Upload checkpoint-1491 contents

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

.ipynb_checkpoints/zero_to_fp32-checkpoint.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

README.md

@@ -0,0 +1,202 @@
+---
+base_model: google/gemma-3-27b-it
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.14.0

adapter_config.json

@@ -0,0 +1,37 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-3-27b-it",
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": [
+    "embed_tokens",
+    "lm_head"
+  ],
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "k_proj",
+    "q_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93c0b35bb2f4d35f6fa6698722098abcee6d9402322bd3ffd2f6d2ffdf2c8a3a
+size 5711640384

chat_template.json

@@ -0,0 +1,3 @@
+{
+  "chat_template": "{{ bos_token }}\n{%- if messages[0]['role'] == 'system' -%}\n    {%- if messages[0]['content'] is string -%}\n        {%- set first_user_prefix = messages[0]['content'] + '\n\n' -%}\n    {%- else -%}\n        {%- set first_user_prefix = messages[0]['content'][0]['text'] + '\n\n' -%}\n    {%- endif -%}\n    {%- set loop_messages = messages[1:] -%}\n{%- else -%}\n    {%- set first_user_prefix = \"\" -%}\n    {%- set loop_messages = messages -%}\n{%- endif -%}\n{%- for message in loop_messages -%}\n    {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}\n        {{ raise_exception(\"Conversation roles must alternate user/assistant/user/assistant/...\") }}\n    {%- endif -%}\n    {%- if (message['role'] == 'assistant') -%}\n        {%- set role = \"model\" -%}\n    {%- else -%}\n        {%- set role = message['role'] -%}\n    {%- endif -%}\n    {{ '<start_of_turn>' + role + '\n' + (first_user_prefix if loop.first else \"\") }}\n    {%- if message['content'] is string -%}\n        {{ message['content'] | trim }}\n    {%- elif message['content'] is iterable -%}\n        {%- for item in message['content'] -%}\n            {%- if item['type'] == 'image' -%}\n                {{ '<start_of_image>' }}\n            {%- elif item['type'] == 'text' -%}\n                {{ item['text'] | trim }}\n            {%- endif -%}\n        {%- endfor -%}\n    {%- else -%}\n        {{ raise_exception(\"Invalid content type\") }}\n    {%- endif -%}\n    {{ '<end_of_turn>\n' }}\n{%- endfor -%}\n{%- if add_generation_prompt -%}\n    {{'<start_of_turn>model\n'}}\n{%- endif -%}\n"
+}

global_step1491/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9088be42fdd195304bd2479329e0be460086e1781f3f785fcd8405175ff3c2fc
+size 4283677776

global_step1491/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:786cd946e7fa2a921b4fe31bb44bbb0b64543eacd00c9894a0c909faf33e1df4
+size 4283655696

global_step1491/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a7c969b8e5e37fb200a4840ab0089b4e5f1ab18a2cdf21da64d6d309e94f65d
+size 4283655696

global_step1491/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e8514fb7770ebca226733faa38c8ca3e349e796eb471f68ecf4a09a9cd116f7
+size 4283684048

global_step1491/bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46e4430135e20d55ea949a6242f48bfa19b371388c06b30d632f2be28358f621
+size 4283689808

global_step1491/bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:534dc9092aa96349817ea4dce8cf5dd800005632fabf268e68e8b7ce552f9185
+size 4283655696

global_step1491/bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85eca2ec0dbc051955695599819daa2e010eb6dc639af7a597759ada8656d673
+size 4283655696

global_step1491/bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:361d25aef70a8089a1f1cf5afa96ac938e6293701d699c17c4ffb4c238f44c0d
+size 4283655696

global_step1491/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8556e68d38c1f01727a1243cdd9e8d56a36c3a032a04c0756d031d745b16039b
+size 8531477944

latest

@@ -0,0 +1 @@
+global_step1491

preprocessor_config.json

@@ -0,0 +1,36 @@
+{
+  "crop_size": null,
+  "data_format": "channels_first",
+  "default_to_square": true,
+  "device": null,
+  "do_center_crop": null,
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_pan_and_scan": null,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "Gemma3ImageProcessorFast",
+  "image_seq_length": 256,
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "input_data_format": null,
+  "pan_and_scan_max_num_crops": null,
+  "pan_and_scan_min_crop_size": null,
+  "pan_and_scan_min_ratio_to_activate": null,
+  "processor_class": "Gemma3Processor",
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "return_tensors": null,
+  "size": {
+    "height": 896,
+    "width": 896
+  }
+}

processor_config.json

@@ -0,0 +1,4 @@
+{
+  "image_seq_length": 256,
+  "processor_class": "Gemma3Processor"
+}

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efca4649b288b8ad5a7993429fbf1df9c8d390d3e6a2fd99c9abac426716a574
+size 15984

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f4319170fbce0a9f543035ef13c7e23bd26040d6391ed5c3a57f13d1b4396b8
+size 15984

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:327036d7728641b60db20720aa0f304e65c9e7c31b217dc9a0749c97f1ffbf21
+size 15984

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e79f1e1aec1dff62dac8c2ee0d2498a5360a8b6e7625b9af9c18fbf44c7021c6
+size 15984

rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:941cf38d14342643415ee2d35c13039c0c9eec406c8c759c7377366fc2692851
+size 15984

rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b52955539d014f9249f1d5ca976e749c952d5d82b447f3fd08ab16c6cf0b267a
+size 15984

rng_state_6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d28b3d402ce58c9d30f488e68bffada0c398774ba8971cf64a7ead8878a828d8
+size 15984

rng_state_7.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2b3b7df2243bb8b9173d3d9a966115ca954c1686c9213a4c94f221cb77eb668
+size 15984

scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5cda86e1665017464cb4091acb5734e3208741b558b0a1874603320d04971f1
+size 1064

special_tokens_map.json

@@ -0,0 +1,33 @@
+{
+  "boi_token": "<start_of_image>",
+  "bos_token": {
+    "content": "<bos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eoi_token": "<end_of_image>",
+  "eos_token": {
+    "content": "<eos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "image_token": "<image_soft_token>",
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4667f2089529e8e7657cfb6d1c19910ae71ff5f28aa7ab2ff2763330affad795
+size 33384568

trainer_state.json

@@ -0,0 +1,2427 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.9994972347913524,
+  "eval_steps": 750,
+  "global_step": 1491,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0033517680576504107,
+      "grad_norm": 14.694869995117188,
+      "learning_rate": 6.666666666666667e-06,
+      "loss": 53.6406,
+      "mean_token_accuracy": 0.5338318642228842,
+      "step": 5
+    },
+    {
+      "epoch": 0.006703536115300821,
+      "grad_norm": 14.033230781555176,
+      "learning_rate": 1.3333333333333333e-05,
+      "loss": 52.3838,
+      "mean_token_accuracy": 0.5248840853571892,
+      "step": 10
+    },
+    {
+      "epoch": 0.010055304172951232,
+      "grad_norm": 6.804769039154053,
+      "learning_rate": 2e-05,
+      "loss": 47.9105,
+      "mean_token_accuracy": 0.5399681400507689,
+      "step": 15
+    },
+    {
+      "epoch": 0.013407072230601643,
+      "grad_norm": 7.750083923339844,
+      "learning_rate": 2.6666666666666667e-05,
+      "loss": 41.8861,
+      "mean_token_accuracy": 0.55653104968369,
+      "step": 20
+    },
+    {
+      "epoch": 0.01675884028825205,
+      "grad_norm": 6.184543132781982,
+      "learning_rate": 3.3333333333333335e-05,
+      "loss": 37.33,
+      "mean_token_accuracy": 0.5655230440199375,
+      "step": 25
+    },
+    {
+      "epoch": 0.020110608345902465,
+      "grad_norm": 4.537179946899414,
+      "learning_rate": 4e-05,
+      "loss": 32.7503,
+      "mean_token_accuracy": 0.587661711126566,
+      "step": 30
+    },
+    {
+      "epoch": 0.023462376403552875,
+      "grad_norm": 3.6645753383636475,
+      "learning_rate": 4.666666666666667e-05,
+      "loss": 29.1892,
+      "mean_token_accuracy": 0.6075583577156067,
+      "step": 35
+    },
+    {
+      "epoch": 0.026814144461203285,
+      "grad_norm": 3.7526533603668213,
+      "learning_rate": 5.333333333333333e-05,
+      "loss": 26.3524,
+      "mean_token_accuracy": 0.6198613092303276,
+      "step": 40
+    },
+    {
+      "epoch": 0.030165912518853696,
+      "grad_norm": 3.0561397075653076,
+      "learning_rate": 6e-05,
+      "loss": 24.1513,
+      "mean_token_accuracy": 0.6353930421173573,
+      "step": 45
+    },
+    {
+      "epoch": 0.0335176805765041,
+      "grad_norm": 2.857618808746338,
+      "learning_rate": 6.666666666666667e-05,
+      "loss": 23.5029,
+      "mean_token_accuracy": 0.6437373287975788,
+      "step": 50
+    },
+    {
+      "epoch": 0.03686944863415452,
+      "grad_norm": 2.7901978492736816,
+      "learning_rate": 7.333333333333333e-05,
+      "loss": 22.9387,
+      "mean_token_accuracy": 0.646886795759201,
+      "step": 55
+    },
+    {
+      "epoch": 0.04022121669180493,
+      "grad_norm": 2.8266501426696777,
+      "learning_rate": 8e-05,
+      "loss": 22.0359,
+      "mean_token_accuracy": 0.6525138475000858,
+      "step": 60
+    },
+    {
+      "epoch": 0.04357298474945534,
+      "grad_norm": 2.5010733604431152,
+      "learning_rate": 8.666666666666667e-05,
+      "loss": 21.5158,
+      "mean_token_accuracy": 0.6548139773309231,
+      "step": 65
+    },
+    {
+      "epoch": 0.04692475280710575,
+      "grad_norm": 2.5834386348724365,
+      "learning_rate": 9.333333333333334e-05,
+      "loss": 21.5409,
+      "mean_token_accuracy": 0.6478891499340534,
+      "step": 70
+    },
+    {
+      "epoch": 0.05027652086475616,
+      "grad_norm": 2.6927576065063477,
+      "learning_rate": 0.0001,
+      "loss": 20.1017,
+      "mean_token_accuracy": 0.6757474772632122,
+      "step": 75
+    },
+    {
+      "epoch": 0.05362828892240657,
+      "grad_norm": 2.0276572704315186,
+      "learning_rate": 9.964689265536724e-05,
+      "loss": 19.9912,
+      "mean_token_accuracy": 0.6763999305665493,
+      "step": 80
+    },
+    {
+      "epoch": 0.05698005698005698,
+      "grad_norm": 2.4628567695617676,
+      "learning_rate": 9.929378531073446e-05,
+      "loss": 19.9089,
+      "mean_token_accuracy": 0.672279854118824,
+      "step": 85
+    },
+    {
+      "epoch": 0.06033182503770739,
+      "grad_norm": 2.258838415145874,
+      "learning_rate": 9.89406779661017e-05,
+      "loss": 19.7132,
+      "mean_token_accuracy": 0.6713059276342392,
+      "step": 90
+    },
+    {
+      "epoch": 0.0636835930953578,
+      "grad_norm": 2.447565793991089,
+      "learning_rate": 9.858757062146892e-05,
+      "loss": 18.7631,
+      "mean_token_accuracy": 0.6825208596885204,
+      "step": 95
+    },
+    {
+      "epoch": 0.0670353611530082,
+      "grad_norm": 2.1105902194976807,
+      "learning_rate": 9.823446327683616e-05,
+      "loss": 19.4631,
+      "mean_token_accuracy": 0.6674435302615166,
+      "step": 100
+    },
+    {
+      "epoch": 0.07038712921065862,
+      "grad_norm": 2.309248447418213,
+      "learning_rate": 9.78813559322034e-05,
+      "loss": 19.0249,
+      "mean_token_accuracy": 0.6734571024775505,
+      "step": 105
+    },
+    {
+      "epoch": 0.07373889726830904,
+      "grad_norm": 2.101681709289551,
+      "learning_rate": 9.752824858757063e-05,
+      "loss": 18.593,
+      "mean_token_accuracy": 0.6875097192823887,
+      "step": 110
+    },
+    {
+      "epoch": 0.07709066532595944,
+      "grad_norm": 2.157726526260376,
+      "learning_rate": 9.717514124293787e-05,
+      "loss": 18.5973,
+      "mean_token_accuracy": 0.6829216606914997,
+      "step": 115
+    },
+    {
+      "epoch": 0.08044243338360986,
+      "grad_norm": 2.0711209774017334,
+      "learning_rate": 9.682203389830509e-05,
+      "loss": 19.1541,
+      "mean_token_accuracy": 0.6785640828311443,
+      "step": 120
+    },
+    {
+      "epoch": 0.08379420144126026,
+      "grad_norm": 2.015594959259033,
+      "learning_rate": 9.646892655367233e-05,
+      "loss": 18.9493,
+      "mean_token_accuracy": 0.6861244946718216,
+      "step": 125
+    },
+    {
+      "epoch": 0.08714596949891068,
+      "grad_norm": 2.1295998096466064,
+      "learning_rate": 9.611581920903955e-05,
+      "loss": 18.5125,
+      "mean_token_accuracy": 0.6793887488543987,
+      "step": 130
+    },
+    {
+      "epoch": 0.09049773755656108,
+      "grad_norm": 2.2496395111083984,
+      "learning_rate": 9.576271186440679e-05,
+      "loss": 18.4019,
+      "mean_token_accuracy": 0.6890006221830844,
+      "step": 135
+    },
+    {
+      "epoch": 0.0938495056142115,
+      "grad_norm": 2.1168577671051025,
+      "learning_rate": 9.540960451977402e-05,
+      "loss": 18.7305,
+      "mean_token_accuracy": 0.6841622419655323,
+      "step": 140
+    },
+    {
+      "epoch": 0.0972012736718619,
+      "grad_norm": 1.8554915189743042,
+      "learning_rate": 9.505649717514125e-05,
+      "loss": 18.6606,
+      "mean_token_accuracy": 0.6859239712357521,
+      "step": 145
+    },
+    {
+      "epoch": 0.10055304172951232,
+      "grad_norm": 1.9698066711425781,
+      "learning_rate": 9.470338983050848e-05,
+      "loss": 19.1065,
+      "mean_token_accuracy": 0.6759489566087723,
+      "step": 150
+    },
+    {
+      "epoch": 0.10390480978716272,
+      "grad_norm": 2.2483623027801514,
+      "learning_rate": 9.43502824858757e-05,
+      "loss": 18.8041,
+      "mean_token_accuracy": 0.68142851293087,
+      "step": 155
+    },
+    {
+      "epoch": 0.10725657784481314,
+      "grad_norm": 1.8570690155029297,
+      "learning_rate": 9.399717514124294e-05,
+      "loss": 18.8862,
+      "mean_token_accuracy": 0.6791303649544715,
+      "step": 160
+    },
+    {
+      "epoch": 0.11060834590246355,
+      "grad_norm": 2.143021583557129,
+      "learning_rate": 9.364406779661016e-05,
+      "loss": 18.7605,
+      "mean_token_accuracy": 0.681893227249384,
+      "step": 165
+    },
+    {
+      "epoch": 0.11396011396011396,
+      "grad_norm": 1.8951307535171509,
+      "learning_rate": 9.32909604519774e-05,
+      "loss": 18.3005,
+      "mean_token_accuracy": 0.6897541806101799,
+      "step": 170
+    },
+    {
+      "epoch": 0.11731188201776437,
+      "grad_norm": 1.971745252609253,
+      "learning_rate": 9.293785310734464e-05,
+      "loss": 18.8995,
+      "mean_token_accuracy": 0.6820204116404056,
+      "step": 175
+    },
+    {
+      "epoch": 0.12066365007541478,
+      "grad_norm": 1.910328984260559,
+      "learning_rate": 9.258474576271187e-05,
+      "loss": 18.8808,
+      "mean_token_accuracy": 0.6812884464859963,
+      "step": 180
+    },
+    {
+      "epoch": 0.12401541813306519,
+      "grad_norm": 1.730974555015564,
+      "learning_rate": 9.223163841807911e-05,
+      "loss": 18.0871,
+      "mean_token_accuracy": 0.6907590143382549,
+      "step": 185
+    },
+    {
+      "epoch": 0.1273671861907156,
+      "grad_norm": 2.125452995300293,
+      "learning_rate": 9.187853107344633e-05,
+      "loss": 18.1569,
+      "mean_token_accuracy": 0.689236406236887,
+      "step": 190
+    },
+    {
+      "epoch": 0.13071895424836602,
+      "grad_norm": 2.0234949588775635,
+      "learning_rate": 9.152542372881357e-05,
+      "loss": 18.3342,
+      "mean_token_accuracy": 0.6902932204306126,
+      "step": 195
+    },
+    {
+      "epoch": 0.1340707223060164,
+      "grad_norm": 1.9802364110946655,
+      "learning_rate": 9.11723163841808e-05,
+      "loss": 18.7942,
+      "mean_token_accuracy": 0.6788501650094986,
+      "step": 200
+    },
+    {
+      "epoch": 0.13742249036366683,
+      "grad_norm": 1.8897534608840942,
+      "learning_rate": 9.081920903954803e-05,
+      "loss": 18.4679,
+      "mean_token_accuracy": 0.6900524459779263,
+      "step": 205
+    },
+    {
+      "epoch": 0.14077425842131724,
+      "grad_norm": 1.9040635824203491,
+      "learning_rate": 9.046610169491526e-05,
+      "loss": 18.0058,
+      "mean_token_accuracy": 0.690093420445919,
+      "step": 210
+    },
+    {
+      "epoch": 0.14412602647896766,
+      "grad_norm": 2.0558955669403076,
+      "learning_rate": 9.011299435028249e-05,
+      "loss": 17.5489,
+      "mean_token_accuracy": 0.7006829999387264,
+      "step": 215
+    },
+    {
+      "epoch": 0.14747779453661808,
+      "grad_norm": 1.7952055931091309,
+      "learning_rate": 8.975988700564972e-05,
+      "loss": 18.2907,
+      "mean_token_accuracy": 0.6876891441643238,
+      "step": 220
+    },
+    {
+      "epoch": 0.15082956259426847,
+      "grad_norm": 1.8588192462921143,
+      "learning_rate": 8.940677966101694e-05,
+      "loss": 18.4005,
+      "mean_token_accuracy": 0.6897859051823616,
+      "step": 225
+    },
+    {
+      "epoch": 0.15418133065191889,
+      "grad_norm": 1.9269477128982544,
+      "learning_rate": 8.905367231638418e-05,
+      "loss": 18.2096,
+      "mean_token_accuracy": 0.6909494370222091,
+      "step": 230
+    },
+    {
+      "epoch": 0.1575330987095693,
+      "grad_norm": 1.8693301677703857,
+      "learning_rate": 8.870056497175142e-05,
+      "loss": 18.394,
+      "mean_token_accuracy": 0.6836515329778194,
+      "step": 235
+    },
+    {
+      "epoch": 0.16088486676721972,
+      "grad_norm": 1.787061333656311,
+      "learning_rate": 8.834745762711864e-05,
+      "loss": 18.1503,
+      "mean_token_accuracy": 0.6907145738601684,
+      "step": 240
+    },
+    {
+      "epoch": 0.1642366348248701,
+      "grad_norm": 1.8895225524902344,
+      "learning_rate": 8.799435028248588e-05,
+      "loss": 18.3026,
+      "mean_token_accuracy": 0.6878940775990486,
+      "step": 245
+    },
+    {
+      "epoch": 0.16758840288252053,
+      "grad_norm": 1.835693120956421,
+      "learning_rate": 8.764124293785311e-05,
+      "loss": 17.9347,
+      "mean_token_accuracy": 0.6917316012084485,
+      "step": 250
+    },
+    {
+      "epoch": 0.17094017094017094,
+      "grad_norm": 1.7408661842346191,
+      "learning_rate": 8.728813559322035e-05,
+      "loss": 18.0051,
+      "mean_token_accuracy": 0.689583633840084,
+      "step": 255
+    },
+    {
+      "epoch": 0.17429193899782136,
+      "grad_norm": 1.9096996784210205,
+      "learning_rate": 8.693502824858759e-05,
+      "loss": 17.6064,
+      "mean_token_accuracy": 0.6965925216674804,
+      "step": 260
+    },
+    {
+      "epoch": 0.17764370705547175,
+      "grad_norm": 1.9822146892547607,
+      "learning_rate": 8.658192090395481e-05,
+      "loss": 17.6301,
+      "mean_token_accuracy": 0.7005406267940998,
+      "step": 265
+    },
+    {
+      "epoch": 0.18099547511312217,
+      "grad_norm": 1.8383901119232178,
+      "learning_rate": 8.622881355932204e-05,
+      "loss": 17.9114,
+      "mean_token_accuracy": 0.6876685306429863,
+      "step": 270
+    },
+    {
+      "epoch": 0.18434724317077258,
+      "grad_norm": 1.7920355796813965,
+      "learning_rate": 8.587570621468927e-05,
+      "loss": 18.1271,
+      "mean_token_accuracy": 0.689356567710638,
+      "step": 275
+    },
+    {
+      "epoch": 0.187699011228423,
+      "grad_norm": 1.6455663442611694,
+      "learning_rate": 8.55225988700565e-05,
+      "loss": 17.787,
+      "mean_token_accuracy": 0.6919776491820813,
+      "step": 280
+    },
+    {
+      "epoch": 0.1910507792860734,
+      "grad_norm": 1.9442647695541382,
+      "learning_rate": 8.516949152542373e-05,
+      "loss": 17.6019,
+      "mean_token_accuracy": 0.6980393722653389,
+      "step": 285
+    },
+    {
+      "epoch": 0.1944025473437238,
+      "grad_norm": 2.294377565383911,
+      "learning_rate": 8.481638418079096e-05,
+      "loss": 17.8778,
+      "mean_token_accuracy": 0.6954585202038288,
+      "step": 290
+    },
+    {
+      "epoch": 0.19775431540137423,
+      "grad_norm": 1.8009259700775146,
+      "learning_rate": 8.44632768361582e-05,
+      "loss": 17.5257,
+      "mean_token_accuracy": 0.6998075112700463,
+      "step": 295
+    },
+    {
+      "epoch": 0.20110608345902464,
+      "grad_norm": 2.015516757965088,
+      "learning_rate": 8.411016949152542e-05,
+      "loss": 17.7554,
+      "mean_token_accuracy": 0.6968327619135379,
+      "step": 300
+    },
+    {
+      "epoch": 0.20445785151667506,
+      "grad_norm": 1.5640082359313965,
+      "learning_rate": 8.375706214689266e-05,
+      "loss": 17.3438,
+      "mean_token_accuracy": 0.69996168166399,
+      "step": 305
+    },
+    {
+      "epoch": 0.20780961957432545,
+      "grad_norm": 1.9527899026870728,
+      "learning_rate": 8.340395480225988e-05,
+      "loss": 17.6883,
+      "mean_token_accuracy": 0.6988407798111439,
+      "step": 310
+    },
+    {
+      "epoch": 0.21116138763197587,
+      "grad_norm": 1.8222606182098389,
+      "learning_rate": 8.305084745762712e-05,
+      "loss": 17.0646,
+      "mean_token_accuracy": 0.7061679445207119,
+      "step": 315
+    },
+    {
+      "epoch": 0.21451315568962628,
+      "grad_norm": 1.8560868501663208,
+      "learning_rate": 8.269774011299435e-05,
+      "loss": 17.8875,
+      "mean_token_accuracy": 0.6941629223525524,
+      "step": 320
+    },
+    {
+      "epoch": 0.2178649237472767,
+      "grad_norm": 1.7588037252426147,
+      "learning_rate": 8.234463276836159e-05,
+      "loss": 17.6412,
+      "mean_token_accuracy": 0.6954927705228329,
+      "step": 325
+    },
+    {
+      "epoch": 0.2212166918049271,
+      "grad_norm": 1.738242268562317,
+      "learning_rate": 8.199152542372883e-05,
+      "loss": 17.8251,
+      "mean_token_accuracy": 0.6898994512856007,
+      "step": 330
+    },
+    {
+      "epoch": 0.2245684598625775,
+      "grad_norm": 1.8485089540481567,
+      "learning_rate": 8.163841807909605e-05,
+      "loss": 17.3078,
+      "mean_token_accuracy": 0.7000270999968052,
+      "step": 335
+    },
+    {
+      "epoch": 0.22792022792022792,
+      "grad_norm": 1.8579105138778687,
+      "learning_rate": 8.128531073446328e-05,
+      "loss": 17.3078,
+      "mean_token_accuracy": 0.6995702408254146,
+      "step": 340
+    },
+    {
+      "epoch": 0.23127199597787834,
+      "grad_norm": 1.7994352579116821,
+      "learning_rate": 8.093220338983051e-05,
+      "loss": 17.7557,
+      "mean_token_accuracy": 0.6928035505115986,
+      "step": 345
+    },
+    {
+      "epoch": 0.23462376403552873,
+      "grad_norm": 1.9240634441375732,
+      "learning_rate": 8.057909604519774e-05,
+      "loss": 17.4329,
+      "mean_token_accuracy": 0.6960855178534985,
+      "step": 350
+    },
+    {
+      "epoch": 0.23797553209317915,
+      "grad_norm": 1.6718952655792236,
+      "learning_rate": 8.022598870056498e-05,
+      "loss": 17.5951,
+      "mean_token_accuracy": 0.6947735913097859,
+      "step": 355
+    },
+    {
+      "epoch": 0.24132730015082957,
+      "grad_norm": 1.6835826635360718,
+      "learning_rate": 7.98728813559322e-05,
+      "loss": 18.1085,
+      "mean_token_accuracy": 0.6882089108228684,
+      "step": 360
+    },
+    {
+      "epoch": 0.24467906820847998,
+      "grad_norm": 1.7387073040008545,
+      "learning_rate": 7.951977401129944e-05,
+      "loss": 17.799,
+      "mean_token_accuracy": 0.6932998545467853,
+      "step": 365
+    },
+    {
+      "epoch": 0.24803083626613037,
+      "grad_norm": 2.0071725845336914,
+      "learning_rate": 7.916666666666666e-05,
+      "loss": 17.4076,
+      "mean_token_accuracy": 0.6961173862218857,
+      "step": 370
+    },
+    {
+      "epoch": 0.2513826043237808,
+      "grad_norm": 2.326915740966797,
+      "learning_rate": 7.88135593220339e-05,
+      "loss": 17.3121,
+      "mean_token_accuracy": 0.7005321949720382,
+      "step": 375
+    },
+    {
+      "epoch": 0.2547343723814312,
+      "grad_norm": 2.1876060962677,
+      "learning_rate": 7.846045197740113e-05,
+      "loss": 17.9069,
+      "mean_token_accuracy": 0.6906426399946213,
+      "step": 380
+    },
+    {
+      "epoch": 0.2580861404390816,
+      "grad_norm": 1.849671483039856,
+      "learning_rate": 7.810734463276837e-05,
+      "loss": 17.483,
+      "mean_token_accuracy": 0.7000573620200157,
+      "step": 385
+    },
+    {
+      "epoch": 0.26143790849673204,
+      "grad_norm": 1.6676862239837646,
+      "learning_rate": 7.775423728813561e-05,
+      "loss": 16.8936,
+      "mean_token_accuracy": 0.7045633904635906,
+      "step": 390
+    },
+    {
+      "epoch": 0.26478967655438246,
+      "grad_norm": 1.6702505350112915,
+      "learning_rate": 7.740112994350283e-05,
+      "loss": 17.904,
+      "mean_token_accuracy": 0.6874841086566448,
+      "step": 395
+    },
+    {
+      "epoch": 0.2681414446120328,
+      "grad_norm": 1.7280704975128174,
+      "learning_rate": 7.704802259887007e-05,
+      "loss": 17.4515,
+      "mean_token_accuracy": 0.7018027983605861,
+      "step": 400
+    },
+    {
+      "epoch": 0.27149321266968324,
+      "grad_norm": 1.8801991939544678,
+      "learning_rate": 7.669491525423729e-05,
+      "loss": 17.43,
+      "mean_token_accuracy": 0.7009049601852894,
+      "step": 405
+    },
+    {
+      "epoch": 0.27484498072733365,
+      "grad_norm": 1.9758073091506958,
+      "learning_rate": 7.634180790960453e-05,
+      "loss": 17.5984,
+      "mean_token_accuracy": 0.6948069363832474,
+      "step": 410
+    },
+    {
+      "epoch": 0.27819674878498407,
+      "grad_norm": 1.5747147798538208,
+      "learning_rate": 7.598870056497176e-05,
+      "loss": 18.3079,
+      "mean_token_accuracy": 0.6853139907121658,
+      "step": 415
+    },
+    {
+      "epoch": 0.2815485168426345,
+      "grad_norm": 1.6292234659194946,
+      "learning_rate": 7.563559322033898e-05,
+      "loss": 17.4527,
+      "mean_token_accuracy": 0.697540608048439,
+      "step": 420
+    },
+    {
+      "epoch": 0.2849002849002849,
+      "grad_norm": 1.6185086965560913,
+      "learning_rate": 7.528248587570622e-05,
+      "loss": 17.4193,
+      "mean_token_accuracy": 0.7012022204697133,
+      "step": 425
+    },
+    {
+      "epoch": 0.2882520529579353,
+      "grad_norm": 1.8361762762069702,
+      "learning_rate": 7.492937853107344e-05,
+      "loss": 17.4544,
+      "mean_token_accuracy": 0.698820473998785,
+      "step": 430
+    },
+    {
+      "epoch": 0.29160382101558574,
+      "grad_norm": 1.7740592956542969,
+      "learning_rate": 7.457627118644068e-05,
+      "loss": 18.0507,
+      "mean_token_accuracy": 0.6881603226065636,
+      "step": 435
+    },
+    {
+      "epoch": 0.29495558907323616,
+      "grad_norm": 1.8252911567687988,
+      "learning_rate": 7.42231638418079e-05,
+      "loss": 17.155,
+      "mean_token_accuracy": 0.7065504610538482,
+      "step": 440
+    },
+    {
+      "epoch": 0.2983073571308865,
+      "grad_norm": 1.8424382209777832,
+      "learning_rate": 7.387005649717514e-05,
+      "loss": 17.3055,
+      "mean_token_accuracy": 0.6978819817304611,
+      "step": 445
+    },
+    {
+      "epoch": 0.30165912518853694,
+      "grad_norm": 1.7494243383407593,
+      "learning_rate": 7.351694915254238e-05,
+      "loss": 16.8365,
+      "mean_token_accuracy": 0.7099504336714745,
+      "step": 450
+    },
+    {
+      "epoch": 0.30501089324618735,
+      "grad_norm": 1.936540961265564,
+      "learning_rate": 7.316384180790961e-05,
+      "loss": 18.2753,
+      "mean_token_accuracy": 0.6913827233016491,
+      "step": 455
+    },
+    {
+      "epoch": 0.30836266130383777,
+      "grad_norm": 1.810272216796875,
+      "learning_rate": 7.281073446327685e-05,
+      "loss": 17.0536,
+      "mean_token_accuracy": 0.6986232809722424,
+      "step": 460
+    },
+    {
+      "epoch": 0.3117144293614882,
+      "grad_norm": 1.6832094192504883,
+      "learning_rate": 7.245762711864407e-05,
+      "loss": 17.2231,
+      "mean_token_accuracy": 0.702030860632658,
+      "step": 465
+    },
+    {
+      "epoch": 0.3150661974191386,
+      "grad_norm": 1.8872151374816895,
+      "learning_rate": 7.21045197740113e-05,
+      "loss": 17.5502,
+      "mean_token_accuracy": 0.6932449921965599,
+      "step": 470
+    },
+    {
+      "epoch": 0.318417965476789,
+      "grad_norm": 1.788021445274353,
+      "learning_rate": 7.175141242937854e-05,
+      "loss": 16.8596,
+      "mean_token_accuracy": 0.7096694305539131,
+      "step": 475
+    },
+    {
+      "epoch": 0.32176973353443944,
+      "grad_norm": 1.8025559186935425,
+      "learning_rate": 7.139830508474577e-05,
+      "loss": 16.662,
+      "mean_token_accuracy": 0.7063573338091373,
+      "step": 480
+    },
+    {
+      "epoch": 0.3251215015920898,
+      "grad_norm": 2.274674654006958,
+      "learning_rate": 7.1045197740113e-05,
+      "loss": 17.5965,
+      "mean_token_accuracy": 0.6934389650821686,
+      "step": 485
+    },
+    {
+      "epoch": 0.3284732696497402,
+      "grad_norm": 1.6426053047180176,
+      "learning_rate": 7.069209039548022e-05,
+      "loss": 17.0914,
+      "mean_token_accuracy": 0.7049042917788029,
+      "step": 490
+    },
+    {
+      "epoch": 0.33182503770739064,
+      "grad_norm": 1.6252586841583252,
+      "learning_rate": 7.033898305084746e-05,
+      "loss": 17.6078,
+      "mean_token_accuracy": 0.6924709647893905,
+      "step": 495
+    },
+    {
+      "epoch": 0.33517680576504105,
+      "grad_norm": 1.7185930013656616,
+      "learning_rate": 6.998587570621468e-05,
+      "loss": 17.314,
+      "mean_token_accuracy": 0.7039985358715057,
+      "step": 500
+    },
+    {
+      "epoch": 0.33852857382269147,
+      "grad_norm": 1.7891852855682373,
+      "learning_rate": 6.963276836158192e-05,
+      "loss": 17.2188,
+      "mean_token_accuracy": 0.6977060906589031,
+      "step": 505
+    },
+    {
+      "epoch": 0.3418803418803419,
+      "grad_norm": 1.9103929996490479,
+      "learning_rate": 6.927966101694916e-05,
+      "loss": 17.4467,
+      "mean_token_accuracy": 0.6982413403689861,
+      "step": 510
+    },
+    {
+      "epoch": 0.3452321099379923,
+      "grad_norm": 1.8996375799179077,
+      "learning_rate": 6.892655367231638e-05,
+      "loss": 16.9608,
+      "mean_token_accuracy": 0.7054095402359962,
+      "step": 515
+    },
+    {
+      "epoch": 0.3485838779956427,
+      "grad_norm": 2.0335419178009033,
+      "learning_rate": 6.857344632768362e-05,
+      "loss": 17.3361,
+      "mean_token_accuracy": 0.7016568422317505,
+      "step": 520
+    },
+    {
+      "epoch": 0.35193564605329314,
+      "grad_norm": 1.9008755683898926,
+      "learning_rate": 6.822033898305085e-05,
+      "loss": 16.9694,
+      "mean_token_accuracy": 0.7059390284121037,
+      "step": 525
+    },
+    {
+      "epoch": 0.3552874141109435,
+      "grad_norm": 1.8340988159179688,
+      "learning_rate": 6.786723163841809e-05,
+      "loss": 17.3528,
+      "mean_token_accuracy": 0.7033507622778415,
+      "step": 530
+    },
+    {
+      "epoch": 0.3586391821685939,
+      "grad_norm": 1.6903594732284546,
+      "learning_rate": 6.751412429378532e-05,
+      "loss": 17.3021,
+      "mean_token_accuracy": 0.7001501135528088,
+      "step": 535
+    },
+    {
+      "epoch": 0.36199095022624433,
+      "grad_norm": 1.8101950883865356,
+      "learning_rate": 6.716101694915255e-05,
+      "loss": 17.938,
+      "mean_token_accuracy": 0.6908830553293228,
+      "step": 540
+    },
+    {
+      "epoch": 0.36534271828389475,
+      "grad_norm": 1.6470075845718384,
+      "learning_rate": 6.680790960451978e-05,
+      "loss": 17.6612,
+      "mean_token_accuracy": 0.6923478744924069,
+      "step": 545
+    },
+    {
+      "epoch": 0.36869448634154517,
+      "grad_norm": 2.1860337257385254,
+      "learning_rate": 6.6454802259887e-05,
+      "loss": 17.5684,
+      "mean_token_accuracy": 0.6983748801052571,
+      "step": 550
+    },
+    {
+      "epoch": 0.3720462543991956,
+      "grad_norm": 1.717653512954712,
+      "learning_rate": 6.610169491525424e-05,
+      "loss": 17.1166,
+      "mean_token_accuracy": 0.7025655619800091,
+      "step": 555
+    },
+    {
+      "epoch": 0.375398022456846,
+      "grad_norm": 1.9525723457336426,
+      "learning_rate": 6.574858757062147e-05,
+      "loss": 17.2908,
+      "mean_token_accuracy": 0.6997996769845486,
+      "step": 560
+    },
+    {
+      "epoch": 0.3787497905144964,
+      "grad_norm": 1.6053602695465088,
+      "learning_rate": 6.53954802259887e-05,
+      "loss": 17.3894,
+      "mean_token_accuracy": 0.698741364479065,
+      "step": 565
+    },
+    {
+      "epoch": 0.3821015585721468,
+      "grad_norm": 1.7356934547424316,
+      "learning_rate": 6.504237288135594e-05,
+      "loss": 17.1546,
+      "mean_token_accuracy": 0.7013543620705605,
+      "step": 570
+    },
+    {
+      "epoch": 0.3854533266297972,
+      "grad_norm": 1.7188559770584106,
+      "learning_rate": 6.468926553672316e-05,
+      "loss": 17.7637,
+      "mean_token_accuracy": 0.6936320647597313,
+      "step": 575
+    },
+    {
+      "epoch": 0.3888050946874476,
+      "grad_norm": 1.8413478136062622,
+      "learning_rate": 6.43361581920904e-05,
+      "loss": 17.8498,
+      "mean_token_accuracy": 0.695782047510147,
+      "step": 580
+    },
+    {
+      "epoch": 0.39215686274509803,
+      "grad_norm": 1.5715190172195435,
+      "learning_rate": 6.398305084745762e-05,
+      "loss": 17.4304,
+      "mean_token_accuracy": 0.6989135831594467,
+      "step": 585
+    },
+    {
+      "epoch": 0.39550863080274845,
+      "grad_norm": 1.8729442358016968,
+      "learning_rate": 6.362994350282486e-05,
+      "loss": 16.9125,
+      "mean_token_accuracy": 0.708356649428606,
+      "step": 590
+    },
+    {
+      "epoch": 0.39886039886039887,
+      "grad_norm": 2.099592685699463,
+      "learning_rate": 6.327683615819209e-05,
+      "loss": 17.542,
+      "mean_token_accuracy": 0.6888726130127907,
+      "step": 595
+    },
+    {
+      "epoch": 0.4022121669180493,
+      "grad_norm": 1.6204314231872559,
+      "learning_rate": 6.292372881355933e-05,
+      "loss": 16.9305,
+      "mean_token_accuracy": 0.7038852870464325,
+      "step": 600
+    },
+    {
+      "epoch": 0.4055639349756997,
+      "grad_norm": 2.12034010887146,
+      "learning_rate": 6.257062146892656e-05,
+      "loss": 17.0389,
+      "mean_token_accuracy": 0.704576326906681,
+      "step": 605
+    },
+    {
+      "epoch": 0.4089157030333501,
+      "grad_norm": 1.6821502447128296,
+      "learning_rate": 6.221751412429379e-05,
+      "loss": 16.788,
+      "mean_token_accuracy": 0.7000284940004349,
+      "step": 610
+    },
+    {
+      "epoch": 0.4122674710910005,
+      "grad_norm": 1.8137435913085938,
+      "learning_rate": 6.186440677966102e-05,
+      "loss": 17.5926,
+      "mean_token_accuracy": 0.6961537927389145,
+      "step": 615
+    },
+    {
+      "epoch": 0.4156192391486509,
+      "grad_norm": 1.6652235984802246,
+      "learning_rate": 6.151129943502825e-05,
+      "loss": 17.3539,
+      "mean_token_accuracy": 0.7028377398848533,
+      "step": 620
+    },
+    {
+      "epoch": 0.4189710072063013,
+      "grad_norm": 1.766480803489685,
+      "learning_rate": 6.115819209039548e-05,
+      "loss": 17.529,
+      "mean_token_accuracy": 0.6905739739537239,
+      "step": 625
+    },
+    {
+      "epoch": 0.42232277526395173,
+      "grad_norm": 1.6319854259490967,
+      "learning_rate": 6.080508474576272e-05,
+      "loss": 16.9847,
+      "mean_token_accuracy": 0.7060947254300117,
+      "step": 630
+    },
+    {
+      "epoch": 0.42567454332160215,
+      "grad_norm": 2.1006696224212646,
+      "learning_rate": 6.045197740112994e-05,
+      "loss": 16.9317,
+      "mean_token_accuracy": 0.7015593230724335,
+      "step": 635
+    },
+    {
+      "epoch": 0.42902631137925257,
+      "grad_norm": 1.7353427410125732,
+      "learning_rate": 6.009887005649718e-05,
+      "loss": 17.4744,
+      "mean_token_accuracy": 0.7001501567661762,
+      "step": 640
+    },
+    {
+      "epoch": 0.432378079436903,
+      "grad_norm": 1.9449700117111206,
+      "learning_rate": 5.974576271186441e-05,
+      "loss": 16.8705,
+      "mean_token_accuracy": 0.7026407413184643,
+      "step": 645
+    },
+    {
+      "epoch": 0.4357298474945534,
+      "grad_norm": 1.6030067205429077,
+      "learning_rate": 5.9392655367231644e-05,
+      "loss": 16.8924,
+      "mean_token_accuracy": 0.702277285605669,
+      "step": 650
+    },
+    {
+      "epoch": 0.43908161555220376,
+      "grad_norm": 1.5722424983978271,
+      "learning_rate": 5.903954802259888e-05,
+      "loss": 17.364,
+      "mean_token_accuracy": 0.6959278948605061,
+      "step": 655
+    },
+    {
+      "epoch": 0.4424333836098542,
+      "grad_norm": 1.8168216943740845,
+      "learning_rate": 5.86864406779661e-05,
+      "loss": 16.704,
+      "mean_token_accuracy": 0.7045813865959645,
+      "step": 660
+    },
+    {
+      "epoch": 0.4457851516675046,
+      "grad_norm": 1.905402660369873,
+      "learning_rate": 5.833333333333334e-05,
+      "loss": 16.8896,
+      "mean_token_accuracy": 0.7026248089969158,
+      "step": 665
+    },
+    {
+      "epoch": 0.449136919725155,
+      "grad_norm": 1.7437454462051392,
+      "learning_rate": 5.798022598870056e-05,
+      "loss": 17.0496,
+      "mean_token_accuracy": 0.702862861007452,
+      "step": 670
+    },
+    {
+      "epoch": 0.45248868778280543,
+      "grad_norm": 1.7496871948242188,
+      "learning_rate": 5.76271186440678e-05,
+      "loss": 16.7024,
+      "mean_token_accuracy": 0.7073140636086463,
+      "step": 675
+    },
+    {
+      "epoch": 0.45584045584045585,
+      "grad_norm": 1.6521803140640259,
+      "learning_rate": 5.727401129943503e-05,
+      "loss": 17.4437,
+      "mean_token_accuracy": 0.6910906590521335,
+      "step": 680
+    },
+    {
+      "epoch": 0.45919222389810627,
+      "grad_norm": 1.7904677391052246,
+      "learning_rate": 5.6920903954802264e-05,
+      "loss": 17.4803,
+      "mean_token_accuracy": 0.6987466789782047,
+      "step": 685
+    },
+    {
+      "epoch": 0.4625439919557567,
+      "grad_norm": 2.4545388221740723,
+      "learning_rate": 5.65677966101695e-05,
+      "loss": 17.2987,
+      "mean_token_accuracy": 0.699196208268404,
+      "step": 690
+    },
+    {
+      "epoch": 0.46589576001340705,
+      "grad_norm": 1.6428866386413574,
+      "learning_rate": 5.6214689265536723e-05,
+      "loss": 16.7636,
+      "mean_token_accuracy": 0.7029999569058418,
+      "step": 695
+    },
+    {
+      "epoch": 0.46924752807105746,
+      "grad_norm": 1.9685977697372437,
+      "learning_rate": 5.586158192090396e-05,
+      "loss": 17.3887,
+      "mean_token_accuracy": 0.6938736639916897,
+      "step": 700
+    },
+    {
+      "epoch": 0.4725992961287079,
+      "grad_norm": 1.5567928552627563,
+      "learning_rate": 5.550847457627118e-05,
+      "loss": 17.1879,
+      "mean_token_accuracy": 0.7024729043245316,
+      "step": 705
+    },
+    {
+      "epoch": 0.4759510641863583,
+      "grad_norm": 1.6846567392349243,
+      "learning_rate": 5.515536723163842e-05,
+      "loss": 16.8679,
+      "mean_token_accuracy": 0.7025640495121479,
+      "step": 710
+    },
+    {
+      "epoch": 0.4793028322440087,
+      "grad_norm": 1.6596832275390625,
+      "learning_rate": 5.480225988700565e-05,
+      "loss": 16.7137,
+      "mean_token_accuracy": 0.7031160019338131,
+      "step": 715
+    },
+    {
+      "epoch": 0.48265460030165913,
+      "grad_norm": 2.04453444480896,
+      "learning_rate": 5.4449152542372885e-05,
+      "loss": 17.0646,
+      "mean_token_accuracy": 0.7018779084086418,
+      "step": 720
+    },
+    {
+      "epoch": 0.48600636835930955,
+      "grad_norm": 1.7244528532028198,
+      "learning_rate": 5.409604519774012e-05,
+      "loss": 17.1897,
+      "mean_token_accuracy": 0.6981223806738853,
+      "step": 725
+    },
+    {
+      "epoch": 0.48935813641695997,
+      "grad_norm": 1.6929802894592285,
+      "learning_rate": 5.3742937853107344e-05,
+      "loss": 17.2678,
+      "mean_token_accuracy": 0.6996262572705746,
+      "step": 730
+    },
+    {
+      "epoch": 0.4927099044746104,
+      "grad_norm": 1.7945303916931152,
+      "learning_rate": 5.338983050847458e-05,
+      "loss": 17.1465,
+      "mean_token_accuracy": 0.7002299666404724,
+      "step": 735
+    },
+    {
+      "epoch": 0.49606167253226074,
+      "grad_norm": 1.5936013460159302,
+      "learning_rate": 5.30367231638418e-05,
+      "loss": 17.0265,
+      "mean_token_accuracy": 0.6998031720519066,
+      "step": 740
+    },
+    {
+      "epoch": 0.49941344058991116,
+      "grad_norm": 1.553004264831543,
+      "learning_rate": 5.268361581920904e-05,
+      "loss": 16.7301,
+      "mean_token_accuracy": 0.7022854961454869,
+      "step": 745
+    },
+    {
+      "epoch": 0.5027652086475616,
+      "grad_norm": 1.7667690515518188,
+      "learning_rate": 5.2330508474576275e-05,
+      "loss": 16.8576,
+      "mean_token_accuracy": 0.7085686258971691,
+      "step": 750
+    },
+    {
+      "epoch": 0.5027652086475616,
+      "eval_loss": 1.0600364208221436,
+      "eval_mean_token_accuracy": 0.7049777010093035,
+      "eval_runtime": 1736.5707,
+      "eval_samples_per_second": 1.392,
+      "eval_steps_per_second": 0.174,
+      "step": 750
+    },
+    {
+      "epoch": 0.506116976705212,
+      "grad_norm": 1.4901829957962036,
+      "learning_rate": 5.1977401129943505e-05,
+      "loss": 17.0004,
+      "mean_token_accuracy": 0.6990960523486137,
+      "step": 755
+    },
+    {
+      "epoch": 0.5094687447628624,
+      "grad_norm": 1.8451662063598633,
+      "learning_rate": 5.162429378531074e-05,
+      "loss": 17.2012,
+      "mean_token_accuracy": 0.7007680244743824,
+      "step": 760
+    },
+    {
+      "epoch": 0.5128205128205128,
+      "grad_norm": 1.6952011585235596,
+      "learning_rate": 5.1271186440677964e-05,
+      "loss": 17.612,
+      "mean_token_accuracy": 0.6927438467741013,
+      "step": 765
+    },
+    {
+      "epoch": 0.5161722808781632,
+      "grad_norm": 1.7307817935943604,
+      "learning_rate": 5.09180790960452e-05,
+      "loss": 16.8776,
+      "mean_token_accuracy": 0.706513649225235,
+      "step": 770
+    },
+    {
+      "epoch": 0.5195240489358136,
+      "grad_norm": 1.6692585945129395,
+      "learning_rate": 5.056497175141243e-05,
+      "loss": 17.0364,
+      "mean_token_accuracy": 0.704279126226902,
+      "step": 775
+    },
+    {
+      "epoch": 0.5228758169934641,
+      "grad_norm": 1.6963402032852173,
+      "learning_rate": 5.0211864406779666e-05,
+      "loss": 16.8957,
+      "mean_token_accuracy": 0.7085353158414364,
+      "step": 780
+    },
+    {
+      "epoch": 0.5262275850511144,
+      "grad_norm": 1.678458571434021,
+      "learning_rate": 4.9858757062146896e-05,
+      "loss": 17.7932,
+      "mean_token_accuracy": 0.6964584030210972,
+      "step": 785
+    },
+    {
+      "epoch": 0.5295793531087649,
+      "grad_norm": 1.7449827194213867,
+      "learning_rate": 4.9505649717514125e-05,
+      "loss": 16.8765,
+      "mean_token_accuracy": 0.7036922007799149,
+      "step": 790
+    },
+    {
+      "epoch": 0.5329311211664153,
+      "grad_norm": 1.7107524871826172,
+      "learning_rate": 4.915254237288136e-05,
+      "loss": 17.243,
+      "mean_token_accuracy": 0.6997682720422744,
+      "step": 795
+    },
+    {
+      "epoch": 0.5362828892240656,
+      "grad_norm": 1.6416223049163818,
+      "learning_rate": 4.879943502824859e-05,
+      "loss": 16.7253,
+      "mean_token_accuracy": 0.7050332672894001,
+      "step": 800
+    },
+    {
+      "epoch": 0.5396346572817161,
+      "grad_norm": 1.867213249206543,
+      "learning_rate": 4.844632768361582e-05,
+      "loss": 16.8566,
+      "mean_token_accuracy": 0.7032786093652248,
+      "step": 805
+    },
+    {
+      "epoch": 0.5429864253393665,
+      "grad_norm": 1.6539360284805298,
+      "learning_rate": 4.809322033898305e-05,
+      "loss": 16.6993,
+      "mean_token_accuracy": 0.7117977932095527,
+      "step": 810
+    },
+    {
+      "epoch": 0.546338193397017,
+      "grad_norm": 1.752715826034546,
+      "learning_rate": 4.7740112994350286e-05,
+      "loss": 17.5809,
+      "mean_token_accuracy": 0.6992670528590679,
+      "step": 815
+    },
+    {
+      "epoch": 0.5496899614546673,
+      "grad_norm": 1.806174397468567,
+      "learning_rate": 4.7387005649717516e-05,
+      "loss": 17.1588,
+      "mean_token_accuracy": 0.6960965767502785,
+      "step": 820
+    },
+    {
+      "epoch": 0.5530417295123178,
+      "grad_norm": 1.719764232635498,
+      "learning_rate": 4.703389830508475e-05,
+      "loss": 16.8685,
+      "mean_token_accuracy": 0.7025568410754204,
+      "step": 825
+    },
+    {
+      "epoch": 0.5563934975699681,
+      "grad_norm": 1.7800629138946533,
+      "learning_rate": 4.668079096045198e-05,
+      "loss": 16.8872,
+      "mean_token_accuracy": 0.6994628652930259,
+      "step": 830
+    },
+    {
+      "epoch": 0.5597452656276186,
+      "grad_norm": 1.7011103630065918,
+      "learning_rate": 4.632768361581921e-05,
+      "loss": 17.2342,
+      "mean_token_accuracy": 0.7006913289427757,
+      "step": 835
+    },
+    {
+      "epoch": 0.563097033685269,
+      "grad_norm": 1.6887695789337158,
+      "learning_rate": 4.597457627118644e-05,
+      "loss": 16.7385,
+      "mean_token_accuracy": 0.7045929700136184,
+      "step": 840
+    },
+    {
+      "epoch": 0.5664488017429193,
+      "grad_norm": 1.9496142864227295,
+      "learning_rate": 4.562146892655367e-05,
+      "loss": 16.8387,
+      "mean_token_accuracy": 0.7083131410181522,
+      "step": 845
+    },
+    {
+      "epoch": 0.5698005698005698,
+      "grad_norm": 1.7757388353347778,
+      "learning_rate": 4.5268361581920906e-05,
+      "loss": 17.3856,
+      "mean_token_accuracy": 0.6994826771318913,
+      "step": 850
+    },
+    {
+      "epoch": 0.5731523378582202,
+      "grad_norm": 1.7115302085876465,
+      "learning_rate": 4.491525423728814e-05,
+      "loss": 16.5993,
+      "mean_token_accuracy": 0.7093915119767189,
+      "step": 855
+    },
+    {
+      "epoch": 0.5765041059158706,
+      "grad_norm": 1.7968231439590454,
+      "learning_rate": 4.456214689265537e-05,
+      "loss": 16.8983,
+      "mean_token_accuracy": 0.7087731070816516,
+      "step": 860
+    },
+    {
+      "epoch": 0.579855873973521,
+      "grad_norm": 1.6066899299621582,
+      "learning_rate": 4.42090395480226e-05,
+      "loss": 16.7126,
+      "mean_token_accuracy": 0.7053335346281528,
+      "step": 865
+    },
+    {
+      "epoch": 0.5832076420311715,
+      "grad_norm": 1.6380205154418945,
+      "learning_rate": 4.385593220338983e-05,
+      "loss": 17.0037,
+      "mean_token_accuracy": 0.7038719221949578,
+      "step": 870
+    },
+    {
+      "epoch": 0.5865594100888218,
+      "grad_norm": 1.8956695795059204,
+      "learning_rate": 4.350282485875706e-05,
+      "loss": 16.9679,
+      "mean_token_accuracy": 0.6983371920883655,
+      "step": 875
+    },
+    {
+      "epoch": 0.5899111781464723,
+      "grad_norm": 1.625135064125061,
+      "learning_rate": 4.314971751412429e-05,
+      "loss": 17.0642,
+      "mean_token_accuracy": 0.7067640118300915,
+      "step": 880
+    },
+    {
+      "epoch": 0.5932629462041227,
+      "grad_norm": 1.6344581842422485,
+      "learning_rate": 4.279661016949153e-05,
+      "loss": 16.3079,
+      "mean_token_accuracy": 0.7225491903722286,
+      "step": 885
+    },
+    {
+      "epoch": 0.596614714261773,
+      "grad_norm": 1.7680976390838623,
+      "learning_rate": 4.244350282485876e-05,
+      "loss": 16.7187,
+      "mean_token_accuracy": 0.7041032016277313,
+      "step": 890
+    },
+    {
+      "epoch": 0.5999664823194235,
+      "grad_norm": 1.8056613206863403,
+      "learning_rate": 4.209039548022599e-05,
+      "loss": 17.3536,
+      "mean_token_accuracy": 0.6975419208407402,
+      "step": 895
+    },
+    {
+      "epoch": 0.6033182503770739,
+      "grad_norm": 1.8398966789245605,
+      "learning_rate": 4.173728813559322e-05,
+      "loss": 16.6245,
+      "mean_token_accuracy": 0.7088275127112865,
+      "step": 900
+    },
+    {
+      "epoch": 0.6066700184347243,
+      "grad_norm": 1.8332566022872925,
+      "learning_rate": 4.138418079096045e-05,
+      "loss": 17.0128,
+      "mean_token_accuracy": 0.7018843114376068,
+      "step": 905
+    },
+    {
+      "epoch": 0.6100217864923747,
+      "grad_norm": 1.6582337617874146,
+      "learning_rate": 4.103107344632768e-05,
+      "loss": 16.8948,
+      "mean_token_accuracy": 0.7051651798188686,
+      "step": 910
+    },
+    {
+      "epoch": 0.6133735545500252,
+      "grad_norm": 1.7373839616775513,
+      "learning_rate": 4.067796610169492e-05,
+      "loss": 16.9138,
+      "mean_token_accuracy": 0.7022108249366283,
+      "step": 915
+    },
+    {
+      "epoch": 0.6167253226076755,
+      "grad_norm": 1.6373577117919922,
+      "learning_rate": 4.0324858757062154e-05,
+      "loss": 17.0573,
+      "mean_token_accuracy": 0.7042267486453057,
+      "step": 920
+    },
+    {
+      "epoch": 0.620077090665326,
+      "grad_norm": 1.581024408340454,
+      "learning_rate": 3.997175141242938e-05,
+      "loss": 16.6234,
+      "mean_token_accuracy": 0.7054463028907776,
+      "step": 925
+    },
+    {
+      "epoch": 0.6234288587229764,
+      "grad_norm": 1.6900616884231567,
+      "learning_rate": 3.961864406779661e-05,
+      "loss": 17.0468,
+      "mean_token_accuracy": 0.7014504976570606,
+      "step": 930
+    },
+    {
+      "epoch": 0.6267806267806267,
+      "grad_norm": 1.6560430526733398,
+      "learning_rate": 3.926553672316384e-05,
+      "loss": 16.909,
+      "mean_token_accuracy": 0.7064756542444229,
+      "step": 935
+    },
+    {
+      "epoch": 0.6301323948382772,
+      "grad_norm": 1.8687000274658203,
+      "learning_rate": 3.891242937853107e-05,
+      "loss": 17.0047,
+      "mean_token_accuracy": 0.7055176287889481,
+      "step": 940
+    },
+    {
+      "epoch": 0.6334841628959276,
+      "grad_norm": 1.777716040611267,
+      "learning_rate": 3.855932203389831e-05,
+      "loss": 16.556,
+      "mean_token_accuracy": 0.7047871246933937,
+      "step": 945
+    },
+    {
+      "epoch": 0.636835930953578,
+      "grad_norm": 1.6830016374588013,
+      "learning_rate": 3.820621468926554e-05,
+      "loss": 16.5832,
+      "mean_token_accuracy": 0.7049862682819367,
+      "step": 950
+    },
+    {
+      "epoch": 0.6401876990112284,
+      "grad_norm": 1.5959638357162476,
+      "learning_rate": 3.7853107344632774e-05,
+      "loss": 16.8336,
+      "mean_token_accuracy": 0.7072055459022522,
+      "step": 955
+    },
+    {
+      "epoch": 0.6435394670688789,
+      "grad_norm": 1.82794189453125,
+      "learning_rate": 3.7500000000000003e-05,
+      "loss": 16.6644,
+      "mean_token_accuracy": 0.7058505766093731,
+      "step": 960
+    },
+    {
+      "epoch": 0.6468912351265292,
+      "grad_norm": 1.6554478406906128,
+      "learning_rate": 3.714689265536723e-05,
+      "loss": 16.2796,
+      "mean_token_accuracy": 0.7101977132260799,
+      "step": 965
+    },
+    {
+      "epoch": 0.6502430031841796,
+      "grad_norm": 1.8698370456695557,
+      "learning_rate": 3.679378531073446e-05,
+      "loss": 16.1934,
+      "mean_token_accuracy": 0.7142874717712402,
+      "step": 970
+    },
+    {
+      "epoch": 0.6535947712418301,
+      "grad_norm": 1.8040566444396973,
+      "learning_rate": 3.644067796610169e-05,
+      "loss": 16.5345,
+      "mean_token_accuracy": 0.7125143676996231,
+      "step": 975
+    },
+    {
+      "epoch": 0.6569465392994804,
+      "grad_norm": 1.6644558906555176,
+      "learning_rate": 3.608757062146893e-05,
+      "loss": 16.508,
+      "mean_token_accuracy": 0.7078846462070942,
+      "step": 980
+    },
+    {
+      "epoch": 0.6602983073571309,
+      "grad_norm": 1.7228506803512573,
+      "learning_rate": 3.573446327683616e-05,
+      "loss": 16.8474,
+      "mean_token_accuracy": 0.7084795109927654,
+      "step": 985
+    },
+    {
+      "epoch": 0.6636500754147813,
+      "grad_norm": 1.486241102218628,
+      "learning_rate": 3.5381355932203394e-05,
+      "loss": 17.1453,
+      "mean_token_accuracy": 0.6975291892886162,
+      "step": 990
+    },
+    {
+      "epoch": 0.6670018434724317,
+      "grad_norm": 1.7130765914916992,
+      "learning_rate": 3.5028248587570624e-05,
+      "loss": 16.458,
+      "mean_token_accuracy": 0.7106956362724304,
+      "step": 995
+    },
+    {
+      "epoch": 0.6703536115300821,
+      "grad_norm": 1.863926649093628,
+      "learning_rate": 3.467514124293785e-05,
+      "loss": 17.3095,
+      "mean_token_accuracy": 0.6962033234536648,
+      "step": 1000
+    },
+    {
+      "epoch": 0.6737053795877326,
+      "grad_norm": 1.6535072326660156,
+      "learning_rate": 3.432203389830508e-05,
+      "loss": 16.6846,
+      "mean_token_accuracy": 0.7084034703671932,
+      "step": 1005
+    },
+    {
+      "epoch": 0.6770571476453829,
+      "grad_norm": 1.7278594970703125,
+      "learning_rate": 3.396892655367232e-05,
+      "loss": 16.9805,
+      "mean_token_accuracy": 0.7026786416769027,
+      "step": 1010
+    },
+    {
+      "epoch": 0.6804089157030333,
+      "grad_norm": 1.9055004119873047,
+      "learning_rate": 3.361581920903955e-05,
+      "loss": 17.2562,
+      "mean_token_accuracy": 0.6977267302572727,
+      "step": 1015
+    },
+    {
+      "epoch": 0.6837606837606838,
+      "grad_norm": 1.6398614645004272,
+      "learning_rate": 3.326271186440678e-05,
+      "loss": 17.3378,
+      "mean_token_accuracy": 0.6958214737474918,
+      "step": 1020
+    },
+    {
+      "epoch": 0.6871124518183341,
+      "grad_norm": 1.926950454711914,
+      "learning_rate": 3.2909604519774014e-05,
+      "loss": 16.6536,
+      "mean_token_accuracy": 0.7083842910826206,
+      "step": 1025
+    },
+    {
+      "epoch": 0.6904642198759846,
+      "grad_norm": 1.8061659336090088,
+      "learning_rate": 3.2556497175141244e-05,
+      "loss": 16.643,
+      "mean_token_accuracy": 0.7093963578343392,
+      "step": 1030
+    },
+    {
+      "epoch": 0.693815987933635,
+      "grad_norm": 1.6816084384918213,
+      "learning_rate": 3.2203389830508473e-05,
+      "loss": 16.9696,
+      "mean_token_accuracy": 0.7000316813588142,
+      "step": 1035
+    },
+    {
+      "epoch": 0.6971677559912854,
+      "grad_norm": 1.630842685699463,
+      "learning_rate": 3.185028248587571e-05,
+      "loss": 16.587,
+      "mean_token_accuracy": 0.7107978977262974,
+      "step": 1040
+    },
+    {
+      "epoch": 0.7005195240489358,
+      "grad_norm": 1.755123257637024,
+      "learning_rate": 3.149717514124294e-05,
+      "loss": 17.0736,
+      "mean_token_accuracy": 0.7017260067164898,
+      "step": 1045
+    },
+    {
+      "epoch": 0.7038712921065863,
+      "grad_norm": 1.4850029945373535,
+      "learning_rate": 3.114406779661017e-05,
+      "loss": 16.3165,
+      "mean_token_accuracy": 0.7119720429182053,
+      "step": 1050
+    },
+    {
+      "epoch": 0.7072230601642366,
+      "grad_norm": 1.916961908340454,
+      "learning_rate": 3.0790960451977405e-05,
+      "loss": 17.0237,
+      "mean_token_accuracy": 0.6976533338427544,
+      "step": 1055
+    },
+    {
+      "epoch": 0.710574828221887,
+      "grad_norm": 1.5003294944763184,
+      "learning_rate": 3.043785310734463e-05,
+      "loss": 16.8504,
+      "mean_token_accuracy": 0.7056308597326278,
+      "step": 1060
+    },
+    {
+      "epoch": 0.7139265962795375,
+      "grad_norm": 1.9166836738586426,
+      "learning_rate": 3.0084745762711864e-05,
+      "loss": 16.8231,
+      "mean_token_accuracy": 0.7023352533578873,
+      "step": 1065
+    },
+    {
+      "epoch": 0.7172783643371878,
+      "grad_norm": 1.7789411544799805,
+      "learning_rate": 2.97316384180791e-05,
+      "loss": 17.3132,
+      "mean_token_accuracy": 0.6994914725422859,
+      "step": 1070
+    },
+    {
+      "epoch": 0.7206301323948383,
+      "grad_norm": 1.7289875745773315,
+      "learning_rate": 2.937853107344633e-05,
+      "loss": 17.3902,
+      "mean_token_accuracy": 0.69447166249156,
+      "step": 1075
+    },
+    {
+      "epoch": 0.7239819004524887,
+      "grad_norm": 1.4835467338562012,
+      "learning_rate": 2.902542372881356e-05,
+      "loss": 16.751,
+      "mean_token_accuracy": 0.7052346661686897,
+      "step": 1080
+    },
+    {
+      "epoch": 0.7273336685101391,
+      "grad_norm": 1.5802119970321655,
+      "learning_rate": 2.8672316384180792e-05,
+      "loss": 16.6574,
+      "mean_token_accuracy": 0.7059398606419564,
+      "step": 1085
+    },
+    {
+      "epoch": 0.7306854365677895,
+      "grad_norm": 1.8420851230621338,
+      "learning_rate": 2.8319209039548022e-05,
+      "loss": 16.9315,
+      "mean_token_accuracy": 0.7063411138951778,
+      "step": 1090
+    },
+    {
+      "epoch": 0.7340372046254399,
+      "grad_norm": 1.7593777179718018,
+      "learning_rate": 2.7966101694915255e-05,
+      "loss": 16.8653,
+      "mean_token_accuracy": 0.7089171193540096,
+      "step": 1095
+    },
+    {
+      "epoch": 0.7373889726830903,
+      "grad_norm": 1.681443452835083,
+      "learning_rate": 2.7612994350282488e-05,
+      "loss": 16.9878,
+      "mean_token_accuracy": 0.7057393230497837,
+      "step": 1100
+    },
+    {
+      "epoch": 0.7407407407407407,
+      "grad_norm": 1.6064281463623047,
+      "learning_rate": 2.725988700564972e-05,
+      "loss": 16.6153,
+      "mean_token_accuracy": 0.7038764618337154,
+      "step": 1105
+    },
+    {
+      "epoch": 0.7440925087983912,
+      "grad_norm": 1.5632483959197998,
+      "learning_rate": 2.690677966101695e-05,
+      "loss": 16.0927,
+      "mean_token_accuracy": 0.7171440742909908,
+      "step": 1110
+    },
+    {
+      "epoch": 0.7474442768560415,
+      "grad_norm": 1.8588156700134277,
+      "learning_rate": 2.6553672316384183e-05,
+      "loss": 16.5765,
+      "mean_token_accuracy": 0.7098327249288559,
+      "step": 1115
+    },
+    {
+      "epoch": 0.750796044913692,
+      "grad_norm": 1.5576221942901611,
+      "learning_rate": 2.6200564971751413e-05,
+      "loss": 16.6568,
+      "mean_token_accuracy": 0.7029327027499676,
+      "step": 1120
+    },
+    {
+      "epoch": 0.7541478129713424,
+      "grad_norm": 1.645244836807251,
+      "learning_rate": 2.5847457627118642e-05,
+      "loss": 16.7294,
+      "mean_token_accuracy": 0.7060277953743934,
+      "step": 1125
+    },
+    {
+      "epoch": 0.7574995810289928,
+      "grad_norm": 1.4038984775543213,
+      "learning_rate": 2.549435028248588e-05,
+      "loss": 16.5925,
+      "mean_token_accuracy": 0.7068064086139202,
+      "step": 1130
+    },
+    {
+      "epoch": 0.7608513490866432,
+      "grad_norm": 1.7987641096115112,
+      "learning_rate": 2.514124293785311e-05,
+      "loss": 16.6834,
+      "mean_token_accuracy": 0.7070130936801433,
+      "step": 1135
+    },
+    {
+      "epoch": 0.7642031171442936,
+      "grad_norm": 1.5423444509506226,
+      "learning_rate": 2.478813559322034e-05,
+      "loss": 16.4551,
+      "mean_token_accuracy": 0.7121224895119667,
+      "step": 1140
+    },
+    {
+      "epoch": 0.767554885201944,
+      "grad_norm": 1.7546942234039307,
+      "learning_rate": 2.443502824858757e-05,
+      "loss": 16.9741,
+      "mean_token_accuracy": 0.7010989025235176,
+      "step": 1145
+    },
+    {
+      "epoch": 0.7709066532595944,
+      "grad_norm": 1.8481935262680054,
+      "learning_rate": 2.4081920903954803e-05,
+      "loss": 16.6323,
+      "mean_token_accuracy": 0.7058765202760696,
+      "step": 1150
+    },
+    {
+      "epoch": 0.7742584213172449,
+      "grad_norm": 1.6855909824371338,
+      "learning_rate": 2.3728813559322036e-05,
+      "loss": 16.6844,
+      "mean_token_accuracy": 0.7119428858160972,
+      "step": 1155
+    },
+    {
+      "epoch": 0.7776101893748952,
+      "grad_norm": 1.9828130006790161,
+      "learning_rate": 2.3375706214689266e-05,
+      "loss": 16.866,
+      "mean_token_accuracy": 0.7036800056695938,
+      "step": 1160
+    },
+    {
+      "epoch": 0.7809619574325457,
+      "grad_norm": 1.5005120038986206,
+      "learning_rate": 2.30225988700565e-05,
+      "loss": 16.3539,
+      "mean_token_accuracy": 0.711839384585619,
+      "step": 1165
+    },
+    {
+      "epoch": 0.7843137254901961,
+      "grad_norm": 2.262735366821289,
+      "learning_rate": 2.266949152542373e-05,
+      "loss": 16.4102,
+      "mean_token_accuracy": 0.7110463745892048,
+      "step": 1170
+    },
+    {
+      "epoch": 0.7876654935478465,
+      "grad_norm": 1.6699568033218384,
+      "learning_rate": 2.231638418079096e-05,
+      "loss": 17.1027,
+      "mean_token_accuracy": 0.7031991191208362,
+      "step": 1175
+    },
+    {
+      "epoch": 0.7910172616054969,
+      "grad_norm": 1.6248890161514282,
+      "learning_rate": 2.196327683615819e-05,
+      "loss": 16.3399,
+      "mean_token_accuracy": 0.7143234215676785,
+      "step": 1180
+    },
+    {
+      "epoch": 0.7943690296631473,
+      "grad_norm": 1.7570775747299194,
+      "learning_rate": 2.1610169491525427e-05,
+      "loss": 16.2255,
+      "mean_token_accuracy": 0.7123358778655529,
+      "step": 1185
+    },
+    {
+      "epoch": 0.7977207977207977,
+      "grad_norm": 1.9391677379608154,
+      "learning_rate": 2.1257062146892657e-05,
+      "loss": 16.3472,
+      "mean_token_accuracy": 0.711616413295269,
+      "step": 1190
+    },
+    {
+      "epoch": 0.8010725657784481,
+      "grad_norm": 1.8997981548309326,
+      "learning_rate": 2.0903954802259886e-05,
+      "loss": 16.5601,
+      "mean_token_accuracy": 0.7071553356945515,
+      "step": 1195
+    },
+    {
+      "epoch": 0.8044243338360986,
+      "grad_norm": 1.6094359159469604,
+      "learning_rate": 2.055084745762712e-05,
+      "loss": 16.622,
+      "mean_token_accuracy": 0.7043877936899662,
+      "step": 1200
+    },
+    {
+      "epoch": 0.8077761018937489,
+      "grad_norm": 1.7940973043441772,
+      "learning_rate": 2.0197740112994352e-05,
+      "loss": 16.6535,
+      "mean_token_accuracy": 0.705554535984993,
+      "step": 1205
+    },
+    {
+      "epoch": 0.8111278699513994,
+      "grad_norm": 1.6890041828155518,
+      "learning_rate": 1.984463276836158e-05,
+      "loss": 17.2328,
+      "mean_token_accuracy": 0.6988375537097454,
+      "step": 1210
+    },
+    {
+      "epoch": 0.8144796380090498,
+      "grad_norm": 1.5568735599517822,
+      "learning_rate": 1.9491525423728814e-05,
+      "loss": 16.9753,
+      "mean_token_accuracy": 0.7015632651746273,
+      "step": 1215
+    },
+    {
+      "epoch": 0.8178314060667002,
+      "grad_norm": 1.7157835960388184,
+      "learning_rate": 1.9138418079096047e-05,
+      "loss": 16.3668,
+      "mean_token_accuracy": 0.7098449252545833,
+      "step": 1220
+    },
+    {
+      "epoch": 0.8211831741243506,
+      "grad_norm": 1.7175644636154175,
+      "learning_rate": 1.8785310734463277e-05,
+      "loss": 16.8061,
+      "mean_token_accuracy": 0.7032932281494141,
+      "step": 1225
+    },
+    {
+      "epoch": 0.824534942182001,
+      "grad_norm": 1.7225829362869263,
+      "learning_rate": 1.843220338983051e-05,
+      "loss": 16.5716,
+      "mean_token_accuracy": 0.7074852548539639,
+      "step": 1230
+    },
+    {
+      "epoch": 0.8278867102396514,
+      "grad_norm": 1.8654727935791016,
+      "learning_rate": 1.8079096045197743e-05,
+      "loss": 16.8172,
+      "mean_token_accuracy": 0.7035241700708866,
+      "step": 1235
+    },
+    {
+      "epoch": 0.8312384782973018,
+      "grad_norm": 1.9604694843292236,
+      "learning_rate": 1.7725988700564972e-05,
+      "loss": 16.2992,
+      "mean_token_accuracy": 0.714275274425745,
+      "step": 1240
+    },
+    {
+      "epoch": 0.8345902463549523,
+      "grad_norm": 1.7569185495376587,
+      "learning_rate": 1.7372881355932205e-05,
+      "loss": 16.6269,
+      "mean_token_accuracy": 0.7052666112780571,
+      "step": 1245
+    },
+    {
+      "epoch": 0.8379420144126026,
+      "grad_norm": 1.6537069082260132,
+      "learning_rate": 1.7019774011299435e-05,
+      "loss": 16.5978,
+      "mean_token_accuracy": 0.708269502967596,
+      "step": 1250
+    },
+    {
+      "epoch": 0.8412937824702531,
+      "grad_norm": 1.8623359203338623,
+      "learning_rate": 1.6666666666666667e-05,
+      "loss": 16.1831,
+      "mean_token_accuracy": 0.7164609245955944,
+      "step": 1255
+    },
+    {
+      "epoch": 0.8446455505279035,
+      "grad_norm": 1.7004101276397705,
+      "learning_rate": 1.63135593220339e-05,
+      "loss": 16.9611,
+      "mean_token_accuracy": 0.7057129152119159,
+      "step": 1260
+    },
+    {
+      "epoch": 0.8479973185855538,
+      "grad_norm": 1.8294973373413086,
+      "learning_rate": 1.596045197740113e-05,
+      "loss": 16.8036,
+      "mean_token_accuracy": 0.7046464517712593,
+      "step": 1265
+    },
+    {
+      "epoch": 0.8513490866432043,
+      "grad_norm": 1.7992702722549438,
+      "learning_rate": 1.5607344632768363e-05,
+      "loss": 16.139,
+      "mean_token_accuracy": 0.7126708298921585,
+      "step": 1270
+    },
+    {
+      "epoch": 0.8547008547008547,
+      "grad_norm": 2.033846855163574,
+      "learning_rate": 1.5254237288135596e-05,
+      "loss": 16.49,
+      "mean_token_accuracy": 0.707030464708805,
+      "step": 1275
+    },
+    {
+      "epoch": 0.8580526227585051,
+      "grad_norm": 1.690617561340332,
+      "learning_rate": 1.4901129943502825e-05,
+      "loss": 16.7829,
+      "mean_token_accuracy": 0.7026272863149643,
+      "step": 1280
+    },
+    {
+      "epoch": 0.8614043908161555,
+      "grad_norm": 1.7161706686019897,
+      "learning_rate": 1.4548022598870056e-05,
+      "loss": 16.4907,
+      "mean_token_accuracy": 0.7054763376712799,
+      "step": 1285
+    },
+    {
+      "epoch": 0.864756158873806,
+      "grad_norm": 1.5910500288009644,
+      "learning_rate": 1.419491525423729e-05,
+      "loss": 16.3073,
+      "mean_token_accuracy": 0.7165283918380737,
+      "step": 1290
+    },
+    {
+      "epoch": 0.8681079269314563,
+      "grad_norm": 1.5939749479293823,
+      "learning_rate": 1.384180790960452e-05,
+      "loss": 16.6524,
+      "mean_token_accuracy": 0.705347529053688,
+      "step": 1295
+    },
+    {
+      "epoch": 0.8714596949891068,
+      "grad_norm": 1.7478996515274048,
+      "learning_rate": 1.3488700564971752e-05,
+      "loss": 17.1832,
+      "mean_token_accuracy": 0.6956523738801479,
+      "step": 1300
+    },
+    {
+      "epoch": 0.8748114630467572,
+      "grad_norm": 1.6442205905914307,
+      "learning_rate": 1.3135593220338985e-05,
+      "loss": 16.3978,
+      "mean_token_accuracy": 0.7132278561592102,
+      "step": 1305
+    },
+    {
+      "epoch": 0.8781632311044075,
+      "grad_norm": 1.7201565504074097,
+      "learning_rate": 1.2782485875706216e-05,
+      "loss": 16.3159,
+      "mean_token_accuracy": 0.711051919311285,
+      "step": 1310
+    },
+    {
+      "epoch": 0.881514999162058,
+      "grad_norm": 1.829209327697754,
+      "learning_rate": 1.2429378531073447e-05,
+      "loss": 16.7987,
+      "mean_token_accuracy": 0.7058401651680469,
+      "step": 1315
+    },
+    {
+      "epoch": 0.8848667672197084,
+      "grad_norm": 1.4660886526107788,
+      "learning_rate": 1.2076271186440678e-05,
+      "loss": 16.7297,
+      "mean_token_accuracy": 0.7092804253101349,
+      "step": 1320
+    },
+    {
+      "epoch": 0.8882185352773588,
+      "grad_norm": 1.4927663803100586,
+      "learning_rate": 1.172316384180791e-05,
+      "loss": 15.9333,
+      "mean_token_accuracy": 0.7158772744238376,
+      "step": 1325
+    },
+    {
+      "epoch": 0.8915703033350092,
+      "grad_norm": 1.6522186994552612,
+      "learning_rate": 1.137005649717514e-05,
+      "loss": 16.4156,
+      "mean_token_accuracy": 0.7134528748691082,
+      "step": 1330
+    },
+    {
+      "epoch": 0.8949220713926597,
+      "grad_norm": 1.7809523344039917,
+      "learning_rate": 1.1016949152542374e-05,
+      "loss": 16.2625,
+      "mean_token_accuracy": 0.7148336976766586,
+      "step": 1335
+    },
+    {
+      "epoch": 0.89827383945031,
+      "grad_norm": 1.8860619068145752,
+      "learning_rate": 1.0663841807909605e-05,
+      "loss": 16.6187,
+      "mean_token_accuracy": 0.7087382405996323,
+      "step": 1340
+    },
+    {
+      "epoch": 0.9016256075079605,
+      "grad_norm": 1.854195475578308,
+      "learning_rate": 1.0310734463276836e-05,
+      "loss": 16.5843,
+      "mean_token_accuracy": 0.7144103929400444,
+      "step": 1345
+    },
+    {
+      "epoch": 0.9049773755656109,
+      "grad_norm": 1.7052239179611206,
+      "learning_rate": 9.957627118644067e-06,
+      "loss": 16.3345,
+      "mean_token_accuracy": 0.7125584341585636,
+      "step": 1350
+    },
+    {
+      "epoch": 0.9083291436232612,
+      "grad_norm": 1.5887420177459717,
+      "learning_rate": 9.6045197740113e-06,
+      "loss": 16.2409,
+      "mean_token_accuracy": 0.7080107174813748,
+      "step": 1355
+    },
+    {
+      "epoch": 0.9116809116809117,
+      "grad_norm": 1.6052732467651367,
+      "learning_rate": 9.251412429378532e-06,
+      "loss": 16.2373,
+      "mean_token_accuracy": 0.7137157171964645,
+      "step": 1360
+    },
+    {
+      "epoch": 0.9150326797385621,
+      "grad_norm": 1.7612617015838623,
+      "learning_rate": 8.898305084745763e-06,
+      "loss": 16.0292,
+      "mean_token_accuracy": 0.7181592255830764,
+      "step": 1365
+    },
+    {
+      "epoch": 0.9183844477962125,
+      "grad_norm": 1.8271749019622803,
+      "learning_rate": 8.545197740112996e-06,
+      "loss": 16.8757,
+      "mean_token_accuracy": 0.701992305368185,
+      "step": 1370
+    },
+    {
+      "epoch": 0.9217362158538629,
+      "grad_norm": 1.6350926160812378,
+      "learning_rate": 8.192090395480225e-06,
+      "loss": 16.6061,
+      "mean_token_accuracy": 0.7089238859713077,
+      "step": 1375
+    },
+    {
+      "epoch": 0.9250879839115134,
+      "grad_norm": 1.7321621179580688,
+      "learning_rate": 7.838983050847458e-06,
+      "loss": 16.2532,
+      "mean_token_accuracy": 0.7115737572312355,
+      "step": 1380
+    },
+    {
+      "epoch": 0.9284397519691637,
+      "grad_norm": 1.8958040475845337,
+      "learning_rate": 7.48587570621469e-06,
+      "loss": 16.5068,
+      "mean_token_accuracy": 0.7108790181577206,
+      "step": 1385
+    },
+    {
+      "epoch": 0.9317915200268141,
+      "grad_norm": 1.629992127418518,
+      "learning_rate": 7.1327683615819206e-06,
+      "loss": 16.2367,
+      "mean_token_accuracy": 0.7134776934981346,
+      "step": 1390
+    },
+    {
+      "epoch": 0.9351432880844646,
+      "grad_norm": 1.904123067855835,
+      "learning_rate": 6.779661016949153e-06,
+      "loss": 16.3444,
+      "mean_token_accuracy": 0.7045241884887219,
+      "step": 1395
+    },
+    {
+      "epoch": 0.9384950561421149,
+      "grad_norm": 1.6319600343704224,
+      "learning_rate": 6.426553672316385e-06,
+      "loss": 16.3,
+      "mean_token_accuracy": 0.7118948072195053,
+      "step": 1400
+    },
+    {
+      "epoch": 0.9418468241997654,
+      "grad_norm": 1.6921709775924683,
+      "learning_rate": 6.073446327683617e-06,
+      "loss": 16.5816,
+      "mean_token_accuracy": 0.7079687170684338,
+      "step": 1405
+    },
+    {
+      "epoch": 0.9451985922574158,
+      "grad_norm": 1.636551856994629,
+      "learning_rate": 5.720338983050848e-06,
+      "loss": 16.785,
+      "mean_token_accuracy": 0.7054948009550571,
+      "step": 1410
+    },
+    {
+      "epoch": 0.9485503603150662,
+      "grad_norm": 1.6171858310699463,
+      "learning_rate": 5.367231638418079e-06,
+      "loss": 16.6877,
+      "mean_token_accuracy": 0.7033485405147075,
+      "step": 1415
+    },
+    {
+      "epoch": 0.9519021283727166,
+      "grad_norm": 1.6833641529083252,
+      "learning_rate": 5.014124293785311e-06,
+      "loss": 16.5803,
+      "mean_token_accuracy": 0.706027402728796,
+      "step": 1420
+    },
+    {
+      "epoch": 0.9552538964303671,
+      "grad_norm": 2.0238494873046875,
+      "learning_rate": 4.6610169491525425e-06,
+      "loss": 16.4305,
+      "mean_token_accuracy": 0.7110757566988468,
+      "step": 1425
+    },
+    {
+      "epoch": 0.9586056644880174,
+      "grad_norm": 1.5262683629989624,
+      "learning_rate": 4.307909604519774e-06,
+      "loss": 16.105,
+      "mean_token_accuracy": 0.7173994883894921,
+      "step": 1430
+    },
+    {
+      "epoch": 0.9619574325456678,
+      "grad_norm": 1.6822128295898438,
+      "learning_rate": 3.954802259887006e-06,
+      "loss": 17.0064,
+      "mean_token_accuracy": 0.7033144362270832,
+      "step": 1435
+    },
+    {
+      "epoch": 0.9653092006033183,
+      "grad_norm": 2.1382946968078613,
+      "learning_rate": 3.6016949152542374e-06,
+      "loss": 16.6567,
+      "mean_token_accuracy": 0.7085098147392273,
+      "step": 1440
+    },
+    {
+      "epoch": 0.9686609686609686,
+      "grad_norm": 1.6137080192565918,
+      "learning_rate": 3.248587570621469e-06,
+      "loss": 16.4193,
+      "mean_token_accuracy": 0.7077061600983143,
+      "step": 1445
+    },
+    {
+      "epoch": 0.9720127367186191,
+      "grad_norm": 1.6318018436431885,
+      "learning_rate": 2.8954802259887007e-06,
+      "loss": 16.5904,
+      "mean_token_accuracy": 0.7037704810500145,
+      "step": 1450
+    },
+    {
+      "epoch": 0.9753645047762695,
+      "grad_norm": 1.6723519563674927,
+      "learning_rate": 2.5423728813559323e-06,
+      "loss": 16.351,
+      "mean_token_accuracy": 0.715372896194458,
+      "step": 1455
+    },
+    {
+      "epoch": 0.9787162728339199,
+      "grad_norm": 2.6915719509124756,
+      "learning_rate": 2.189265536723164e-06,
+      "loss": 16.5627,
+      "mean_token_accuracy": 0.706637478619814,
+      "step": 1460
+    },
+    {
+      "epoch": 0.9820680408915703,
+      "grad_norm": 1.9349390268325806,
+      "learning_rate": 1.8361581920903956e-06,
+      "loss": 16.7821,
+      "mean_token_accuracy": 0.7010103747248649,
+      "step": 1465
+    },
+    {
+      "epoch": 0.9854198089492208,
+      "grad_norm": 1.6685172319412231,
+      "learning_rate": 1.4830508474576273e-06,
+      "loss": 16.7016,
+      "mean_token_accuracy": 0.7086931586265564,
+      "step": 1470
+    },
+    {
+      "epoch": 0.9887715770068711,
+      "grad_norm": 1.7148998975753784,
+      "learning_rate": 1.129943502824859e-06,
+      "loss": 16.4809,
+      "mean_token_accuracy": 0.7131018862128258,
+      "step": 1475
+    },
+    {
+      "epoch": 0.9921233450645215,
+      "grad_norm": 1.8873836994171143,
+      "learning_rate": 7.768361581920904e-07,
+      "loss": 16.5183,
+      "mean_token_accuracy": 0.7111847102642059,
+      "step": 1480
+    },
+    {
+      "epoch": 0.995475113122172,
+      "grad_norm": 1.8390552997589111,
+      "learning_rate": 4.2372881355932204e-07,
+      "loss": 16.1742,
+      "mean_token_accuracy": 0.7128683432936669,
+      "step": 1485
+    },
+    {
+      "epoch": 0.9988268811798223,
+      "grad_norm": 1.8799461126327515,
+      "learning_rate": 7.062146892655368e-08,
+      "loss": 17.1633,
+      "mean_token_accuracy": 0.6963419988751411,
+      "step": 1490
+    }
+  ],
+  "logging_steps": 5,
+  "max_steps": 1491,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 750,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.5012213304045076e+19,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4615ef1249a88d7982238a4f457d81abd14a1991eab8956de34c050cdf3445ef
+size 7224

zero_to_fp32.py

@@ -0,0 +1,761 @@
+
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)