      "loss": 4.86,
+      "step": 23
+    },
+    {
+      "epoch": 0.02465331278890601,
+      "grad_norm": 26.709815979003906,
+      "learning_rate": 6.84931506849315e-08,
+      "loss": 4.8659,
+      "step": 24
+    },
+    {
+      "epoch": 0.025680534155110426,
+      "grad_norm": 28.10965919494629,
+      "learning_rate": 8.561643835616439e-08,
+      "loss": 4.8789,
+      "step": 25
+    },
+    {
+      "epoch": 0.02670775552131484,
+      "grad_norm": 27.105060577392578,
+      "learning_rate": 1.0273972602739726e-07,
+      "loss": 4.8581,
+      "step": 26
+    },
+    {
+      "epoch": 0.02773497688751926,
+      "grad_norm": 28.362043380737305,
+      "learning_rate": 1.1986301369863014e-07,
+      "loss": 4.8626,
+      "step": 27
+    },
+    {
+      "epoch": 0.028762198253723677,
+      "grad_norm": 26.980588912963867,
+      "learning_rate": 1.36986301369863e-07,
+      "loss": 4.873,
+      "step": 28
+    },
+    {
+      "epoch": 0.029789419619928096,
+      "grad_norm": 27.26178550720215,
+      "learning_rate": 1.541095890410959e-07,
+      "loss": 4.9076,
+      "step": 29
+    },
+    {
+      "epoch": 0.030816640986132512,
+      "grad_norm": 27.26181983947754,
+      "learning_rate": 1.7123287671232878e-07,
+      "loss": 4.9069,
+      "step": 30
+    },
+    {
+      "epoch": 0.030816640986132512,
+      "eval_loss": 4.63671875,
+      "eval_runtime": 8.1739,
+      "eval_samples_per_second": 122.341,
+      "eval_steps_per_second": 5.138,
+      "step": 30
+    },
+    {
+      "epoch": 0.03184386235233693,
+      "grad_norm": 27.67842674255371,
+      "learning_rate": 1.8835616438356165e-07,
+      "loss": 4.8991,
+      "step": 31
+    },
+    {
+      "epoch": 0.03287108371854135,
+      "grad_norm": 27.32350730895996,
+      "learning_rate": 2.0547945205479452e-07,
+      "loss": 4.9108,
+      "step": 32
+    },
+    {
+      "epoch": 0.03389830508474576,
+      "grad_norm": 27.842466354370117,
+      "learning_rate": 2.226027397260274e-07,
+      "loss": 4.8939,
+      "step": 33
+    },
+    {
+      "epoch": 0.03492552645095018,
+      "grad_norm": 28.03348159790039,
+      "learning_rate": 2.397260273972603e-07,
+      "loss": 4.8783,
+      "step": 34
+    },
+    {
+      "epoch": 0.035952747817154594,
+      "grad_norm": 27.308210372924805,
+      "learning_rate": 2.568493150684932e-07,
+      "loss": 4.877,
+      "step": 35
+    },
+    {
+      "epoch": 0.03697996918335902,
+      "grad_norm": 27.26631736755371,
+      "learning_rate": 2.73972602739726e-07,
+      "loss": 4.8281,
+      "step": 36
+    },
+    {
+      "epoch": 0.03800719054956343,
+      "grad_norm": 26.74373435974121,
+      "learning_rate": 2.910958904109589e-07,
+      "loss": 4.9264,
+      "step": 37
+    },
+    {
+      "epoch": 0.03903441191576785,
+      "grad_norm": 26.75712013244629,
+      "learning_rate": 3.082191780821918e-07,
+      "loss": 4.8672,
+      "step": 38
+    },
+    {
+      "epoch": 0.040061633281972264,
+      "grad_norm": 27.08635711669922,
+      "learning_rate": 3.2534246575342466e-07,
+      "loss": 4.8711,
+      "step": 39
+    },
+    {
+      "epoch": 0.04108885464817668,
+      "grad_norm": 27.5158634185791,
+      "learning_rate": 3.4246575342465755e-07,
+      "loss": 4.8984,
+      "step": 40
+    },
+    {
+      "epoch": 0.04108885464817668,
+      "eval_loss": 4.63671875,
+      "eval_runtime": 8.1684,
+      "eval_samples_per_second": 122.423,
+      "eval_steps_per_second": 5.142,
+      "step": 40
+    },
+    {
+      "epoch": 0.042116076014381096,
+      "grad_norm": 25.9942684173584,
+      "learning_rate": 3.595890410958904e-07,
+      "loss": 4.8079,
+      "step": 41
+    },
+    {
+      "epoch": 0.04314329738058552,
+      "grad_norm": 28.041112899780273,
+      "learning_rate": 3.767123287671233e-07,
+      "loss": 4.8405,
+      "step": 42
+    },
+    {
+      "epoch": 0.044170518746789934,
+      "grad_norm": 27.068256378173828,
+      "learning_rate": 3.938356164383562e-07,
+      "loss": 4.9121,
+      "step": 43
+    },
+    {
+      "epoch": 0.04519774011299435,
+      "grad_norm": 27.018674850463867,
+      "learning_rate": 4.1095890410958903e-07,
+      "loss": 4.8529,
+      "step": 44
+    },
+    {
+      "epoch": 0.046224961479198766,
+      "grad_norm": 27.157318115234375,
+      "learning_rate": 4.2808219178082193e-07,
+      "loss": 4.8307,
+      "step": 45
+    },
+    {
+      "epoch": 0.04725218284540318,
+      "grad_norm": 26.827917098999023,
+      "learning_rate": 4.452054794520548e-07,
+      "loss": 4.8385,
+      "step": 46
+    },
+    {
+      "epoch": 0.048279404211607604,
+      "grad_norm": 26.432662963867188,
+      "learning_rate": 4.6232876712328767e-07,
+      "loss": 4.8503,
+      "step": 47
+    },
+    {
+      "epoch": 0.04930662557781202,
+      "grad_norm": 26.891084671020508,
+      "learning_rate": 4.794520547945206e-07,
+      "loss": 4.9082,
+      "step": 48
+    },
+    {
+      "epoch": 0.050333846944016436,
+      "grad_norm": 26.872167587280273,
+      "learning_rate": 4.965753424657534e-07,
+      "loss": 4.86,
+      "step": 49
+    },
+    {
+      "epoch": 0.05136106831022085,
+      "grad_norm": 26.370975494384766,
+      "learning_rate": 5.136986301369864e-07,
+      "loss": 4.819,
+      "step": 50
+    },
+    {
+      "epoch": 0.05136106831022085,
+      "eval_loss": 4.6328125,
+      "eval_runtime": 8.1128,
+      "eval_samples_per_second": 123.262,
+      "eval_steps_per_second": 5.177,
+      "step": 50
+    },
+    {
+      "epoch": 0.05238828967642527,
+      "grad_norm": 27.79764747619629,
+      "learning_rate": 5.308219178082192e-07,
+      "loss": 4.8737,
+      "step": 51
+    },
+    {
+      "epoch": 0.05341551104262968,
+      "grad_norm": 25.080102920532227,
+      "learning_rate": 5.47945205479452e-07,
+      "loss": 4.8568,
+      "step": 52
+    },
+    {
+      "epoch": 0.054442732408834106,
+      "grad_norm": 26.24574089050293,
+      "learning_rate": 5.65068493150685e-07,
+      "loss": 4.7917,
+      "step": 53
+    },
+    {
+      "epoch": 0.05546995377503852,
+      "grad_norm": 25.477985382080078,
+      "learning_rate": 5.821917808219178e-07,
+      "loss": 4.8301,
+      "step": 54
+    },
+    {
+      "epoch": 0.05649717514124294,
+      "grad_norm": 25.411081314086914,
+      "learning_rate": 5.993150684931507e-07,
+      "loss": 4.8164,
+      "step": 55
+    },
+    {
+      "epoch": 0.05752439650744735,
+      "grad_norm": 24.9583797454834,
+      "learning_rate": 6.164383561643836e-07,
+      "loss": 4.7949,
+      "step": 56
+    },
+    {
+      "epoch": 0.05855161787365177,
+      "grad_norm": 26.102153778076172,
+      "learning_rate": 6.335616438356165e-07,
+      "loss": 4.8086,
+      "step": 57
+    },
+    {
+      "epoch": 0.05957883923985619,
+      "grad_norm": 26.286907196044922,
+      "learning_rate": 6.506849315068493e-07,
+      "loss": 4.8099,
+      "step": 58
+    },
+    {
+      "epoch": 0.06060606060606061,
+      "grad_norm": 25.195068359375,
+      "learning_rate": 6.678082191780823e-07,
+      "loss": 4.8685,
+      "step": 59
+    },
+    {
+      "epoch": 0.061633281972265024,
+      "grad_norm": 25.254196166992188,
+      "learning_rate": 6.849315068493151e-07,
+      "loss": 4.8203,
+      "step": 60
+    },
+    {
+      "epoch": 0.061633281972265024,
+      "eval_loss": 4.625,
+      "eval_runtime": 8.1304,
+      "eval_samples_per_second": 122.995,
+      "eval_steps_per_second": 5.166,
+      "step": 60
+    },
+    {
+      "epoch": 0.06266050333846944,
+      "grad_norm": 25.127748489379883,
+      "learning_rate": 7.020547945205481e-07,
+      "loss": 4.7956,
+      "step": 61
+    },
+    {
+      "epoch": 0.06368772470467386,
+      "grad_norm": 25.56139373779297,
+      "learning_rate": 7.191780821917808e-07,
+      "loss": 4.8073,
+      "step": 62
+    },
+    {
+      "epoch": 0.06471494607087827,
+      "grad_norm": 25.190975189208984,
+      "learning_rate": 7.363013698630137e-07,
+      "loss": 4.8353,
+      "step": 63
+    },
+    {
+      "epoch": 0.0657421674370827,
+      "grad_norm": 24.852304458618164,
+      "learning_rate": 7.534246575342466e-07,
+      "loss": 4.8236,
+      "step": 64
+    },
+    {
+      "epoch": 0.0667693888032871,
+      "grad_norm": 24.120141983032227,
+      "learning_rate": 7.705479452054795e-07,
+      "loss": 4.6797,
+      "step": 65
+    },
+    {
+      "epoch": 0.06779661016949153,
+      "grad_norm": 23.0483455657959,
+      "learning_rate": 7.876712328767124e-07,
+      "loss": 4.6842,
+      "step": 66
+    },
+    {
+      "epoch": 0.06882383153569595,
+      "grad_norm": 22.356775283813477,
+      "learning_rate": 8.047945205479453e-07,
+      "loss": 4.7311,
+      "step": 67
+    },
+    {
+      "epoch": 0.06985105290190036,
+      "grad_norm": 22.69537925720215,
+      "learning_rate": 8.219178082191781e-07,
+      "loss": 4.7272,
+      "step": 68
+    },
+    {
+      "epoch": 0.07087827426810478,
+      "grad_norm": 22.291152954101562,
+      "learning_rate": 8.39041095890411e-07,
+      "loss": 4.7109,
+      "step": 69
+    },
+    {
+      "epoch": 0.07190549563430919,
+      "grad_norm": 21.66524314880371,
+      "learning_rate": 8.561643835616439e-07,
+      "loss": 4.7904,
+      "step": 70
+    },
+    {
+      "epoch": 0.07190549563430919,
+      "eval_loss": 4.6015625,
+      "eval_runtime": 8.3582,
+      "eval_samples_per_second": 119.642,
+      "eval_steps_per_second": 5.025,
+      "step": 70
+    },
+    {
+      "epoch": 0.07293271700051361,
+      "grad_norm": 22.571956634521484,
+      "learning_rate": 8.732876712328768e-07,
+      "loss": 4.7285,
+      "step": 71
+    },
+    {
+      "epoch": 0.07395993836671803,
+      "grad_norm": 22.55990219116211,
+      "learning_rate": 8.904109589041097e-07,
+      "loss": 4.6576,
+      "step": 72
+    },
+    {
+      "epoch": 0.07498715973292244,
+      "grad_norm": 22.43810272216797,
+      "learning_rate": 9.075342465753426e-07,
+      "loss": 4.7552,
+      "step": 73
+    },
+    {
+      "epoch": 0.07601438109912687,
+      "grad_norm": 21.978988647460938,
+      "learning_rate": 9.246575342465753e-07,
+      "loss": 4.7461,
+      "step": 74
+    },
+    {
+      "epoch": 0.07704160246533127,
+      "grad_norm": 22.340654373168945,
+      "learning_rate": 9.417808219178083e-07,
+      "loss": 4.6986,
+      "step": 75
+    },
+    {
+      "epoch": 0.0780688238315357,
+      "grad_norm": 21.94169044494629,
+      "learning_rate": 9.589041095890411e-07,
+      "loss": 4.7038,
+      "step": 76
+    },
+    {
+      "epoch": 0.07909604519774012,
+      "grad_norm": 21.635208129882812,
+      "learning_rate": 9.76027397260274e-07,
+      "loss": 4.6777,
+      "step": 77
+    },
+    {
+      "epoch": 0.08012326656394453,
+      "grad_norm": 21.09151268005371,
+      "learning_rate": 9.931506849315068e-07,
+      "loss": 4.7773,
+      "step": 78
+    },
+    {
+      "epoch": 0.08115048793014895,
+      "grad_norm": 21.541440963745117,
+      "learning_rate": 1.0102739726027399e-06,
+      "loss": 4.6608,
+      "step": 79
+    },
+    {
+      "epoch": 0.08217770929635336,
+      "grad_norm": 21.233154296875,
+      "learning_rate": 1.0273972602739727e-06,
+      "loss": 4.7337,
+      "step": 80
+    },
+    {
+      "epoch": 0.08217770929635336,
+      "eval_loss": 4.58203125,
+      "eval_runtime": 8.0829,
+      "eval_samples_per_second": 123.719,
+      "eval_steps_per_second": 5.196,
+      "step": 80
+    },
+    {
+      "epoch": 0.08320493066255778,
+      "grad_norm": 21.584636688232422,
+      "learning_rate": 1.0445205479452056e-06,
+      "loss": 4.6458,
+      "step": 81
+    },
+    {
+      "epoch": 0.08423215202876219,
+      "grad_norm": 21.521387100219727,
+      "learning_rate": 1.0616438356164384e-06,
+      "loss": 4.6198,
+      "step": 82
+    },
+    {
+      "epoch": 0.08525937339496661,
+      "grad_norm": 21.139148712158203,
+      "learning_rate": 1.0787671232876712e-06,
+      "loss": 4.6927,
+      "step": 83
+    },
+    {
+      "epoch": 0.08628659476117104,
+      "grad_norm": 20.5504150390625,
+      "learning_rate": 1.095890410958904e-06,
+      "loss": 4.6426,
+      "step": 84
+    },
+    {
+      "epoch": 0.08731381612737545,
+      "grad_norm": 20.150310516357422,
+      "learning_rate": 1.1130136986301371e-06,
+      "loss": 4.6146,
+      "step": 85
+    },
+    {
+      "epoch": 0.08834103749357987,
+      "grad_norm": 19.10382652282715,
+      "learning_rate": 1.13013698630137e-06,
+      "loss": 4.6178,
+      "step": 86
+    },
+    {
+      "epoch": 0.08936825885978428,
+      "grad_norm": 18.845382690429688,
+      "learning_rate": 1.1472602739726028e-06,
+      "loss": 4.5703,
+      "step": 87
+    },
+    {
+      "epoch": 0.0903954802259887,
+      "grad_norm": 18.1544189453125,
+      "learning_rate": 1.1643835616438357e-06,
+      "loss": 4.6048,
+      "step": 88
+    },
+    {
+      "epoch": 0.09142270159219312,
+      "grad_norm": 17.512283325195312,
+      "learning_rate": 1.1815068493150685e-06,
+      "loss": 4.5866,
+      "step": 89
+    },
+    {
+      "epoch": 0.09244992295839753,
+      "grad_norm": 18.081344604492188,
+      "learning_rate": 1.1986301369863014e-06,
+      "loss": 4.582,
+      "step": 90
+    },
+    {
+      "epoch": 0.09244992295839753,
+      "eval_loss": 4.53125,
+      "eval_runtime": 8.1682,
+      "eval_samples_per_second": 122.426,
+      "eval_steps_per_second": 5.142,
+      "step": 90
+    },
+    {
+      "epoch": 0.09347714432460195,
+      "grad_norm": 17.38676643371582,
+      "learning_rate": 1.2157534246575344e-06,
+      "loss": 4.5514,
+      "step": 91
+    },
+    {
+      "epoch": 0.09450436569080636,
+      "grad_norm": 17.519012451171875,
+      "learning_rate": 1.2328767123287673e-06,
+      "loss": 4.5638,
+      "step": 92
+    },
+    {
+      "epoch": 0.09553158705701079,
+      "grad_norm": 17.74829864501953,
+      "learning_rate": 1.25e-06,
+      "loss": 4.5716,
+      "step": 93
+    },
+    {
+      "epoch": 0.09655880842321521,
+      "grad_norm": 17.831249237060547,
+      "learning_rate": 1.267123287671233e-06,
+      "loss": 4.5697,
+      "step": 94
+    },
+    {
+      "epoch": 0.09758602978941962,
+      "grad_norm": 17.400928497314453,
+      "learning_rate": 1.284246575342466e-06,
+      "loss": 4.5592,
+      "step": 95
+    },
+    {
+      "epoch": 0.09861325115562404,
+      "grad_norm": 17.711442947387695,
+      "learning_rate": 1.3013698630136986e-06,
+      "loss": 4.5078,
+      "step": 96
+    },
+    {
+      "epoch": 0.09964047252182845,
+      "grad_norm": 17.26605796813965,
+      "learning_rate": 1.3184931506849317e-06,
+      "loss": 4.5892,
+      "step": 97
+    },
+    {
+      "epoch": 0.10066769388803287,
+      "grad_norm": 17.803537368774414,
+      "learning_rate": 1.3356164383561645e-06,
+      "loss": 4.4668,
+      "step": 98
+    },
+    {
+      "epoch": 0.1016949152542373,
+      "grad_norm": 17.142738342285156,
+      "learning_rate": 1.3527397260273976e-06,
+      "loss": 4.5391,
+      "step": 99
+    },
+    {
+      "epoch": 0.1027221366204417,
+      "grad_norm": 17.18332290649414,
+      "learning_rate": 1.3698630136986302e-06,
+      "loss": 4.5384,
+      "step": 100
+    },
+    {
+      "epoch": 0.1027221366204417,
+      "eval_loss": 4.49609375,
+      "eval_runtime": 8.101,
+      "eval_samples_per_second": 123.442,
+      "eval_steps_per_second": 5.185,
+      "step": 100
+    },
+    {
+      "epoch": 0.10374935798664613,
+      "grad_norm": 17.503074645996094,
+      "learning_rate": 1.386986301369863e-06,
+      "loss": 4.5436,
+      "step": 101
+    },
+    {
+      "epoch": 0.10477657935285054,
+      "grad_norm": 17.03574562072754,
+      "learning_rate": 1.4041095890410961e-06,
+      "loss": 4.5189,
+      "step": 102
+    },
+    {
+      "epoch": 0.10580380071905496,
+      "grad_norm": 17.297348022460938,
+      "learning_rate": 1.421232876712329e-06,
+      "loss": 4.4512,
+      "step": 103
+    },
+    {
+      "epoch": 0.10683102208525937,
+      "grad_norm": 17.349184036254883,
+      "learning_rate": 1.4383561643835616e-06,
+      "loss": 4.5365,
+      "step": 104
+    },
+    {
+      "epoch": 0.10785824345146379,
+      "grad_norm": 16.68321990966797,
+      "learning_rate": 1.4554794520547946e-06,
+      "loss": 4.5026,
+      "step": 105
+    },
+    {
+      "epoch": 0.10888546481766821,
+      "grad_norm": 17.097679138183594,
+      "learning_rate": 1.4726027397260275e-06,
+      "loss": 4.5306,
+      "step": 106
+    },
+    {
+      "epoch": 0.10991268618387262,
+      "grad_norm": 16.458786010742188,
+      "learning_rate": 1.4897260273972605e-06,
+      "loss": 4.5091,
+      "step": 107
+    },
+    {
+      "epoch": 0.11093990755007704,
+      "grad_norm": 16.699987411499023,
+      "learning_rate": 1.5068493150684932e-06,
+      "loss": 4.5488,
+      "step": 108
+    },
+    {
+      "epoch": 0.11196712891628145,
+      "grad_norm": 16.068634033203125,
+      "learning_rate": 1.5239726027397262e-06,
+      "loss": 4.4023,
+      "step": 109
+    },
+    {
+      "epoch": 0.11299435028248588,
+      "grad_norm": 16.19753646850586,
+      "learning_rate": 1.541095890410959e-06,
+      "loss": 4.5267,
+      "step": 110
+    },
+    {
+      "epoch": 0.11299435028248588,
+      "eval_loss": 4.4296875,
+      "eval_runtime": 8.1416,
+      "eval_samples_per_second": 122.826,
+      "eval_steps_per_second": 5.159,
+      "step": 110
+    },
+    {
+      "epoch": 0.1140215716486903,
+      "grad_norm": 16.418561935424805,
+      "learning_rate": 1.5582191780821921e-06,
+      "loss": 4.418,
+      "step": 111
+    },
+    {
+      "epoch": 0.1150487930148947,
+      "grad_norm": 15.953530311584473,
+      "learning_rate": 1.5753424657534248e-06,
+      "loss": 4.4883,
+      "step": 112
+    },
+    {
+      "epoch": 0.11607601438109913,
+      "grad_norm": 15.907369613647461,
+      "learning_rate": 1.5924657534246576e-06,
+      "loss": 4.3965,
+      "step": 113
+    },
+    {
+      "epoch": 0.11710323574730354,
+      "grad_norm": 16.10055923461914,
+      "learning_rate": 1.6095890410958907e-06,
+      "loss": 4.4512,
+      "step": 114
+    },
+    {
+      "epoch": 0.11813045711350796,
+      "grad_norm": 15.759989738464355,
+      "learning_rate": 1.6267123287671235e-06,
+      "loss": 4.3919,
+      "step": 115
+    },
+    {
+      "epoch": 0.11915767847971238,
+      "grad_norm": 15.749765396118164,
+      "learning_rate": 1.6438356164383561e-06,
+      "loss": 4.5052,
+      "step": 116
+    },
+    {
+      "epoch": 0.12018489984591679,
+      "grad_norm": 15.092001914978027,
+      "learning_rate": 1.6609589041095892e-06,
+      "loss": 4.4375,
+      "step": 117
+    },
+    {
+      "epoch": 0.12121212121212122,
+      "grad_norm": 14.8633451461792,
+      "learning_rate": 1.678082191780822e-06,
+      "loss": 4.3997,
+      "step": 118
+    },
+    {
+      "epoch": 0.12223934257832562,
+      "grad_norm": 14.971479415893555,
+      "learning_rate": 1.695205479452055e-06,
+      "loss": 4.4102,
+      "step": 119
+    },
+    {
+      "epoch": 0.12326656394453005,
+      "grad_norm": 14.7716064453125,
+      "learning_rate": 1.7123287671232877e-06,
+      "loss": 4.4434,
+      "step": 120
+    },
+    {
+      "epoch": 0.12326656394453005,
+      "eval_loss": 4.38671875,
+      "eval_runtime": 8.1933,
+      "eval_samples_per_second": 122.051,
+      "eval_steps_per_second": 5.126,
+      "step": 120
+    },
+    {
+      "epoch": 0.12429378531073447,
+      "grad_norm": 14.274008750915527,
+      "learning_rate": 1.7294520547945206e-06,
+      "loss": 4.4596,
+      "step": 121
+    },
+    {
+      "epoch": 0.12532100667693888,
+      "grad_norm": 14.329883575439453,
+      "learning_rate": 1.7465753424657536e-06,
+      "loss": 4.3704,
+      "step": 122
+    },
+    {
+      "epoch": 0.1263482280431433,
+      "grad_norm": 14.691496849060059,
+      "learning_rate": 1.7636986301369865e-06,
+      "loss": 4.3444,
+      "step": 123
+    },
+    {
+      "epoch": 0.12737544940934772,
+      "grad_norm": 14.369722366333008,
+      "learning_rate": 1.7808219178082193e-06,
+      "loss": 4.3372,
+      "step": 124
+    },
+    {
+      "epoch": 0.12840267077555212,
+      "grad_norm": 13.581815719604492,
+      "learning_rate": 1.7979452054794521e-06,
+      "loss": 4.375,
+      "step": 125
+    },
+    {
+      "epoch": 0.12942989214175654,
+      "grad_norm": 13.743803024291992,
+      "learning_rate": 1.8150684931506852e-06,
+      "loss": 4.4056,
+      "step": 126
+    },
+    {
+      "epoch": 0.13045711350796096,
+      "grad_norm": 13.931293487548828,
+      "learning_rate": 1.832191780821918e-06,
+      "loss": 4.3522,
+      "step": 127
+    },
+    {
+      "epoch": 0.1314843348741654,
+      "grad_norm": 13.874761581420898,
+      "learning_rate": 1.8493150684931507e-06,
+      "loss": 4.3783,
+      "step": 128
+    },
+    {
+      "epoch": 0.1325115562403698,
+      "grad_norm": 14.007351875305176,
+      "learning_rate": 1.8664383561643837e-06,
+      "loss": 4.3535,
+      "step": 129
+    },
+    {
+      "epoch": 0.1335387776065742,
+      "grad_norm": 14.182611465454102,
+      "learning_rate": 1.8835616438356166e-06,
+      "loss": 4.3978,
+      "step": 130
+    },
+    {
+      "epoch": 0.1335387776065742,
+      "eval_loss": 4.35546875,
+      "eval_runtime": 8.1643,
+      "eval_samples_per_second": 122.485,
+      "eval_steps_per_second": 5.144,
+      "step": 130
+    },
+    {
+      "epoch": 0.13456599897277863,
+      "grad_norm": 13.796817779541016,
+      "learning_rate": 1.9006849315068496e-06,
+      "loss": 4.3529,
+      "step": 131
+    },
+    {
+      "epoch": 0.13559322033898305,
+      "grad_norm": 13.67389965057373,
+      "learning_rate": 1.9178082191780823e-06,
+      "loss": 4.4186,
+      "step": 132
+    },
+    {
+      "epoch": 0.13662044170518747,
+      "grad_norm": 13.272880554199219,
+      "learning_rate": 1.9349315068493153e-06,
+      "loss": 4.3965,
+      "step": 133
+    },
+    {
+      "epoch": 0.1376476630713919,
+      "grad_norm": 13.589006423950195,
+      "learning_rate": 1.952054794520548e-06,
+      "loss": 4.3307,
+      "step": 134
+    },
+    {
+      "epoch": 0.1386748844375963,
+      "grad_norm": 13.390036582946777,
+      "learning_rate": 1.969178082191781e-06,
+      "loss": 4.3639,
+      "step": 135
+    },
+    {
+      "epoch": 0.1397021058038007,
+      "grad_norm": 12.994545936584473,
+      "learning_rate": 1.9863013698630136e-06,
+      "loss": 4.3145,
+      "step": 136
+    },
+    {
+      "epoch": 0.14072932717000514,
+      "grad_norm": 13.275940895080566,
+      "learning_rate": 2.0034246575342467e-06,
+      "loss": 4.2676,
+      "step": 137
+    },
+    {
+      "epoch": 0.14175654853620956,
+      "grad_norm": 12.996981620788574,
+      "learning_rate": 2.0205479452054797e-06,
+      "loss": 4.4238,
+      "step": 138
+    },
+    {
+      "epoch": 0.14278376990241398,
+      "grad_norm": 12.925374031066895,
+      "learning_rate": 2.037671232876713e-06,
+      "loss": 4.2786,
+      "step": 139
+    },
+    {
+      "epoch": 0.14381099126861838,
+      "grad_norm": 12.948338508605957,
+      "learning_rate": 2.0547945205479454e-06,
+      "loss": 4.3424,
+      "step": 140
+    },
+    {
+      "epoch": 0.14381099126861838,
+      "eval_loss": 4.328125,
+      "eval_runtime": 8.3013,
+      "eval_samples_per_second": 120.462,
+      "eval_steps_per_second": 5.059,
+      "step": 140
+    },
+    {
+      "epoch": 0.1448382126348228,
+      "grad_norm": 12.943131446838379,
+      "learning_rate": 2.071917808219178e-06,
+      "loss": 4.2682,
+      "step": 141
+    },
+    {
+      "epoch": 0.14586543400102722,
+      "grad_norm": 11.763097763061523,
+      "learning_rate": 2.089041095890411e-06,
+      "loss": 4.276,
+      "step": 142
+    },
+    {
+      "epoch": 0.14689265536723164,
+      "grad_norm": 12.934331893920898,
+      "learning_rate": 2.106164383561644e-06,
+      "loss": 4.4095,
+      "step": 143
+    },
+    {
+      "epoch": 0.14791987673343607,
+      "grad_norm": 12.17612361907959,
+      "learning_rate": 2.123287671232877e-06,
+      "loss": 4.3053,
+      "step": 144
+    },
+    {
+      "epoch": 0.14894709809964046,
+      "grad_norm": 12.38232135772705,
+      "learning_rate": 2.14041095890411e-06,
+      "loss": 4.2467,
+      "step": 145
+    },
+    {
+      "epoch": 0.14997431946584489,
+      "grad_norm": 12.592203140258789,
+      "learning_rate": 2.1575342465753425e-06,
+      "loss": 4.2773,
+      "step": 146
+    },
+    {
+      "epoch": 0.1510015408320493,
+      "grad_norm": 12.006210327148438,
+      "learning_rate": 2.1746575342465755e-06,
+      "loss": 4.2572,
+      "step": 147
+    },
+    {
+      "epoch": 0.15202876219825373,
+      "grad_norm": 12.58430290222168,
+      "learning_rate": 2.191780821917808e-06,
+      "loss": 4.2995,
+      "step": 148
+    },
+    {
+      "epoch": 0.15305598356445815,
+      "grad_norm": 12.709675788879395,
+      "learning_rate": 2.2089041095890412e-06,
+      "loss": 4.3392,
+      "step": 149
+    },
+    {
+      "epoch": 0.15408320493066255,
+      "grad_norm": 12.174428939819336,
+      "learning_rate": 2.2260273972602743e-06,
+      "loss": 4.2734,
+      "step": 150
+    },
+    {
+      "epoch": 0.15408320493066255,
+      "eval_loss": 4.29296875,
+      "eval_runtime": 8.1394,
+      "eval_samples_per_second": 122.859,
+      "eval_steps_per_second": 5.16,
+      "step": 150
+    },
+    {
+      "epoch": 0.15511042629686697,
+      "grad_norm": 11.123571395874023,
+      "learning_rate": 2.243150684931507e-06,
+      "loss": 4.151,
+      "step": 151
+    },
+    {
+      "epoch": 0.1561376476630714,
+      "grad_norm": 11.71768569946289,
+      "learning_rate": 2.26027397260274e-06,
+      "loss": 4.2695,
+      "step": 152
+    },
+    {
+      "epoch": 0.15716486902927582,
+      "grad_norm": 11.423957824707031,
+      "learning_rate": 2.2773972602739726e-06,
+      "loss": 4.2884,
+      "step": 153
+    },
+    {
+      "epoch": 0.15819209039548024,
+      "grad_norm": 11.415791511535645,
+      "learning_rate": 2.2945205479452057e-06,
+      "loss": 4.2487,
+      "step": 154
+    },
+    {
+      "epoch": 0.15921931176168463,
+      "grad_norm": 11.772719383239746,
+      "learning_rate": 2.3116438356164387e-06,
+      "loss": 4.2415,
+      "step": 155
+    },
+    {
+      "epoch": 0.16024653312788906,
+      "grad_norm": 11.59867000579834,
+      "learning_rate": 2.3287671232876713e-06,
+      "loss": 4.1628,
+      "step": 156
+    },
+    {
+      "epoch": 0.16127375449409348,
+      "grad_norm": 11.289015769958496,
+      "learning_rate": 2.3458904109589044e-06,
+      "loss": 4.3132,
+      "step": 157
+    },
+    {
+      "epoch": 0.1623009758602979,
+      "grad_norm": 11.444629669189453,
+      "learning_rate": 2.363013698630137e-06,
+      "loss": 4.3164,
+      "step": 158
+    },
+    {
+      "epoch": 0.1633281972265023,
+      "grad_norm": 11.693326950073242,
+      "learning_rate": 2.38013698630137e-06,
+      "loss": 4.2539,
+      "step": 159
+    },
+    {
+      "epoch": 0.16435541859270672,
+      "grad_norm": 11.368536949157715,
+      "learning_rate": 2.3972602739726027e-06,
+      "loss": 4.2708,
+      "step": 160
+    },
+    {
+      "epoch": 0.16435541859270672,
+      "eval_loss": 4.26171875,
+      "eval_runtime": 8.1405,
+      "eval_samples_per_second": 122.843,
+      "eval_steps_per_second": 5.159,
+      "step": 160
+    },
+    {
+      "epoch": 0.16538263995891114,
+      "grad_norm": 11.117278099060059,
+      "learning_rate": 2.4143835616438358e-06,
+      "loss": 4.2174,
+      "step": 161
+    },
+    {
+      "epoch": 0.16640986132511557,
+      "grad_norm": 10.721233367919922,
+      "learning_rate": 2.431506849315069e-06,
+      "loss": 4.2493,
+      "step": 162
+    },
+    {
+      "epoch": 0.16743708269132,
+      "grad_norm": 11.11534595489502,
+      "learning_rate": 2.4486301369863015e-06,
+      "loss": 4.2194,
+      "step": 163
+    },
+    {
+      "epoch": 0.16846430405752438,
+      "grad_norm": 10.975207328796387,
+      "learning_rate": 2.4657534246575345e-06,
+      "loss": 4.2109,
+      "step": 164
+    },
+    {
+      "epoch": 0.1694915254237288,
+      "grad_norm": 10.390185356140137,
+      "learning_rate": 2.482876712328767e-06,
+      "loss": 4.2051,
+      "step": 165
+    },
+    {
+      "epoch": 0.17051874678993323,
+      "grad_norm": 10.932448387145996,
+      "learning_rate": 2.5e-06,
+      "loss": 4.2376,
+      "step": 166
+    },
+    {
+      "epoch": 0.17154596815613765,
+      "grad_norm": 10.311753273010254,
+      "learning_rate": 2.5171232876712333e-06,
+      "loss": 4.1654,
+      "step": 167
+    },
+    {
+      "epoch": 0.17257318952234207,
+      "grad_norm": 10.680136680603027,
+      "learning_rate": 2.534246575342466e-06,
+      "loss": 4.222,
+      "step": 168
+    },
+    {
+      "epoch": 0.17360041088854647,
+      "grad_norm": 11.132901191711426,
+      "learning_rate": 2.551369863013699e-06,
+      "loss": 4.2832,
+      "step": 169
+    },
+    {
+      "epoch": 0.1746276322547509,
+      "grad_norm": 10.849762916564941,
+      "learning_rate": 2.568493150684932e-06,
+      "loss": 4.2734,
+      "step": 170
+    },
+    {
+      "epoch": 0.1746276322547509,
+      "eval_loss": 4.23046875,
+      "eval_runtime": 8.1466,
+      "eval_samples_per_second": 122.751,
+      "eval_steps_per_second": 5.156,
+      "step": 170
+    },
+    {
+      "epoch": 0.17565485362095531,
+      "grad_norm": 10.316183090209961,
+      "learning_rate": 2.585616438356164e-06,
+      "loss": 4.1986,
+      "step": 171
+    },
+    {
+      "epoch": 0.17668207498715974,
+      "grad_norm": 10.631659507751465,
+      "learning_rate": 2.6027397260273973e-06,
+      "loss": 4.166,
+      "step": 172
+    },
+    {
+      "epoch": 0.17770929635336416,
+      "grad_norm": 10.439342498779297,
+      "learning_rate": 2.6198630136986303e-06,
+      "loss": 4.2148,
+      "step": 173
+    },
+    {
+      "epoch": 0.17873651771956856,
+      "grad_norm": 10.113866806030273,
+      "learning_rate": 2.6369863013698634e-06,
+      "loss": 4.1445,
+      "step": 174
+    },
+    {
+      "epoch": 0.17976373908577298,
+      "grad_norm": 9.995186805725098,
+      "learning_rate": 2.654109589041096e-06,
+      "loss": 4.2269,
+      "step": 175
+    },
+    {
+      "epoch": 0.1807909604519774,
+      "grad_norm": 10.036297798156738,
+      "learning_rate": 2.671232876712329e-06,
+      "loss": 4.224,
+      "step": 176
+    },
+    {
+      "epoch": 0.18181818181818182,
+      "grad_norm": 10.111323356628418,
+      "learning_rate": 2.688356164383562e-06,
+      "loss": 4.1562,
+      "step": 177
+    },
+    {
+      "epoch": 0.18284540318438625,
+      "grad_norm": 10.693925857543945,
+      "learning_rate": 2.705479452054795e-06,
+      "loss": 4.2617,
+      "step": 178
+    },
+    {
+      "epoch": 0.18387262455059064,
+      "grad_norm": 10.322890281677246,
+      "learning_rate": 2.7226027397260274e-06,
+      "loss": 4.1016,
+      "step": 179
+    },
+    {
+      "epoch": 0.18489984591679506,
+      "grad_norm": 10.171163558959961,
+      "learning_rate": 2.7397260273972604e-06,
+      "loss": 4.2878,
+      "step": 180
+    },
+    {
+      "epoch": 0.18489984591679506,
+      "eval_loss": 4.203125,
+      "eval_runtime": 8.1188,
+      "eval_samples_per_second": 123.171,
+      "eval_steps_per_second": 5.173,
+      "step": 180
+    },
+    {
+      "epoch": 0.1859270672829995,
+      "grad_norm": 10.186366081237793,
+      "learning_rate": 2.7568493150684935e-06,
+      "loss": 4.2207,
+      "step": 181
+    },
+    {
+      "epoch": 0.1869542886492039,
+      "grad_norm": 9.998559951782227,
+      "learning_rate": 2.773972602739726e-06,
+      "loss": 4.224,
+      "step": 182
+    },
+    {
+      "epoch": 0.18798151001540833,
+      "grad_norm": 10.428750991821289,
+      "learning_rate": 2.791095890410959e-06,
+      "loss": 4.2266,
+      "step": 183
+    },
+    {
+      "epoch": 0.18900873138161273,
+      "grad_norm": 10.26636791229248,
+      "learning_rate": 2.8082191780821922e-06,
+      "loss": 4.1309,
+      "step": 184
+    },
+    {
+      "epoch": 0.19003595274781715,
+      "grad_norm": 9.86103630065918,
+      "learning_rate": 2.8253424657534253e-06,
+      "loss": 4.1169,
+      "step": 185
+    },
+    {
+      "epoch": 0.19106317411402157,
+      "grad_norm": 9.904764175415039,
+      "learning_rate": 2.842465753424658e-06,
+      "loss": 4.1419,
+      "step": 186
+    },
+    {
+      "epoch": 0.192090395480226,
+      "grad_norm": 9.750298500061035,
+      "learning_rate": 2.8595890410958905e-06,
+      "loss": 4.1335,
+      "step": 187
+    },
+    {
+      "epoch": 0.19311761684643042,
+      "grad_norm": 10.10220718383789,
+      "learning_rate": 2.876712328767123e-06,
+      "loss": 4.1452,
+      "step": 188
+    },
+    {
+      "epoch": 0.1941448382126348,
+      "grad_norm": 10.194372177124023,
+      "learning_rate": 2.8938356164383562e-06,
+      "loss": 4.1732,
+      "step": 189
+    },
+    {
+      "epoch": 0.19517205957883924,
+      "grad_norm": 9.45039176940918,
+      "learning_rate": 2.9109589041095893e-06,
+      "loss": 4.124,
+      "step": 190
+    },
+    {
+      "epoch": 0.19517205957883924,
+      "eval_loss": 4.1796875,
+      "eval_runtime": 8.1489,
+      "eval_samples_per_second": 122.715,
+      "eval_steps_per_second": 5.154,
+      "step": 190
+    },
+    {
+      "epoch": 0.19619928094504366,
+      "grad_norm": 9.065196990966797,
+      "learning_rate": 2.9280821917808223e-06,
+      "loss": 4.1689,
+      "step": 191
+    },
+    {
+      "epoch": 0.19722650231124808,
+      "grad_norm": 8.964832305908203,
+      "learning_rate": 2.945205479452055e-06,
+      "loss": 4.1621,
+      "step": 192
+    },
+    {
+      "epoch": 0.1982537236774525,
+      "grad_norm": 9.150090217590332,
+      "learning_rate": 2.962328767123288e-06,
+      "loss": 4.1302,
+      "step": 193
+    },
+    {
+      "epoch": 0.1992809450436569,
+      "grad_norm": 9.464834213256836,
+      "learning_rate": 2.979452054794521e-06,
+      "loss": 4.1276,
+      "step": 194
+    },
+    {
+      "epoch": 0.20030816640986132,
+      "grad_norm": 9.080924987792969,
+      "learning_rate": 2.9965753424657533e-06,
+      "loss": 4.0879,
+      "step": 195
+    },
+    {
+      "epoch": 0.20133538777606574,
+      "grad_norm": 8.936680793762207,
+      "learning_rate": 3.0136986301369864e-06,
+      "loss": 4.1009,
+      "step": 196
+    },
+    {
+      "epoch": 0.20236260914227017,
+      "grad_norm": 9.066742897033691,
+      "learning_rate": 3.0308219178082194e-06,
+      "loss": 4.1257,
+      "step": 197
+    },
+    {
+      "epoch": 0.2033898305084746,
+      "grad_norm": 9.437700271606445,
+      "learning_rate": 3.0479452054794525e-06,
+      "loss": 4.1055,
+      "step": 198
+    },
+    {
+      "epoch": 0.20441705187467898,
+      "grad_norm": 9.56484317779541,
+      "learning_rate": 3.065068493150685e-06,
+      "loss": 4.1836,
+      "step": 199
+    },
+    {
+      "epoch": 0.2054442732408834,
+      "grad_norm": 9.032593727111816,
+      "learning_rate": 3.082191780821918e-06,
+      "loss": 4.0485,
+      "step": 200
+    },
+    {
+      "epoch": 0.2054442732408834,
+      "eval_loss": 4.15625,
+      "eval_runtime": 8.1569,
+      "eval_samples_per_second": 122.595,
+      "eval_steps_per_second": 5.149,
+      "step": 200
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 2919,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 100,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1881302630400000.0,
+  "train_batch_size": 3,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/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)

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2000/added_tokens.json

@@ -0,0 +1,4 @@
+{
+  "[PAD]": 50257,
+  "🅁": 50258
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2000/config.json

@@ -0,0 +1,39 @@
+{
+  "_name_or_path": "gpt2",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "torch_dtype": "float16",
+  "transformers_version": "4.45.1",
+  "use_cache": true,
+  "vocab_size": 50257
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2000/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.45.1"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2000/latest

@@ -0,0 +1 @@
+global_step2000

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2000/special_tokens_map.json

@@ -0,0 +1,12 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2000/tokenizer_config.json

@@ -0,0 +1,36 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "50256": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50257": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50258": {
+      "content": "🅁",
+      "lstrip": true,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2000/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)

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2800/added_tokens.json

@@ -0,0 +1,4 @@
+{
+  "[PAD]": 50257,
+  "🅁": 50258
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2800/config.json

@@ -0,0 +1,39 @@
+{
+  "_name_or_path": "gpt2",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "torch_dtype": "float16",
+  "transformers_version": "4.45.1",
+  "use_cache": true,
+  "vocab_size": 50257
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2800/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.45.1"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2800/latest

@@ -0,0 +1 @@
+global_step2800

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2800/special_tokens_map.json

@@ -0,0 +1,12 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": "<|endoftext|>"
+}

checkpoints/GPT-2/babylm_reverse_control_10M_seed0/runs/checkpoint-2800/tokenizer_config.json

@@ -0,0 +1,36 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "50256": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50257": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50258": {
+      "content": "🅁",
+      "lstrip": true,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}