Add files using upload-large-folder tool

README.md

@@ -0,0 +1,61 @@
+---
+library_name: transformers
+license: apache-2.0
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+tags:
+- llama-factory
+- full
+- generated_from_trainer
+model-index:
+- name: Qwen2.5-1.5B-Instruct-OT3-8K-QwQ
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# Qwen2.5-1.5B-Instruct-OT3-8K-QwQ
+
+This model is a fine-tuned version of [Qwen/Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) on the ot3_8k_subset_qwq dataset.
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 1e-05
+- train_batch_size: 1
+- eval_batch_size: 8
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 2
+- gradient_accumulation_steps: 8
+- total_train_batch_size: 16
+- total_eval_batch_size: 16
+- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_ratio: 0.1
+- num_epochs: 5.0
+
+### Training results
+
+
+
+### Framework versions
+
+- Transformers 4.52.4
+- Pytorch 2.7.1+cu126
+- Datasets 3.6.0
+- Tokenizers 0.21.1

added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

all_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 5.0,
+    "total_flos": 451881746563072.0,
+    "train_loss": 1.2663256504714147,
+    "train_runtime": 105590.0324,
+    "train_samples_per_second": 0.394,
+    "train_steps_per_second": 0.025
+}

chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

checkpoint-1042/config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-1042/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.1,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "4.52.4"
+}

checkpoint-1042/latest

@@ -0,0 +1 @@
+global_step1041

checkpoint-1042/zero_to_fp32.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)

checkpoint-2084/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoint-2084/chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

checkpoint-2084/config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-2084/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.1,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "4.52.4"
+}

checkpoint-2084/latest

@@ -0,0 +1 @@
+global_step2083

checkpoint-2084/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-2084/tokenizer_config.json

@@ -0,0 +1,208 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-2084/zero_to_fp32.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)

checkpoint-2605/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoint-2605/chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

checkpoint-2605/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.1,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "4.52.4"
+}

checkpoint-2605/latest

@@ -0,0 +1 @@
+global_step2603

checkpoint-2605/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-2605/tokenizer_config.json

@@ -0,0 +1,208 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-521/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoint-521/chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

checkpoint-521/config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-521/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.1,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "4.52.4"
+}

checkpoint-521/latest

@@ -0,0 +1 @@
+global_step520

checkpoint-521/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-521/tokenizer_config.json

@@ -0,0 +1,208 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-521/trainer_state.json

@@ -0,0 +1,3681 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.0,
+  "eval_steps": 500,
+  "global_step": 521,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0019203072491598655,
+      "grad_norm": 3.050241640622783,
+      "learning_rate": 0.0,
+      "loss": 1.5054,
+      "step": 1
+    },
+    {
+      "epoch": 0.003840614498319731,
+      "grad_norm": 2.974162230197891,
+      "learning_rate": 3.831417624521073e-08,
+      "loss": 1.7399,
+      "step": 2
+    },
+    {
+      "epoch": 0.005760921747479597,
+      "grad_norm": 3.006494012208738,
+      "learning_rate": 7.662835249042146e-08,
+      "loss": 1.8793,
+      "step": 3
+    },
+    {
+      "epoch": 0.007681228996639462,
+      "grad_norm": 2.910641370990376,
+      "learning_rate": 1.1494252873563219e-07,
+      "loss": 1.7187,
+      "step": 4
+    },
+    {
+      "epoch": 0.009601536245799328,
+      "grad_norm": 2.6996013829444134,
+      "learning_rate": 1.5325670498084292e-07,
+      "loss": 1.6475,
+      "step": 5
+    },
+    {
+      "epoch": 0.011521843494959194,
+      "grad_norm": 3.0173297525530525,
+      "learning_rate": 1.9157088122605365e-07,
+      "loss": 1.6964,
+      "step": 6
+    },
+    {
+      "epoch": 0.01344215074411906,
+      "grad_norm": 3.0082364904112566,
+      "learning_rate": 2.2988505747126437e-07,
+      "loss": 1.6851,
+      "step": 7
+    },
+    {
+      "epoch": 0.015362457993278924,
+      "grad_norm": 2.880521601062656,
+      "learning_rate": 2.681992337164751e-07,
+      "loss": 1.557,
+      "step": 8
+    },
+    {
+      "epoch": 0.01728276524243879,
+      "grad_norm": 2.8379079611890896,
+      "learning_rate": 3.0651340996168583e-07,
+      "loss": 1.6569,
+      "step": 9
+    },
+    {
+      "epoch": 0.019203072491598656,
+      "grad_norm": 2.902207545955143,
+      "learning_rate": 3.4482758620689656e-07,
+      "loss": 1.637,
+      "step": 10
+    },
+    {
+      "epoch": 0.02112337974075852,
+      "grad_norm": 2.8146383036248883,
+      "learning_rate": 3.831417624521073e-07,
+      "loss": 1.7074,
+      "step": 11
+    },
+    {
+      "epoch": 0.023043686989918388,
+      "grad_norm": 2.7702439023075422,
+      "learning_rate": 4.2145593869731807e-07,
+      "loss": 1.6621,
+      "step": 12
+    },
+    {
+      "epoch": 0.024963994239078253,
+      "grad_norm": 2.948897328700205,
+      "learning_rate": 4.5977011494252875e-07,
+      "loss": 1.7141,
+      "step": 13
+    },
+    {
+      "epoch": 0.02688430148823812,
+      "grad_norm": 2.6155066202719666,
+      "learning_rate": 4.980842911877395e-07,
+      "loss": 1.6133,
+      "step": 14
+    },
+    {
+      "epoch": 0.028804608737397985,
+      "grad_norm": 2.7754744384918335,
+      "learning_rate": 5.363984674329502e-07,
+      "loss": 1.6922,
+      "step": 15
+    },
+    {
+      "epoch": 0.030724915986557848,
+      "grad_norm": 3.005540548334177,
+      "learning_rate": 5.747126436781609e-07,
+      "loss": 1.5753,
+      "step": 16
+    },
+    {
+      "epoch": 0.03264522323571772,
+      "grad_norm": 2.6675051377737486,
+      "learning_rate": 6.130268199233717e-07,
+      "loss": 1.6697,
+      "step": 17
+    },
+    {
+      "epoch": 0.03456553048487758,
+      "grad_norm": 2.7911452731594326,
+      "learning_rate": 6.513409961685824e-07,
+      "loss": 1.6166,
+      "step": 18
+    },
+    {
+      "epoch": 0.03648583773403745,
+      "grad_norm": 2.754640389635153,
+      "learning_rate": 6.896551724137931e-07,
+      "loss": 1.4773,
+      "step": 19
+    },
+    {
+      "epoch": 0.03840614498319731,
+      "grad_norm": 2.58056430792743,
+      "learning_rate": 7.27969348659004e-07,
+      "loss": 1.5533,
+      "step": 20
+    },
+    {
+      "epoch": 0.040326452232357174,
+      "grad_norm": 2.5780862888672047,
+      "learning_rate": 7.662835249042146e-07,
+      "loss": 1.5408,
+      "step": 21
+    },
+    {
+      "epoch": 0.04224675948151704,
+      "grad_norm": 2.2939155489081293,
+      "learning_rate": 8.045977011494253e-07,
+      "loss": 1.4477,
+      "step": 22
+    },
+    {
+      "epoch": 0.044167066730676906,
+      "grad_norm": 2.326783109749172,
+      "learning_rate": 8.429118773946361e-07,
+      "loss": 1.6446,
+      "step": 23
+    },
+    {
+      "epoch": 0.046087373979836775,
+      "grad_norm": 1.9784500676559427,
+      "learning_rate": 8.812260536398468e-07,
+      "loss": 1.4741,
+      "step": 24
+    },
+    {
+      "epoch": 0.04800768122899664,
+      "grad_norm": 2.158116045848602,
+      "learning_rate": 9.195402298850575e-07,
+      "loss": 1.8359,
+      "step": 25
+    },
+    {
+      "epoch": 0.04992798847815651,
+      "grad_norm": 2.035070253571019,
+      "learning_rate": 9.578544061302683e-07,
+      "loss": 1.6079,
+      "step": 26
+    },
+    {
+      "epoch": 0.05184829572731637,
+      "grad_norm": 1.9828772550714657,
+      "learning_rate": 9.96168582375479e-07,
+      "loss": 1.6275,
+      "step": 27
+    },
+    {
+      "epoch": 0.05376860297647624,
+      "grad_norm": 2.063763559262756,
+      "learning_rate": 1.0344827586206898e-06,
+      "loss": 1.6298,
+      "step": 28
+    },
+    {
+      "epoch": 0.0556889102256361,
+      "grad_norm": 1.9339129722646953,
+      "learning_rate": 1.0727969348659004e-06,
+      "loss": 1.5816,
+      "step": 29
+    },
+    {
+      "epoch": 0.05760921747479597,
+      "grad_norm": 1.916925980149754,
+      "learning_rate": 1.111111111111111e-06,
+      "loss": 1.6064,
+      "step": 30
+    },
+    {
+      "epoch": 0.05952952472395583,
+      "grad_norm": 1.6450444187274702,
+      "learning_rate": 1.1494252873563219e-06,
+      "loss": 1.5159,
+      "step": 31
+    },
+    {
+      "epoch": 0.061449831973115696,
+      "grad_norm": 1.6132237873302806,
+      "learning_rate": 1.1877394636015327e-06,
+      "loss": 1.5367,
+      "step": 32
+    },
+    {
+      "epoch": 0.06337013922227556,
+      "grad_norm": 1.818936433606168,
+      "learning_rate": 1.2260536398467433e-06,
+      "loss": 1.4623,
+      "step": 33
+    },
+    {
+      "epoch": 0.06529044647143543,
+      "grad_norm": 1.7241994573893415,
+      "learning_rate": 1.2643678160919542e-06,
+      "loss": 1.6151,
+      "step": 34
+    },
+    {
+      "epoch": 0.06721075372059529,
+      "grad_norm": 1.5649464939481088,
+      "learning_rate": 1.3026819923371648e-06,
+      "loss": 1.6588,
+      "step": 35
+    },
+    {
+      "epoch": 0.06913106096975516,
+      "grad_norm": 1.5636025180091495,
+      "learning_rate": 1.3409961685823756e-06,
+      "loss": 1.3946,
+      "step": 36
+    },
+    {
+      "epoch": 0.07105136821891503,
+      "grad_norm": 1.5461645880192942,
+      "learning_rate": 1.3793103448275862e-06,
+      "loss": 1.6768,
+      "step": 37
+    },
+    {
+      "epoch": 0.0729716754680749,
+      "grad_norm": 1.4344137662856442,
+      "learning_rate": 1.417624521072797e-06,
+      "loss": 1.4397,
+      "step": 38
+    },
+    {
+      "epoch": 0.07489198271723475,
+      "grad_norm": 1.3960130764357097,
+      "learning_rate": 1.455938697318008e-06,
+      "loss": 1.6793,
+      "step": 39
+    },
+    {
+      "epoch": 0.07681228996639462,
+      "grad_norm": 1.2396850770578887,
+      "learning_rate": 1.4942528735632185e-06,
+      "loss": 1.5595,
+      "step": 40
+    },
+    {
+      "epoch": 0.07873259721555449,
+      "grad_norm": 1.2226941074613882,
+      "learning_rate": 1.5325670498084292e-06,
+      "loss": 1.4513,
+      "step": 41
+    },
+    {
+      "epoch": 0.08065290446471435,
+      "grad_norm": 1.2420772331499856,
+      "learning_rate": 1.57088122605364e-06,
+      "loss": 1.5815,
+      "step": 42
+    },
+    {
+      "epoch": 0.08257321171387422,
+      "grad_norm": 1.1487654644464287,
+      "learning_rate": 1.6091954022988506e-06,
+      "loss": 1.5939,
+      "step": 43
+    },
+    {
+      "epoch": 0.08449351896303409,
+      "grad_norm": 1.2124391317795524,
+      "learning_rate": 1.6475095785440615e-06,
+      "loss": 1.3123,
+      "step": 44
+    },
+    {
+      "epoch": 0.08641382621219396,
+      "grad_norm": 1.2527034268005153,
+      "learning_rate": 1.6858237547892723e-06,
+      "loss": 1.6962,
+      "step": 45
+    },
+    {
+      "epoch": 0.08833413346135381,
+      "grad_norm": 1.22540150664902,
+      "learning_rate": 1.724137931034483e-06,
+      "loss": 1.5573,
+      "step": 46
+    },
+    {
+      "epoch": 0.09025444071051368,
+      "grad_norm": 1.0822330995590186,
+      "learning_rate": 1.7624521072796935e-06,
+      "loss": 1.6174,
+      "step": 47
+    },
+    {
+      "epoch": 0.09217474795967355,
+      "grad_norm": 1.3725369113494932,
+      "learning_rate": 1.8007662835249044e-06,
+      "loss": 1.5724,
+      "step": 48
+    },
+    {
+      "epoch": 0.09409505520883342,
+      "grad_norm": 1.1246522439222844,
+      "learning_rate": 1.839080459770115e-06,
+      "loss": 1.6577,
+      "step": 49
+    },
+    {
+      "epoch": 0.09601536245799328,
+      "grad_norm": 1.153435955007641,
+      "learning_rate": 1.8773946360153258e-06,
+      "loss": 1.6045,
+      "step": 50
+    },
+    {
+      "epoch": 0.09793566970715314,
+      "grad_norm": 1.1142741754051628,
+      "learning_rate": 1.9157088122605367e-06,
+      "loss": 1.6562,
+      "step": 51
+    },
+    {
+      "epoch": 0.09985597695631301,
+      "grad_norm": 1.0608003189926583,
+      "learning_rate": 1.9540229885057475e-06,
+      "loss": 1.4951,
+      "step": 52
+    },
+    {
+      "epoch": 0.10177628420547287,
+      "grad_norm": 1.0049278590265711,
+      "learning_rate": 1.992337164750958e-06,
+      "loss": 1.5432,
+      "step": 53
+    },
+    {
+      "epoch": 0.10369659145463274,
+      "grad_norm": 0.895715467779371,
+      "learning_rate": 2.0306513409961687e-06,
+      "loss": 1.4494,
+      "step": 54
+    },
+    {
+      "epoch": 0.10561689870379261,
+      "grad_norm": 0.948460307989256,
+      "learning_rate": 2.0689655172413796e-06,
+      "loss": 1.6377,
+      "step": 55
+    },
+    {
+      "epoch": 0.10753720595295248,
+      "grad_norm": 0.8960302855343928,
+      "learning_rate": 2.1072796934865904e-06,
+      "loss": 1.5476,
+      "step": 56
+    },
+    {
+      "epoch": 0.10945751320211233,
+      "grad_norm": 0.8555220673502346,
+      "learning_rate": 2.145593869731801e-06,
+      "loss": 1.2937,
+      "step": 57
+    },
+    {
+      "epoch": 0.1113778204512722,
+      "grad_norm": 0.9297918052847216,
+      "learning_rate": 2.1839080459770117e-06,
+      "loss": 1.5018,
+      "step": 58
+    },
+    {
+      "epoch": 0.11329812770043207,
+      "grad_norm": 0.8615767845690797,
+      "learning_rate": 2.222222222222222e-06,
+      "loss": 1.4411,
+      "step": 59
+    },
+    {
+      "epoch": 0.11521843494959194,
+      "grad_norm": 0.9800340953282465,
+      "learning_rate": 2.260536398467433e-06,
+      "loss": 1.6592,
+      "step": 60
+    },
+    {
+      "epoch": 0.1171387421987518,
+      "grad_norm": 0.8319044483038768,
+      "learning_rate": 2.2988505747126437e-06,
+      "loss": 1.4999,
+      "step": 61
+    },
+    {
+      "epoch": 0.11905904944791167,
+      "grad_norm": 0.8113904065691636,
+      "learning_rate": 2.3371647509578546e-06,
+      "loss": 1.4058,
+      "step": 62
+    },
+    {
+      "epoch": 0.12097935669707154,
+      "grad_norm": 0.7584474578599453,
+      "learning_rate": 2.3754789272030654e-06,
+      "loss": 1.1995,
+      "step": 63
+    },
+    {
+      "epoch": 0.12289966394623139,
+      "grad_norm": 0.8112593744821786,
+      "learning_rate": 2.4137931034482762e-06,
+      "loss": 1.622,
+      "step": 64
+    },
+    {
+      "epoch": 0.12481997119539126,
+      "grad_norm": 0.8465973433211744,
+      "learning_rate": 2.4521072796934867e-06,
+      "loss": 1.5474,
+      "step": 65
+    },
+    {
+      "epoch": 0.12674027844455113,
+      "grad_norm": 0.829645722256521,
+      "learning_rate": 2.4904214559386975e-06,
+      "loss": 1.485,
+      "step": 66
+    },
+    {
+      "epoch": 0.128660585693711,
+      "grad_norm": 0.8480123859041653,
+      "learning_rate": 2.5287356321839083e-06,
+      "loss": 1.5641,
+      "step": 67
+    },
+    {
+      "epoch": 0.13058089294287087,
+      "grad_norm": 0.7945937151249043,
+      "learning_rate": 2.567049808429119e-06,
+      "loss": 1.3511,
+      "step": 68
+    },
+    {
+      "epoch": 0.13250120019203074,
+      "grad_norm": 0.8452583555772669,
+      "learning_rate": 2.6053639846743296e-06,
+      "loss": 1.4936,
+      "step": 69
+    },
+    {
+      "epoch": 0.13442150744119058,
+      "grad_norm": 0.7913167472525273,
+      "learning_rate": 2.6436781609195404e-06,
+      "loss": 1.524,
+      "step": 70
+    },
+    {
+      "epoch": 0.13634181469035045,
+      "grad_norm": 0.7632883226077274,
+      "learning_rate": 2.6819923371647512e-06,
+      "loss": 1.4957,
+      "step": 71
+    },
+    {
+      "epoch": 0.13826212193951032,
+      "grad_norm": 0.8040467176753551,
+      "learning_rate": 2.720306513409962e-06,
+      "loss": 1.5472,
+      "step": 72
+    },
+    {
+      "epoch": 0.1401824291886702,
+      "grad_norm": 0.8063495654842968,
+      "learning_rate": 2.7586206896551725e-06,
+      "loss": 1.4355,
+      "step": 73
+    },
+    {
+      "epoch": 0.14210273643783006,
+      "grad_norm": 0.8632589003342163,
+      "learning_rate": 2.796934865900383e-06,
+      "loss": 1.6251,
+      "step": 74
+    },
+    {
+      "epoch": 0.14402304368698993,
+      "grad_norm": 0.9065110727033228,
+      "learning_rate": 2.835249042145594e-06,
+      "loss": 1.3482,
+      "step": 75
+    },
+    {
+      "epoch": 0.1459433509361498,
+      "grad_norm": 0.7358621824934779,
+      "learning_rate": 2.8735632183908046e-06,
+      "loss": 1.4579,
+      "step": 76
+    },
+    {
+      "epoch": 0.14786365818530964,
+      "grad_norm": 0.7882740690499183,
+      "learning_rate": 2.911877394636016e-06,
+      "loss": 1.584,
+      "step": 77
+    },
+    {
+      "epoch": 0.1497839654344695,
+      "grad_norm": 0.8028390602285591,
+      "learning_rate": 2.9501915708812262e-06,
+      "loss": 1.5799,
+      "step": 78
+    },
+    {
+      "epoch": 0.15170427268362938,
+      "grad_norm": 0.7087212262202383,
+      "learning_rate": 2.988505747126437e-06,
+      "loss": 1.3796,
+      "step": 79
+    },
+    {
+      "epoch": 0.15362457993278925,
+      "grad_norm": 0.7477246745009772,
+      "learning_rate": 3.026819923371648e-06,
+      "loss": 1.4577,
+      "step": 80
+    },
+    {
+      "epoch": 0.15554488718194912,
+      "grad_norm": 0.7362367028641865,
+      "learning_rate": 3.0651340996168583e-06,
+      "loss": 1.3986,
+      "step": 81
+    },
+    {
+      "epoch": 0.15746519443110898,
+      "grad_norm": 0.7542749931181064,
+      "learning_rate": 3.103448275862069e-06,
+      "loss": 1.3923,
+      "step": 82
+    },
+    {
+      "epoch": 0.15938550168026885,
+      "grad_norm": 0.7591236429919996,
+      "learning_rate": 3.14176245210728e-06,
+      "loss": 1.468,
+      "step": 83
+    },
+    {
+      "epoch": 0.1613058089294287,
+      "grad_norm": 0.750073455610315,
+      "learning_rate": 3.180076628352491e-06,
+      "loss": 1.5728,
+      "step": 84
+    },
+    {
+      "epoch": 0.16322611617858857,
+      "grad_norm": 0.7463483577071213,
+      "learning_rate": 3.2183908045977012e-06,
+      "loss": 1.5751,
+      "step": 85
+    },
+    {
+      "epoch": 0.16514642342774843,
+      "grad_norm": 0.7459336348523459,
+      "learning_rate": 3.256704980842912e-06,
+      "loss": 1.5821,
+      "step": 86
+    },
+    {
+      "epoch": 0.1670667306769083,
+      "grad_norm": 0.7503745075545026,
+      "learning_rate": 3.295019157088123e-06,
+      "loss": 1.3811,
+      "step": 87
+    },
+    {
+      "epoch": 0.16898703792606817,
+      "grad_norm": 0.6781019129991772,
+      "learning_rate": 3.3333333333333333e-06,
+      "loss": 1.4967,
+      "step": 88
+    },
+    {
+      "epoch": 0.17090734517522804,
+      "grad_norm": 0.6772160890677477,
+      "learning_rate": 3.3716475095785446e-06,
+      "loss": 1.4602,
+      "step": 89
+    },
+    {
+      "epoch": 0.1728276524243879,
+      "grad_norm": 0.7414559131468873,
+      "learning_rate": 3.409961685823755e-06,
+      "loss": 1.5384,
+      "step": 90
+    },
+    {
+      "epoch": 0.17474795967354778,
+      "grad_norm": 0.7339713894607118,
+      "learning_rate": 3.448275862068966e-06,
+      "loss": 1.3478,
+      "step": 91
+    },
+    {
+      "epoch": 0.17666826692270762,
+      "grad_norm": 0.6741291120903733,
+      "learning_rate": 3.4865900383141767e-06,
+      "loss": 1.3051,
+      "step": 92
+    },
+    {
+      "epoch": 0.1785885741718675,
+      "grad_norm": 0.8466746593898973,
+      "learning_rate": 3.524904214559387e-06,
+      "loss": 1.6414,
+      "step": 93
+    },
+    {
+      "epoch": 0.18050888142102736,
+      "grad_norm": 0.6235998792480368,
+      "learning_rate": 3.563218390804598e-06,
+      "loss": 1.3774,
+      "step": 94
+    },
+    {
+      "epoch": 0.18242918867018723,
+      "grad_norm": 0.7888875616341429,
+      "learning_rate": 3.6015325670498087e-06,
+      "loss": 1.2273,
+      "step": 95
+    },
+    {
+      "epoch": 0.1843494959193471,
+      "grad_norm": 0.7184060206606023,
+      "learning_rate": 3.6398467432950196e-06,
+      "loss": 1.4414,
+      "step": 96
+    },
+    {
+      "epoch": 0.18626980316850697,
+      "grad_norm": 0.6931282624303644,
+      "learning_rate": 3.67816091954023e-06,
+      "loss": 1.1996,
+      "step": 97
+    },
+    {
+      "epoch": 0.18819011041766684,
+      "grad_norm": 0.6898899817332669,
+      "learning_rate": 3.7164750957854412e-06,
+      "loss": 1.153,
+      "step": 98
+    },
+    {
+      "epoch": 0.19011041766682668,
+      "grad_norm": 0.6706065855351341,
+      "learning_rate": 3.7547892720306517e-06,
+      "loss": 1.3219,
+      "step": 99
+    },
+    {
+      "epoch": 0.19203072491598655,
+      "grad_norm": 0.6618017863324709,
+      "learning_rate": 3.793103448275862e-06,
+      "loss": 1.3694,
+      "step": 100
+    },
+    {
+      "epoch": 0.19395103216514642,
+      "grad_norm": 0.7084267342732283,
+      "learning_rate": 3.831417624521073e-06,
+      "loss": 1.5242,
+      "step": 101
+    },
+    {
+      "epoch": 0.1958713394143063,
+      "grad_norm": 0.726545416016091,
+      "learning_rate": 3.869731800766283e-06,
+      "loss": 1.4521,
+      "step": 102
+    },
+    {
+      "epoch": 0.19779164666346616,
+      "grad_norm": 0.7462891561612502,
+      "learning_rate": 3.908045977011495e-06,
+      "loss": 1.4687,
+      "step": 103
+    },
+    {
+      "epoch": 0.19971195391262603,
+      "grad_norm": 0.7186363274122183,
+      "learning_rate": 3.946360153256705e-06,
+      "loss": 1.4222,
+      "step": 104
+    },
+    {
+      "epoch": 0.2016322611617859,
+      "grad_norm": 0.6562522942973675,
+      "learning_rate": 3.984674329501916e-06,
+      "loss": 1.3302,
+      "step": 105
+    },
+    {
+      "epoch": 0.20355256841094574,
+      "grad_norm": 0.7593764693035522,
+      "learning_rate": 4.022988505747127e-06,
+      "loss": 1.4029,
+      "step": 106
+    },
+    {
+      "epoch": 0.2054728756601056,
+      "grad_norm": 0.674187794860422,
+      "learning_rate": 4.0613026819923375e-06,
+      "loss": 1.2851,
+      "step": 107
+    },
+    {
+      "epoch": 0.20739318290926548,
+      "grad_norm": 0.6771453444719904,
+      "learning_rate": 4.099616858237548e-06,
+      "loss": 1.3873,
+      "step": 108
+    },
+    {
+      "epoch": 0.20931349015842535,
+      "grad_norm": 0.6904184501885661,
+      "learning_rate": 4.137931034482759e-06,
+      "loss": 1.4671,
+      "step": 109
+    },
+    {
+      "epoch": 0.21123379740758522,
+      "grad_norm": 0.7386010515763627,
+      "learning_rate": 4.17624521072797e-06,
+      "loss": 1.4114,
+      "step": 110
+    },
+    {
+      "epoch": 0.21315410465674509,
+      "grad_norm": 0.6330108602988729,
+      "learning_rate": 4.214559386973181e-06,
+      "loss": 1.5344,
+      "step": 111
+    },
+    {
+      "epoch": 0.21507441190590496,
+      "grad_norm": 0.6476316086673372,
+      "learning_rate": 4.252873563218391e-06,
+      "loss": 1.4134,
+      "step": 112
+    },
+    {
+      "epoch": 0.2169947191550648,
+      "grad_norm": 0.7369608607793586,
+      "learning_rate": 4.291187739463602e-06,
+      "loss": 1.4656,
+      "step": 113
+    },
+    {
+      "epoch": 0.21891502640422467,
+      "grad_norm": 0.6991819737178624,
+      "learning_rate": 4.3295019157088125e-06,
+      "loss": 1.3801,
+      "step": 114
+    },
+    {
+      "epoch": 0.22083533365338454,
+      "grad_norm": 0.6670879445681203,
+      "learning_rate": 4.367816091954023e-06,
+      "loss": 1.4012,
+      "step": 115
+    },
+    {
+      "epoch": 0.2227556409025444,
+      "grad_norm": 0.7091321651558993,
+      "learning_rate": 4.406130268199234e-06,
+      "loss": 1.369,
+      "step": 116
+    },
+    {
+      "epoch": 0.22467594815170427,
+      "grad_norm": 0.6538069130880931,
+      "learning_rate": 4.444444444444444e-06,
+      "loss": 1.207,
+      "step": 117
+    },
+    {
+      "epoch": 0.22659625540086414,
+      "grad_norm": 0.7414604205506811,
+      "learning_rate": 4.482758620689656e-06,
+      "loss": 1.4466,
+      "step": 118
+    },
+    {
+      "epoch": 0.228516562650024,
+      "grad_norm": 0.6754242592483168,
+      "learning_rate": 4.521072796934866e-06,
+      "loss": 1.345,
+      "step": 119
+    },
+    {
+      "epoch": 0.23043686989918388,
+      "grad_norm": 0.7890890418316723,
+      "learning_rate": 4.5593869731800775e-06,
+      "loss": 1.4887,
+      "step": 120
+    },
+    {
+      "epoch": 0.23235717714834372,
+      "grad_norm": 0.7142672200972927,
+      "learning_rate": 4.5977011494252875e-06,
+      "loss": 1.5663,
+      "step": 121
+    },
+    {
+      "epoch": 0.2342774843975036,
+      "grad_norm": 0.6607576396465201,
+      "learning_rate": 4.636015325670498e-06,
+      "loss": 1.5097,
+      "step": 122
+    },
+    {
+      "epoch": 0.23619779164666346,
+      "grad_norm": 0.7801245141924045,
+      "learning_rate": 4.674329501915709e-06,
+      "loss": 1.5577,
+      "step": 123
+    },
+    {
+      "epoch": 0.23811809889582333,
+      "grad_norm": 0.7853237789383211,
+      "learning_rate": 4.71264367816092e-06,
+      "loss": 1.5542,
+      "step": 124
+    },
+    {
+      "epoch": 0.2400384061449832,
+      "grad_norm": 0.6731916164633106,
+      "learning_rate": 4.750957854406131e-06,
+      "loss": 1.5991,
+      "step": 125
+    },
+    {
+      "epoch": 0.24195871339414307,
+      "grad_norm": 0.7081285505123779,
+      "learning_rate": 4.789272030651342e-06,
+      "loss": 1.4219,
+      "step": 126
+    },
+    {
+      "epoch": 0.24387902064330294,
+      "grad_norm": 0.6523124091427233,
+      "learning_rate": 4.8275862068965525e-06,
+      "loss": 1.3967,
+      "step": 127
+    },
+    {
+      "epoch": 0.24579932789246278,
+      "grad_norm": 0.7010318704056578,
+      "learning_rate": 4.8659003831417625e-06,
+      "loss": 1.4079,
+      "step": 128
+    },
+    {
+      "epoch": 0.24771963514162265,
+      "grad_norm": 0.7310188880844346,
+      "learning_rate": 4.904214559386973e-06,
+      "loss": 1.5645,
+      "step": 129
+    },
+    {
+      "epoch": 0.24963994239078252,
+      "grad_norm": 0.7370196750980766,
+      "learning_rate": 4.942528735632184e-06,
+      "loss": 1.3163,
+      "step": 130
+    },
+    {
+      "epoch": 0.2515602496399424,
+      "grad_norm": 0.5988613538056734,
+      "learning_rate": 4.980842911877395e-06,
+      "loss": 1.2164,
+      "step": 131
+    },
+    {
+      "epoch": 0.25348055688910226,
+      "grad_norm": 0.6686452758995931,
+      "learning_rate": 5.019157088122606e-06,
+      "loss": 1.3216,
+      "step": 132
+    },
+    {
+      "epoch": 0.25540086413826213,
+      "grad_norm": 0.6283032311907456,
+      "learning_rate": 5.057471264367817e-06,
+      "loss": 1.3004,
+      "step": 133
+    },
+    {
+      "epoch": 0.257321171387422,
+      "grad_norm": 0.7131274515928692,
+      "learning_rate": 5.095785440613027e-06,
+      "loss": 1.3574,
+      "step": 134
+    },
+    {
+      "epoch": 0.25924147863658187,
+      "grad_norm": 0.7562023685666645,
+      "learning_rate": 5.134099616858238e-06,
+      "loss": 1.4755,
+      "step": 135
+    },
+    {
+      "epoch": 0.26116178588574174,
+      "grad_norm": 0.7260126586378673,
+      "learning_rate": 5.172413793103449e-06,
+      "loss": 1.3512,
+      "step": 136
+    },
+    {
+      "epoch": 0.2630820931349016,
+      "grad_norm": 0.7027786093003081,
+      "learning_rate": 5.210727969348659e-06,
+      "loss": 1.3988,
+      "step": 137
+    },
+    {
+      "epoch": 0.2650024003840615,
+      "grad_norm": 0.6536542941035637,
+      "learning_rate": 5.24904214559387e-06,
+      "loss": 1.3403,
+      "step": 138
+    },
+    {
+      "epoch": 0.2669227076332213,
+      "grad_norm": 0.6887374540995002,
+      "learning_rate": 5.287356321839081e-06,
+      "loss": 1.4009,
+      "step": 139
+    },
+    {
+      "epoch": 0.26884301488238116,
+      "grad_norm": 0.6904650924816887,
+      "learning_rate": 5.3256704980842925e-06,
+      "loss": 1.3674,
+      "step": 140
+    },
+    {
+      "epoch": 0.27076332213154103,
+      "grad_norm": 0.7396766783545521,
+      "learning_rate": 5.3639846743295025e-06,
+      "loss": 1.2224,
+      "step": 141
+    },
+    {
+      "epoch": 0.2726836293807009,
+      "grad_norm": 0.6618698222762107,
+      "learning_rate": 5.402298850574713e-06,
+      "loss": 1.3261,
+      "step": 142
+    },
+    {
+      "epoch": 0.27460393662986077,
+      "grad_norm": 0.7346869512327489,
+      "learning_rate": 5.440613026819924e-06,
+      "loss": 1.5539,
+      "step": 143
+    },
+    {
+      "epoch": 0.27652424387902064,
+      "grad_norm": 0.6542272405748857,
+      "learning_rate": 5.478927203065134e-06,
+      "loss": 1.2301,
+      "step": 144
+    },
+    {
+      "epoch": 0.2784445511281805,
+      "grad_norm": 0.652844033827654,
+      "learning_rate": 5.517241379310345e-06,
+      "loss": 1.3203,
+      "step": 145
+    },
+    {
+      "epoch": 0.2803648583773404,
+      "grad_norm": 0.7093255205591261,
+      "learning_rate": 5.555555555555557e-06,
+      "loss": 1.4194,
+      "step": 146
+    },
+    {
+      "epoch": 0.28228516562650025,
+      "grad_norm": 0.7713263587807493,
+      "learning_rate": 5.593869731800766e-06,
+      "loss": 1.3986,
+      "step": 147
+    },
+    {
+      "epoch": 0.2842054728756601,
+      "grad_norm": 0.7273045777114594,
+      "learning_rate": 5.6321839080459775e-06,
+      "loss": 1.5226,
+      "step": 148
+    },
+    {
+      "epoch": 0.28612578012482,
+      "grad_norm": 0.6416135588874926,
+      "learning_rate": 5.670498084291188e-06,
+      "loss": 1.4148,
+      "step": 149
+    },
+    {
+      "epoch": 0.28804608737397985,
+      "grad_norm": 0.6767941047038493,
+      "learning_rate": 5.708812260536399e-06,
+      "loss": 1.4741,
+      "step": 150
+    },
+    {
+      "epoch": 0.2899663946231397,
+      "grad_norm": 0.7125976753134923,
+      "learning_rate": 5.747126436781609e-06,
+      "loss": 1.4689,
+      "step": 151
+    },
+    {
+      "epoch": 0.2918867018722996,
+      "grad_norm": 0.6848434615690435,
+      "learning_rate": 5.78544061302682e-06,
+      "loss": 1.3208,
+      "step": 152
+    },
+    {
+      "epoch": 0.2938070091214594,
+      "grad_norm": 0.7637154008633908,
+      "learning_rate": 5.823754789272032e-06,
+      "loss": 1.5237,
+      "step": 153
+    },
+    {
+      "epoch": 0.2957273163706193,
+      "grad_norm": 0.6483331857341429,
+      "learning_rate": 5.862068965517242e-06,
+      "loss": 1.3073,
+      "step": 154
+    },
+    {
+      "epoch": 0.29764762361977914,
+      "grad_norm": 0.684818235289912,
+      "learning_rate": 5.9003831417624525e-06,
+      "loss": 1.4076,
+      "step": 155
+    },
+    {
+      "epoch": 0.299567930868939,
+      "grad_norm": 0.6683114123494819,
+      "learning_rate": 5.938697318007663e-06,
+      "loss": 1.3518,
+      "step": 156
+    },
+    {
+      "epoch": 0.3014882381180989,
+      "grad_norm": 0.8301668940114854,
+      "learning_rate": 5.977011494252874e-06,
+      "loss": 1.5087,
+      "step": 157
+    },
+    {
+      "epoch": 0.30340854536725875,
+      "grad_norm": 0.7306107739349151,
+      "learning_rate": 6.015325670498084e-06,
+      "loss": 1.4187,
+      "step": 158
+    },
+    {
+      "epoch": 0.3053288526164186,
+      "grad_norm": 0.692887187718929,
+      "learning_rate": 6.053639846743296e-06,
+      "loss": 1.5365,
+      "step": 159
+    },
+    {
+      "epoch": 0.3072491598655785,
+      "grad_norm": 0.6502853635559999,
+      "learning_rate": 6.091954022988507e-06,
+      "loss": 1.3497,
+      "step": 160
+    },
+    {
+      "epoch": 0.30916946711473836,
+      "grad_norm": 0.7309674020853324,
+      "learning_rate": 6.130268199233717e-06,
+      "loss": 1.4965,
+      "step": 161
+    },
+    {
+      "epoch": 0.31108977436389823,
+      "grad_norm": 0.7124865800378807,
+      "learning_rate": 6.1685823754789275e-06,
+      "loss": 1.4449,
+      "step": 162
+    },
+    {
+      "epoch": 0.3130100816130581,
+      "grad_norm": 0.692769353628125,
+      "learning_rate": 6.206896551724138e-06,
+      "loss": 1.2823,
+      "step": 163
+    },
+    {
+      "epoch": 0.31493038886221797,
+      "grad_norm": 0.8120061941639997,
+      "learning_rate": 6.24521072796935e-06,
+      "loss": 1.5322,
+      "step": 164
+    },
+    {
+      "epoch": 0.31685069611137784,
+      "grad_norm": 0.6506906315887734,
+      "learning_rate": 6.28352490421456e-06,
+      "loss": 1.4138,
+      "step": 165
+    },
+    {
+      "epoch": 0.3187710033605377,
+      "grad_norm": 0.7021274588006857,
+      "learning_rate": 6.321839080459771e-06,
+      "loss": 1.5534,
+      "step": 166
+    },
+    {
+      "epoch": 0.3206913106096976,
+      "grad_norm": 0.6965648094070332,
+      "learning_rate": 6.360153256704982e-06,
+      "loss": 1.3055,
+      "step": 167
+    },
+    {
+      "epoch": 0.3226116178588574,
+      "grad_norm": 0.7498732720147595,
+      "learning_rate": 6.398467432950192e-06,
+      "loss": 1.2775,
+      "step": 168
+    },
+    {
+      "epoch": 0.32453192510801726,
+      "grad_norm": 0.7097564261194121,
+      "learning_rate": 6.4367816091954025e-06,
+      "loss": 1.3029,
+      "step": 169
+    },
+    {
+      "epoch": 0.32645223235717713,
+      "grad_norm": 0.6382917432482148,
+      "learning_rate": 6.475095785440614e-06,
+      "loss": 1.3569,
+      "step": 170
+    },
+    {
+      "epoch": 0.328372539606337,
+      "grad_norm": 0.6745359700702276,
+      "learning_rate": 6.513409961685824e-06,
+      "loss": 1.3339,
+      "step": 171
+    },
+    {
+      "epoch": 0.33029284685549687,
+      "grad_norm": 0.8056549461489378,
+      "learning_rate": 6.551724137931035e-06,
+      "loss": 1.4649,
+      "step": 172
+    },
+    {
+      "epoch": 0.33221315410465674,
+      "grad_norm": 0.7305964712888819,
+      "learning_rate": 6.590038314176246e-06,
+      "loss": 1.4451,
+      "step": 173
+    },
+    {
+      "epoch": 0.3341334613538166,
+      "grad_norm": 0.6201165387321146,
+      "learning_rate": 6.628352490421457e-06,
+      "loss": 1.3352,
+      "step": 174
+    },
+    {
+      "epoch": 0.3360537686029765,
+      "grad_norm": 0.8265036713279577,
+      "learning_rate": 6.666666666666667e-06,
+      "loss": 1.4794,
+      "step": 175
+    },
+    {
+      "epoch": 0.33797407585213635,
+      "grad_norm": 0.6893681404195715,
+      "learning_rate": 6.7049808429118775e-06,
+      "loss": 1.4419,
+      "step": 176
+    },
+    {
+      "epoch": 0.3398943831012962,
+      "grad_norm": 0.7509743815259918,
+      "learning_rate": 6.743295019157089e-06,
+      "loss": 1.3666,
+      "step": 177
+    },
+    {
+      "epoch": 0.3418146903504561,
+      "grad_norm": 0.7122960885573658,
+      "learning_rate": 6.781609195402299e-06,
+      "loss": 1.4108,
+      "step": 178
+    },
+    {
+      "epoch": 0.34373499759961595,
+      "grad_norm": 0.848930385570575,
+      "learning_rate": 6.81992337164751e-06,
+      "loss": 1.4827,
+      "step": 179
+    },
+    {
+      "epoch": 0.3456553048487758,
+      "grad_norm": 0.7602028733722221,
+      "learning_rate": 6.858237547892721e-06,
+      "loss": 1.3414,
+      "step": 180
+    },
+    {
+      "epoch": 0.3475756120979357,
+      "grad_norm": 0.8288114979771807,
+      "learning_rate": 6.896551724137932e-06,
+      "loss": 1.2727,
+      "step": 181
+    },
+    {
+      "epoch": 0.34949591934709556,
+      "grad_norm": 0.7693760683257455,
+      "learning_rate": 6.934865900383142e-06,
+      "loss": 1.3706,
+      "step": 182
+    },
+    {
+      "epoch": 0.3514162265962554,
+      "grad_norm": 0.7808563776887848,
+      "learning_rate": 6.973180076628353e-06,
+      "loss": 1.4723,
+      "step": 183
+    },
+    {
+      "epoch": 0.35333653384541525,
+      "grad_norm": 0.8734769267258163,
+      "learning_rate": 7.011494252873564e-06,
+      "loss": 1.2891,
+      "step": 184
+    },
+    {
+      "epoch": 0.3552568410945751,
+      "grad_norm": 0.8867229876837837,
+      "learning_rate": 7.049808429118774e-06,
+      "loss": 1.4359,
+      "step": 185
+    },
+    {
+      "epoch": 0.357177148343735,
+      "grad_norm": 0.8349083935516824,
+      "learning_rate": 7.088122605363985e-06,
+      "loss": 1.5565,
+      "step": 186
+    },
+    {
+      "epoch": 0.35909745559289485,
+      "grad_norm": 0.7400284977493244,
+      "learning_rate": 7.126436781609196e-06,
+      "loss": 1.456,
+      "step": 187
+    },
+    {
+      "epoch": 0.3610177628420547,
+      "grad_norm": 0.6439707610744461,
+      "learning_rate": 7.1647509578544075e-06,
+      "loss": 1.2718,
+      "step": 188
+    },
+    {
+      "epoch": 0.3629380700912146,
+      "grad_norm": 0.757428349385491,
+      "learning_rate": 7.2030651340996175e-06,
+      "loss": 1.3176,
+      "step": 189
+    },
+    {
+      "epoch": 0.36485837734037446,
+      "grad_norm": 0.6979336476743889,
+      "learning_rate": 7.241379310344828e-06,
+      "loss": 1.1578,
+      "step": 190
+    },
+    {
+      "epoch": 0.36677868458953433,
+      "grad_norm": 0.7318936685975646,
+      "learning_rate": 7.279693486590039e-06,
+      "loss": 1.5306,
+      "step": 191
+    },
+    {
+      "epoch": 0.3686989918386942,
+      "grad_norm": 0.6968684758048505,
+      "learning_rate": 7.318007662835249e-06,
+      "loss": 1.3176,
+      "step": 192
+    },
+    {
+      "epoch": 0.37061929908785407,
+      "grad_norm": 0.8360636142316585,
+      "learning_rate": 7.35632183908046e-06,
+      "loss": 1.6162,
+      "step": 193
+    },
+    {
+      "epoch": 0.37253960633701394,
+      "grad_norm": 0.6842163000752954,
+      "learning_rate": 7.394636015325672e-06,
+      "loss": 1.3775,
+      "step": 194
+    },
+    {
+      "epoch": 0.3744599135861738,
+      "grad_norm": 0.7405931672623491,
+      "learning_rate": 7.4329501915708825e-06,
+      "loss": 1.4155,
+      "step": 195
+    },
+    {
+      "epoch": 0.3763802208353337,
+      "grad_norm": 0.6724666872249973,
+      "learning_rate": 7.4712643678160925e-06,
+      "loss": 1.6041,
+      "step": 196
+    },
+    {
+      "epoch": 0.3783005280844935,
+      "grad_norm": 0.6600816017352631,
+      "learning_rate": 7.509578544061303e-06,
+      "loss": 1.3836,
+      "step": 197
+    },
+    {
+      "epoch": 0.38022083533365336,
+      "grad_norm": 0.7294649653837092,
+      "learning_rate": 7.547892720306514e-06,
+      "loss": 1.3381,
+      "step": 198
+    },
+    {
+      "epoch": 0.38214114258281323,
+      "grad_norm": 0.6849006152291969,
+      "learning_rate": 7.586206896551724e-06,
+      "loss": 1.3448,
+      "step": 199
+    },
+    {
+      "epoch": 0.3840614498319731,
+      "grad_norm": 0.6413698682933802,
+      "learning_rate": 7.624521072796936e-06,
+      "loss": 1.1691,
+      "step": 200
+    },
+    {
+      "epoch": 0.38598175708113297,
+      "grad_norm": 0.6761946714227771,
+      "learning_rate": 7.662835249042147e-06,
+      "loss": 1.2401,
+      "step": 201
+    },
+    {
+      "epoch": 0.38790206433029284,
+      "grad_norm": 0.6604987264007839,
+      "learning_rate": 7.701149425287356e-06,
+      "loss": 1.4297,
+      "step": 202
+    },
+    {
+      "epoch": 0.3898223715794527,
+      "grad_norm": 0.6655444711020334,
+      "learning_rate": 7.739463601532567e-06,
+      "loss": 1.2069,
+      "step": 203
+    },
+    {
+      "epoch": 0.3917426788286126,
+      "grad_norm": 0.716356344463303,
+      "learning_rate": 7.77777777777778e-06,
+      "loss": 1.3456,
+      "step": 204
+    },
+    {
+      "epoch": 0.39366298607777245,
+      "grad_norm": 0.7371288713977551,
+      "learning_rate": 7.81609195402299e-06,
+      "loss": 1.3331,
+      "step": 205
+    },
+    {
+      "epoch": 0.3955832933269323,
+      "grad_norm": 0.6679678204725035,
+      "learning_rate": 7.854406130268199e-06,
+      "loss": 1.4027,
+      "step": 206
+    },
+    {
+      "epoch": 0.3975036005760922,
+      "grad_norm": 0.6255494545199866,
+      "learning_rate": 7.89272030651341e-06,
+      "loss": 1.4204,
+      "step": 207
+    },
+    {
+      "epoch": 0.39942390782525206,
+      "grad_norm": 0.7776077277070255,
+      "learning_rate": 7.93103448275862e-06,
+      "loss": 1.4262,
+      "step": 208
+    },
+    {
+      "epoch": 0.4013442150744119,
+      "grad_norm": 0.7495735486802019,
+      "learning_rate": 7.969348659003832e-06,
+      "loss": 1.3803,
+      "step": 209
+    },
+    {
+      "epoch": 0.4032645223235718,
+      "grad_norm": 0.7359991087372613,
+      "learning_rate": 8.007662835249042e-06,
+      "loss": 1.3407,
+      "step": 210
+    },
+    {
+      "epoch": 0.40518482957273166,
+      "grad_norm": 0.74502169700267,
+      "learning_rate": 8.045977011494253e-06,
+      "loss": 1.2109,
+      "step": 211
+    },
+    {
+      "epoch": 0.4071051368218915,
+      "grad_norm": 0.672919730166899,
+      "learning_rate": 8.084291187739464e-06,
+      "loss": 1.4416,
+      "step": 212
+    },
+    {
+      "epoch": 0.40902544407105135,
+      "grad_norm": 0.9039099669663537,
+      "learning_rate": 8.122605363984675e-06,
+      "loss": 1.5869,
+      "step": 213
+    },
+    {
+      "epoch": 0.4109457513202112,
+      "grad_norm": 0.609809507238103,
+      "learning_rate": 8.160919540229886e-06,
+      "loss": 1.392,
+      "step": 214
+    },
+    {
+      "epoch": 0.4128660585693711,
+      "grad_norm": 0.7209905089139385,
+      "learning_rate": 8.199233716475097e-06,
+      "loss": 1.5286,
+      "step": 215
+    },
+    {
+      "epoch": 0.41478636581853096,
+      "grad_norm": 0.7641098257228572,
+      "learning_rate": 8.237547892720307e-06,
+      "loss": 1.4556,
+      "step": 216
+    },
+    {
+      "epoch": 0.4167066730676908,
+      "grad_norm": 0.6671329519012665,
+      "learning_rate": 8.275862068965518e-06,
+      "loss": 1.4011,
+      "step": 217
+    },
+    {
+      "epoch": 0.4186269803168507,
+      "grad_norm": 0.7571002797488336,
+      "learning_rate": 8.31417624521073e-06,
+      "loss": 1.3907,
+      "step": 218
+    },
+    {
+      "epoch": 0.42054728756601056,
+      "grad_norm": 0.6765159050714699,
+      "learning_rate": 8.35249042145594e-06,
+      "loss": 1.411,
+      "step": 219
+    },
+    {
+      "epoch": 0.42246759481517043,
+      "grad_norm": 0.7631812758174655,
+      "learning_rate": 8.390804597701149e-06,
+      "loss": 1.4568,
+      "step": 220
+    },
+    {
+      "epoch": 0.4243879020643303,
+      "grad_norm": 0.819638920429698,
+      "learning_rate": 8.429118773946362e-06,
+      "loss": 1.2607,
+      "step": 221
+    },
+    {
+      "epoch": 0.42630820931349017,
+      "grad_norm": 0.6856567916211197,
+      "learning_rate": 8.467432950191573e-06,
+      "loss": 1.3665,
+      "step": 222
+    },
+    {
+      "epoch": 0.42822851656265004,
+      "grad_norm": 0.756851173923122,
+      "learning_rate": 8.505747126436782e-06,
+      "loss": 1.4053,
+      "step": 223
+    },
+    {
+      "epoch": 0.4301488238118099,
+      "grad_norm": 0.7234111087737882,
+      "learning_rate": 8.544061302681992e-06,
+      "loss": 1.2986,
+      "step": 224
+    },
+    {
+      "epoch": 0.4320691310609698,
+      "grad_norm": 0.7822272702322975,
+      "learning_rate": 8.582375478927203e-06,
+      "loss": 1.4298,
+      "step": 225
+    },
+    {
+      "epoch": 0.4339894383101296,
+      "grad_norm": 0.6810248686630753,
+      "learning_rate": 8.620689655172414e-06,
+      "loss": 1.3399,
+      "step": 226
+    },
+    {
+      "epoch": 0.43590974555928946,
+      "grad_norm": 0.8051128970700839,
+      "learning_rate": 8.659003831417625e-06,
+      "loss": 1.5267,
+      "step": 227
+    },
+    {
+      "epoch": 0.43783005280844933,
+      "grad_norm": 0.732357170450308,
+      "learning_rate": 8.697318007662836e-06,
+      "loss": 1.4766,
+      "step": 228
+    },
+    {
+      "epoch": 0.4397503600576092,
+      "grad_norm": 0.8096174835706853,
+      "learning_rate": 8.735632183908047e-06,
+      "loss": 1.5061,
+      "step": 229
+    },
+    {
+      "epoch": 0.44167066730676907,
+      "grad_norm": 0.7498363788557091,
+      "learning_rate": 8.773946360153257e-06,
+      "loss": 1.5328,
+      "step": 230
+    },
+    {
+      "epoch": 0.44359097455592894,
+      "grad_norm": 0.6639308793432624,
+      "learning_rate": 8.812260536398468e-06,
+      "loss": 1.5014,
+      "step": 231
+    },
+    {
+      "epoch": 0.4455112818050888,
+      "grad_norm": 0.7955581619122697,
+      "learning_rate": 8.85057471264368e-06,
+      "loss": 1.4619,
+      "step": 232
+    },
+    {
+      "epoch": 0.4474315890542487,
+      "grad_norm": 0.6665415125116451,
+      "learning_rate": 8.888888888888888e-06,
+      "loss": 1.4401,
+      "step": 233
+    },
+    {
+      "epoch": 0.44935189630340855,
+      "grad_norm": 0.7497106975893845,
+      "learning_rate": 8.9272030651341e-06,
+      "loss": 1.4444,
+      "step": 234
+    },
+    {
+      "epoch": 0.4512722035525684,
+      "grad_norm": 0.7655250611774667,
+      "learning_rate": 8.965517241379312e-06,
+      "loss": 1.4183,
+      "step": 235
+    },
+    {
+      "epoch": 0.4531925108017283,
+      "grad_norm": 0.7069725743129521,
+      "learning_rate": 9.003831417624522e-06,
+      "loss": 1.4695,
+      "step": 236
+    },
+    {
+      "epoch": 0.45511281805088816,
+      "grad_norm": 0.7033810098540295,
+      "learning_rate": 9.042145593869732e-06,
+      "loss": 1.432,
+      "step": 237
+    },
+    {
+      "epoch": 0.457033125300048,
+      "grad_norm": 0.7657511117560273,
+      "learning_rate": 9.080459770114942e-06,
+      "loss": 1.3106,
+      "step": 238
+    },
+    {
+      "epoch": 0.4589534325492079,
+      "grad_norm": 0.6734934664683107,
+      "learning_rate": 9.118773946360155e-06,
+      "loss": 1.2657,
+      "step": 239
+    },
+    {
+      "epoch": 0.46087373979836777,
+      "grad_norm": 0.6793414583903499,
+      "learning_rate": 9.157088122605364e-06,
+      "loss": 1.2823,
+      "step": 240
+    },
+    {
+      "epoch": 0.4627940470475276,
+      "grad_norm": 0.6868551209136969,
+      "learning_rate": 9.195402298850575e-06,
+      "loss": 1.1981,
+      "step": 241
+    },
+    {
+      "epoch": 0.46471435429668745,
+      "grad_norm": 0.7999498220323118,
+      "learning_rate": 9.233716475095786e-06,
+      "loss": 1.3222,
+      "step": 242
+    },
+    {
+      "epoch": 0.4666346615458473,
+      "grad_norm": 0.6765218175071678,
+      "learning_rate": 9.272030651340997e-06,
+      "loss": 1.3159,
+      "step": 243
+    },
+    {
+      "epoch": 0.4685549687950072,
+      "grad_norm": 0.7633174577339077,
+      "learning_rate": 9.310344827586207e-06,
+      "loss": 1.4493,
+      "step": 244
+    },
+    {
+      "epoch": 0.47047527604416706,
+      "grad_norm": 0.7434731700432043,
+      "learning_rate": 9.348659003831418e-06,
+      "loss": 1.3771,
+      "step": 245
+    },
+    {
+      "epoch": 0.4723955832933269,
+      "grad_norm": 0.7887594763056447,
+      "learning_rate": 9.386973180076629e-06,
+      "loss": 1.2978,
+      "step": 246
+    },
+    {
+      "epoch": 0.4743158905424868,
+      "grad_norm": 0.6816410513843812,
+      "learning_rate": 9.42528735632184e-06,
+      "loss": 1.3221,
+      "step": 247
+    },
+    {
+      "epoch": 0.47623619779164666,
+      "grad_norm": 0.6726044542699641,
+      "learning_rate": 9.46360153256705e-06,
+      "loss": 1.3954,
+      "step": 248
+    },
+    {
+      "epoch": 0.47815650504080653,
+      "grad_norm": 0.8209401718350966,
+      "learning_rate": 9.501915708812262e-06,
+      "loss": 1.2964,
+      "step": 249
+    },
+    {
+      "epoch": 0.4800768122899664,
+      "grad_norm": 0.7150467022402414,
+      "learning_rate": 9.54022988505747e-06,
+      "loss": 1.2462,
+      "step": 250
+    },
+    {
+      "epoch": 0.4819971195391263,
+      "grad_norm": 0.7473286809494867,
+      "learning_rate": 9.578544061302683e-06,
+      "loss": 1.3987,
+      "step": 251
+    },
+    {
+      "epoch": 0.48391742678828614,
+      "grad_norm": 0.8919512518954593,
+      "learning_rate": 9.616858237547894e-06,
+      "loss": 1.3271,
+      "step": 252
+    },
+    {
+      "epoch": 0.485837734037446,
+      "grad_norm": 0.7235436092130889,
+      "learning_rate": 9.655172413793105e-06,
+      "loss": 1.4006,
+      "step": 253
+    },
+    {
+      "epoch": 0.4877580412866059,
+      "grad_norm": 0.7496985395380509,
+      "learning_rate": 9.693486590038314e-06,
+      "loss": 1.3485,
+      "step": 254
+    },
+    {
+      "epoch": 0.4896783485357657,
+      "grad_norm": 0.7558104853132458,
+      "learning_rate": 9.731800766283525e-06,
+      "loss": 1.4354,
+      "step": 255
+    },
+    {
+      "epoch": 0.49159865578492556,
+      "grad_norm": 0.8026084646949407,
+      "learning_rate": 9.770114942528738e-06,
+      "loss": 1.3458,
+      "step": 256
+    },
+    {
+      "epoch": 0.49351896303408543,
+      "grad_norm": 0.6871351528584623,
+      "learning_rate": 9.808429118773947e-06,
+      "loss": 1.2849,
+      "step": 257
+    },
+    {
+      "epoch": 0.4954392702832453,
+      "grad_norm": 0.6733156882936058,
+      "learning_rate": 9.846743295019157e-06,
+      "loss": 1.3416,
+      "step": 258
+    },
+    {
+      "epoch": 0.4973595775324052,
+      "grad_norm": 0.8040078483269378,
+      "learning_rate": 9.885057471264368e-06,
+      "loss": 1.3933,
+      "step": 259
+    },
+    {
+      "epoch": 0.49927988478156504,
+      "grad_norm": 0.7019620968242437,
+      "learning_rate": 9.923371647509579e-06,
+      "loss": 1.4092,
+      "step": 260
+    },
+    {
+      "epoch": 0.501200192030725,
+      "grad_norm": 0.7993119421084648,
+      "learning_rate": 9.96168582375479e-06,
+      "loss": 1.4496,
+      "step": 261
+    },
+    {
+      "epoch": 0.5031204992798848,
+      "grad_norm": 0.7500420731214819,
+      "learning_rate": 1e-05,
+      "loss": 1.275,
+      "step": 262
+    },
+    {
+      "epoch": 0.5050408065290446,
+      "grad_norm": 0.7096092331830312,
+      "learning_rate": 9.999995509192137e-06,
+      "loss": 1.4075,
+      "step": 263
+    },
+    {
+      "epoch": 0.5069611137782045,
+      "grad_norm": 0.6880277184241935,
+      "learning_rate": 9.999982036776617e-06,
+      "loss": 1.3853,
+      "step": 264
+    },
+    {
+      "epoch": 0.5088814210273643,
+      "grad_norm": 0.6887794428820988,
+      "learning_rate": 9.999959582777638e-06,
+      "loss": 1.2526,
+      "step": 265
+    },
+    {
+      "epoch": 0.5108017282765243,
+      "grad_norm": 0.6962486538004352,
+      "learning_rate": 9.999928147235536e-06,
+      "loss": 1.385,
+      "step": 266
+    },
+    {
+      "epoch": 0.5127220355256841,
+      "grad_norm": 0.7314717827625464,
+      "learning_rate": 9.99988773020678e-06,
+      "loss": 1.5364,
+      "step": 267
+    },
+    {
+      "epoch": 0.514642342774844,
+      "grad_norm": 0.6755436676174023,
+      "learning_rate": 9.99983833176397e-06,
+      "loss": 1.1625,
+      "step": 268
+    },
+    {
+      "epoch": 0.5165626500240038,
+      "grad_norm": 0.6799257698916303,
+      "learning_rate": 9.999779951995845e-06,
+      "loss": 1.3743,
+      "step": 269
+    },
+    {
+      "epoch": 0.5184829572731637,
+      "grad_norm": 0.6967760972692357,
+      "learning_rate": 9.99971259100727e-06,
+      "loss": 1.4619,
+      "step": 270
+    },
+    {
+      "epoch": 0.5204032645223235,
+      "grad_norm": 0.6876004824004736,
+      "learning_rate": 9.99963624891925e-06,
+      "loss": 1.2302,
+      "step": 271
+    },
+    {
+      "epoch": 0.5223235717714835,
+      "grad_norm": 0.741525025016876,
+      "learning_rate": 9.999550925868919e-06,
+      "loss": 1.4944,
+      "step": 272
+    },
+    {
+      "epoch": 0.5242438790206433,
+      "grad_norm": 0.7044462970993084,
+      "learning_rate": 9.999456622009545e-06,
+      "loss": 1.3233,
+      "step": 273
+    },
+    {
+      "epoch": 0.5261641862698032,
+      "grad_norm": 0.7147463675731596,
+      "learning_rate": 9.999353337510526e-06,
+      "loss": 1.4728,
+      "step": 274
+    },
+    {
+      "epoch": 0.528084493518963,
+      "grad_norm": 0.7580689885973482,
+      "learning_rate": 9.9992410725574e-06,
+      "loss": 1.3894,
+      "step": 275
+    },
+    {
+      "epoch": 0.530004800768123,
+      "grad_norm": 0.627159916011877,
+      "learning_rate": 9.999119827351824e-06,
+      "loss": 1.1499,
+      "step": 276
+    },
+    {
+      "epoch": 0.5319251080172828,
+      "grad_norm": 0.7347721430138173,
+      "learning_rate": 9.998989602111599e-06,
+      "loss": 1.3863,
+      "step": 277
+    },
+    {
+      "epoch": 0.5338454152664426,
+      "grad_norm": 0.7561542947388183,
+      "learning_rate": 9.99885039707065e-06,
+      "loss": 1.4094,
+      "step": 278
+    },
+    {
+      "epoch": 0.5357657225156025,
+      "grad_norm": 0.7257428578657052,
+      "learning_rate": 9.998702212479031e-06,
+      "loss": 1.3558,
+      "step": 279
+    },
+    {
+      "epoch": 0.5376860297647623,
+      "grad_norm": 0.7156710718415759,
+      "learning_rate": 9.998545048602938e-06,
+      "loss": 1.3932,
+      "step": 280
+    },
+    {
+      "epoch": 0.5396063370139222,
+      "grad_norm": 0.8324536861856755,
+      "learning_rate": 9.998378905724677e-06,
+      "loss": 1.4631,
+      "step": 281
+    },
+    {
+      "epoch": 0.5415266442630821,
+      "grad_norm": 0.769865333129953,
+      "learning_rate": 9.998203784142701e-06,
+      "loss": 1.4778,
+      "step": 282
+    },
+    {
+      "epoch": 0.543446951512242,
+      "grad_norm": 0.7558894226395197,
+      "learning_rate": 9.998019684171585e-06,
+      "loss": 1.3401,
+      "step": 283
+    },
+    {
+      "epoch": 0.5453672587614018,
+      "grad_norm": 0.745453902467428,
+      "learning_rate": 9.997826606142031e-06,
+      "loss": 1.376,
+      "step": 284
+    },
+    {
+      "epoch": 0.5472875660105617,
+      "grad_norm": 0.8525307677948906,
+      "learning_rate": 9.997624550400869e-06,
+      "loss": 1.3269,
+      "step": 285
+    },
+    {
+      "epoch": 0.5492078732597215,
+      "grad_norm": 0.8993675673573773,
+      "learning_rate": 9.997413517311055e-06,
+      "loss": 1.3973,
+      "step": 286
+    },
+    {
+      "epoch": 0.5511281805088815,
+      "grad_norm": 0.7637767723846134,
+      "learning_rate": 9.997193507251676e-06,
+      "loss": 1.4107,
+      "step": 287
+    },
+    {
+      "epoch": 0.5530484877580413,
+      "grad_norm": 0.7334145684503066,
+      "learning_rate": 9.996964520617938e-06,
+      "loss": 1.2548,
+      "step": 288
+    },
+    {
+      "epoch": 0.5549687950072012,
+      "grad_norm": 0.8286763467838726,
+      "learning_rate": 9.996726557821177e-06,
+      "loss": 1.3915,
+      "step": 289
+    },
+    {
+      "epoch": 0.556889102256361,
+      "grad_norm": 0.786223417485439,
+      "learning_rate": 9.996479619288853e-06,
+      "loss": 1.2113,
+      "step": 290
+    },
+    {
+      "epoch": 0.5588094095055209,
+      "grad_norm": 0.6629161249458739,
+      "learning_rate": 9.996223705464542e-06,
+      "loss": 1.383,
+      "step": 291
+    },
+    {
+      "epoch": 0.5607297167546808,
+      "grad_norm": 0.7763568743151492,
+      "learning_rate": 9.995958816807951e-06,
+      "loss": 1.511,
+      "step": 292
+    },
+    {
+      "epoch": 0.5626500240038406,
+      "grad_norm": 0.8706771128343518,
+      "learning_rate": 9.995684953794905e-06,
+      "loss": 1.301,
+      "step": 293
+    },
+    {
+      "epoch": 0.5645703312530005,
+      "grad_norm": 0.7186979734065922,
+      "learning_rate": 9.995402116917353e-06,
+      "loss": 1.3137,
+      "step": 294
+    },
+    {
+      "epoch": 0.5664906385021603,
+      "grad_norm": 0.7197477949946243,
+      "learning_rate": 9.995110306683358e-06,
+      "loss": 1.3924,
+      "step": 295
+    },
+    {
+      "epoch": 0.5684109457513202,
+      "grad_norm": 0.7990094094001562,
+      "learning_rate": 9.994809523617109e-06,
+      "loss": 1.3595,
+      "step": 296
+    },
+    {
+      "epoch": 0.57033125300048,
+      "grad_norm": 0.7548256910113419,
+      "learning_rate": 9.994499768258905e-06,
+      "loss": 1.2627,
+      "step": 297
+    },
+    {
+      "epoch": 0.57225156024964,
+      "grad_norm": 0.6080847660246027,
+      "learning_rate": 9.994181041165169e-06,
+      "loss": 1.233,
+      "step": 298
+    },
+    {
+      "epoch": 0.5741718674987998,
+      "grad_norm": 0.7522747291038051,
+      "learning_rate": 9.99385334290844e-06,
+      "loss": 1.4473,
+      "step": 299
+    },
+    {
+      "epoch": 0.5760921747479597,
+      "grad_norm": 0.693071662167599,
+      "learning_rate": 9.993516674077367e-06,
+      "loss": 1.3112,
+      "step": 300
+    },
+    {
+      "epoch": 0.5780124819971195,
+      "grad_norm": 0.7289575683988432,
+      "learning_rate": 9.993171035276717e-06,
+      "loss": 1.4447,
+      "step": 301
+    },
+    {
+      "epoch": 0.5799327892462794,
+      "grad_norm": 0.7219332107019164,
+      "learning_rate": 9.992816427127367e-06,
+      "loss": 1.3059,
+      "step": 302
+    },
+    {
+      "epoch": 0.5818530964954393,
+      "grad_norm": 0.7884436507160677,
+      "learning_rate": 9.992452850266313e-06,
+      "loss": 1.4828,
+      "step": 303
+    },
+    {
+      "epoch": 0.5837734037445992,
+      "grad_norm": 0.6835750513473184,
+      "learning_rate": 9.992080305346652e-06,
+      "loss": 1.3351,
+      "step": 304
+    },
+    {
+      "epoch": 0.585693710993759,
+      "grad_norm": 0.7455682427352298,
+      "learning_rate": 9.991698793037596e-06,
+      "loss": 1.3393,
+      "step": 305
+    },
+    {
+      "epoch": 0.5876140182429188,
+      "grad_norm": 0.6344234674402865,
+      "learning_rate": 9.991308314024466e-06,
+      "loss": 1.1186,
+      "step": 306
+    },
+    {
+      "epoch": 0.5895343254920787,
+      "grad_norm": 0.6463680302231726,
+      "learning_rate": 9.990908869008685e-06,
+      "loss": 1.3536,
+      "step": 307
+    },
+    {
+      "epoch": 0.5914546327412386,
+      "grad_norm": 0.7763640180054763,
+      "learning_rate": 9.99050045870779e-06,
+      "loss": 1.4034,
+      "step": 308
+    },
+    {
+      "epoch": 0.5933749399903985,
+      "grad_norm": 0.7729679972786335,
+      "learning_rate": 9.990083083855413e-06,
+      "loss": 1.518,
+      "step": 309
+    },
+    {
+      "epoch": 0.5952952472395583,
+      "grad_norm": 0.7293910429882318,
+      "learning_rate": 9.9896567452013e-06,
+      "loss": 1.4168,
+      "step": 310
+    },
+    {
+      "epoch": 0.5972155544887182,
+      "grad_norm": 0.6938721572185663,
+      "learning_rate": 9.989221443511286e-06,
+      "loss": 1.5813,
+      "step": 311
+    },
+    {
+      "epoch": 0.599135861737878,
+      "grad_norm": 0.6823348506968784,
+      "learning_rate": 9.98877717956732e-06,
+      "loss": 1.396,
+      "step": 312
+    },
+    {
+      "epoch": 0.601056168987038,
+      "grad_norm": 0.6470598419144085,
+      "learning_rate": 9.988323954167438e-06,
+      "loss": 1.3138,
+      "step": 313
+    },
+    {
+      "epoch": 0.6029764762361978,
+      "grad_norm": 0.6955870872672735,
+      "learning_rate": 9.987861768125783e-06,
+      "loss": 1.4043,
+      "step": 314
+    },
+    {
+      "epoch": 0.6048967834853577,
+      "grad_norm": 0.7788214111440995,
+      "learning_rate": 9.98739062227259e-06,
+      "loss": 1.5305,
+      "step": 315
+    },
+    {
+      "epoch": 0.6068170907345175,
+      "grad_norm": 0.6598071275711943,
+      "learning_rate": 9.986910517454188e-06,
+      "loss": 1.185,
+      "step": 316
+    },
+    {
+      "epoch": 0.6087373979836774,
+      "grad_norm": 0.7478469808999688,
+      "learning_rate": 9.986421454533001e-06,
+      "loss": 1.2875,
+      "step": 317
+    },
+    {
+      "epoch": 0.6106577052328372,
+      "grad_norm": 0.7102386979345604,
+      "learning_rate": 9.985923434387545e-06,
+      "loss": 1.2623,
+      "step": 318
+    },
+    {
+      "epoch": 0.6125780124819972,
+      "grad_norm": 0.8159684281974479,
+      "learning_rate": 9.985416457912423e-06,
+      "loss": 1.3859,
+      "step": 319
+    },
+    {
+      "epoch": 0.614498319731157,
+      "grad_norm": 0.7359831696017346,
+      "learning_rate": 9.984900526018331e-06,
+      "loss": 1.3127,
+      "step": 320
+    },
+    {
+      "epoch": 0.6164186269803168,
+      "grad_norm": 0.7406050107164467,
+      "learning_rate": 9.984375639632047e-06,
+      "loss": 1.3526,
+      "step": 321
+    },
+    {
+      "epoch": 0.6183389342294767,
+      "grad_norm": 0.7171563092825125,
+      "learning_rate": 9.98384179969644e-06,
+      "loss": 1.3718,
+      "step": 322
+    },
+    {
+      "epoch": 0.6202592414786365,
+      "grad_norm": 0.7444992833186831,
+      "learning_rate": 9.983299007170454e-06,
+      "loss": 1.2605,
+      "step": 323
+    },
+    {
+      "epoch": 0.6221795487277965,
+      "grad_norm": 0.6718703991955216,
+      "learning_rate": 9.982747263029123e-06,
+      "loss": 1.2696,
+      "step": 324
+    },
+    {
+      "epoch": 0.6240998559769563,
+      "grad_norm": 0.6953480856589247,
+      "learning_rate": 9.982186568263558e-06,
+      "loss": 1.402,
+      "step": 325
+    },
+    {
+      "epoch": 0.6260201632261162,
+      "grad_norm": 0.6814410132323387,
+      "learning_rate": 9.981616923880948e-06,
+      "loss": 1.357,
+      "step": 326
+    },
+    {
+      "epoch": 0.627940470475276,
+      "grad_norm": 0.6789399695802566,
+      "learning_rate": 9.981038330904556e-06,
+      "loss": 1.3563,
+      "step": 327
+    },
+    {
+      "epoch": 0.6298607777244359,
+      "grad_norm": 0.7084278817620212,
+      "learning_rate": 9.980450790373724e-06,
+      "loss": 1.3983,
+      "step": 328
+    },
+    {
+      "epoch": 0.6317810849735958,
+      "grad_norm": 0.7561797459599714,
+      "learning_rate": 9.979854303343866e-06,
+      "loss": 1.3998,
+      "step": 329
+    },
+    {
+      "epoch": 0.6337013922227557,
+      "grad_norm": 0.7063013474407709,
+      "learning_rate": 9.979248870886463e-06,
+      "loss": 1.4128,
+      "step": 330
+    },
+    {
+      "epoch": 0.6356216994719155,
+      "grad_norm": 0.7285839877290075,
+      "learning_rate": 9.978634494089066e-06,
+      "loss": 1.5026,
+      "step": 331
+    },
+    {
+      "epoch": 0.6375420067210754,
+      "grad_norm": 0.7631063764422019,
+      "learning_rate": 9.9780111740553e-06,
+      "loss": 1.3353,
+      "step": 332
+    },
+    {
+      "epoch": 0.6394623139702352,
+      "grad_norm": 0.8022733412976211,
+      "learning_rate": 9.977378911904843e-06,
+      "loss": 1.3017,
+      "step": 333
+    },
+    {
+      "epoch": 0.6413826212193952,
+      "grad_norm": 0.7397584519382538,
+      "learning_rate": 9.976737708773445e-06,
+      "loss": 1.4274,
+      "step": 334
+    },
+    {
+      "epoch": 0.643302928468555,
+      "grad_norm": 0.7498254808267637,
+      "learning_rate": 9.976087565812913e-06,
+      "loss": 1.2821,
+      "step": 335
+    },
+    {
+      "epoch": 0.6452232357177148,
+      "grad_norm": 0.8440233129581642,
+      "learning_rate": 9.975428484191117e-06,
+      "loss": 1.507,
+      "step": 336
+    },
+    {
+      "epoch": 0.6471435429668747,
+      "grad_norm": 0.7140420729824992,
+      "learning_rate": 9.974760465091975e-06,
+      "loss": 1.3541,
+      "step": 337
+    },
+    {
+      "epoch": 0.6490638502160345,
+      "grad_norm": 0.7935959679049022,
+      "learning_rate": 9.974083509715471e-06,
+      "loss": 1.3704,
+      "step": 338
+    },
+    {
+      "epoch": 0.6509841574651944,
+      "grad_norm": 0.8128531878052235,
+      "learning_rate": 9.973397619277631e-06,
+      "loss": 1.3941,
+      "step": 339
+    },
+    {
+      "epoch": 0.6529044647143543,
+      "grad_norm": 0.7087809281812139,
+      "learning_rate": 9.972702795010539e-06,
+      "loss": 1.3024,
+      "step": 340
+    },
+    {
+      "epoch": 0.6548247719635142,
+      "grad_norm": 0.8779292688681924,
+      "learning_rate": 9.971999038162322e-06,
+      "loss": 1.275,
+      "step": 341
+    },
+    {
+      "epoch": 0.656745079212674,
+      "grad_norm": 0.8321554450658822,
+      "learning_rate": 9.971286349997155e-06,
+      "loss": 1.4001,
+      "step": 342
+    },
+    {
+      "epoch": 0.6586653864618339,
+      "grad_norm": 0.6785835670278665,
+      "learning_rate": 9.970564731795259e-06,
+      "loss": 1.3563,
+      "step": 343
+    },
+    {
+      "epoch": 0.6605856937109937,
+      "grad_norm": 0.8154880274231641,
+      "learning_rate": 9.96983418485289e-06,
+      "loss": 1.2015,
+      "step": 344
+    },
+    {
+      "epoch": 0.6625060009601537,
+      "grad_norm": 0.7877809686842095,
+      "learning_rate": 9.969094710482345e-06,
+      "loss": 1.3176,
+      "step": 345
+    },
+    {
+      "epoch": 0.6644263082093135,
+      "grad_norm": 0.737869545031943,
+      "learning_rate": 9.968346310011965e-06,
+      "loss": 1.2681,
+      "step": 346
+    },
+    {
+      "epoch": 0.6663466154584734,
+      "grad_norm": 0.8367016706595558,
+      "learning_rate": 9.967588984786113e-06,
+      "loss": 1.3146,
+      "step": 347
+    },
+    {
+      "epoch": 0.6682669227076332,
+      "grad_norm": 0.7157484065111132,
+      "learning_rate": 9.966822736165194e-06,
+      "loss": 1.2279,
+      "step": 348
+    },
+    {
+      "epoch": 0.6701872299567931,
+      "grad_norm": 0.8340115611171157,
+      "learning_rate": 9.966047565525636e-06,
+      "loss": 1.346,
+      "step": 349
+    },
+    {
+      "epoch": 0.672107537205953,
+      "grad_norm": 0.7964501684625079,
+      "learning_rate": 9.965263474259896e-06,
+      "loss": 1.2604,
+      "step": 350
+    },
+    {
+      "epoch": 0.6740278444551128,
+      "grad_norm": 0.7644952492186728,
+      "learning_rate": 9.964470463776457e-06,
+      "loss": 1.2808,
+      "step": 351
+    },
+    {
+      "epoch": 0.6759481517042727,
+      "grad_norm": 0.715875104396104,
+      "learning_rate": 9.96366853549982e-06,
+      "loss": 1.3779,
+      "step": 352
+    },
+    {
+      "epoch": 0.6778684589534325,
+      "grad_norm": 0.6658897567313692,
+      "learning_rate": 9.962857690870507e-06,
+      "loss": 1.3005,
+      "step": 353
+    },
+    {
+      "epoch": 0.6797887662025924,
+      "grad_norm": 0.7869576623216241,
+      "learning_rate": 9.962037931345058e-06,
+      "loss": 1.3772,
+      "step": 354
+    },
+    {
+      "epoch": 0.6817090734517522,
+      "grad_norm": 0.9271972422960602,
+      "learning_rate": 9.96120925839603e-06,
+      "loss": 1.346,
+      "step": 355
+    },
+    {
+      "epoch": 0.6836293807009122,
+      "grad_norm": 0.7660166493856924,
+      "learning_rate": 9.96037167351198e-06,
+      "loss": 1.2745,
+      "step": 356
+    },
+    {
+      "epoch": 0.685549687950072,
+      "grad_norm": 0.6278698318719681,
+      "learning_rate": 9.959525178197484e-06,
+      "loss": 1.1849,
+      "step": 357
+    },
+    {
+      "epoch": 0.6874699951992319,
+      "grad_norm": 0.8957448047534449,
+      "learning_rate": 9.958669773973124e-06,
+      "loss": 1.47,
+      "step": 358
+    },
+    {
+      "epoch": 0.6893903024483917,
+      "grad_norm": 0.7893290798004599,
+      "learning_rate": 9.95780546237548e-06,
+      "loss": 1.4411,
+      "step": 359
+    },
+    {
+      "epoch": 0.6913106096975516,
+      "grad_norm": 0.6674442252477479,
+      "learning_rate": 9.956932244957135e-06,
+      "loss": 1.3468,
+      "step": 360
+    },
+    {
+      "epoch": 0.6932309169467115,
+      "grad_norm": 0.8637622685527974,
+      "learning_rate": 9.95605012328667e-06,
+      "loss": 1.4951,
+      "step": 361
+    },
+    {
+      "epoch": 0.6951512241958714,
+      "grad_norm": 0.7032170307148827,
+      "learning_rate": 9.95515909894866e-06,
+      "loss": 1.2168,
+      "step": 362
+    },
+    {
+      "epoch": 0.6970715314450312,
+      "grad_norm": 0.7604809716338964,
+      "learning_rate": 9.954259173543671e-06,
+      "loss": 1.3405,
+      "step": 363
+    },
+    {
+      "epoch": 0.6989918386941911,
+      "grad_norm": 0.7356596357988364,
+      "learning_rate": 9.953350348688264e-06,
+      "loss": 1.4851,
+      "step": 364
+    },
+    {
+      "epoch": 0.7009121459433509,
+      "grad_norm": 0.7106099387583814,
+      "learning_rate": 9.952432626014979e-06,
+      "loss": 1.4105,
+      "step": 365
+    },
+    {
+      "epoch": 0.7028324531925108,
+      "grad_norm": 0.7680103674421779,
+      "learning_rate": 9.951506007172344e-06,
+      "loss": 1.3767,
+      "step": 366
+    },
+    {
+      "epoch": 0.7047527604416707,
+      "grad_norm": 0.7045547217675817,
+      "learning_rate": 9.950570493824864e-06,
+      "loss": 1.3534,
+      "step": 367
+    },
+    {
+      "epoch": 0.7066730676908305,
+      "grad_norm": 0.8618867023288301,
+      "learning_rate": 9.949626087653026e-06,
+      "loss": 1.5932,
+      "step": 368
+    },
+    {
+      "epoch": 0.7085933749399904,
+      "grad_norm": 0.7514120877501722,
+      "learning_rate": 9.948672790353287e-06,
+      "loss": 1.4226,
+      "step": 369
+    },
+    {
+      "epoch": 0.7105136821891502,
+      "grad_norm": 0.7707655494026907,
+      "learning_rate": 9.947710603638078e-06,
+      "loss": 1.3086,
+      "step": 370
+    },
+    {
+      "epoch": 0.7124339894383102,
+      "grad_norm": 0.7353314917000675,
+      "learning_rate": 9.946739529235797e-06,
+      "loss": 1.3498,
+      "step": 371
+    },
+    {
+      "epoch": 0.71435429668747,
+      "grad_norm": 0.7677899312066072,
+      "learning_rate": 9.945759568890804e-06,
+      "loss": 1.337,
+      "step": 372
+    },
+    {
+      "epoch": 0.7162746039366299,
+      "grad_norm": 0.7323170602281932,
+      "learning_rate": 9.944770724363428e-06,
+      "loss": 1.2262,
+      "step": 373
+    },
+    {
+      "epoch": 0.7181949111857897,
+      "grad_norm": 0.7030503301748048,
+      "learning_rate": 9.943772997429955e-06,
+      "loss": 1.2604,
+      "step": 374
+    },
+    {
+      "epoch": 0.7201152184349496,
+      "grad_norm": 0.8803804845996765,
+      "learning_rate": 9.942766389882621e-06,
+      "loss": 1.3465,
+      "step": 375
+    },
+    {
+      "epoch": 0.7220355256841094,
+      "grad_norm": 0.765754153505594,
+      "learning_rate": 9.94175090352962e-06,
+      "loss": 1.4785,
+      "step": 376
+    },
+    {
+      "epoch": 0.7239558329332694,
+      "grad_norm": 0.7412100725496786,
+      "learning_rate": 9.940726540195093e-06,
+      "loss": 1.3886,
+      "step": 377
+    },
+    {
+      "epoch": 0.7258761401824292,
+      "grad_norm": 0.7352092180670398,
+      "learning_rate": 9.939693301719131e-06,
+      "loss": 1.3787,
+      "step": 378
+    },
+    {
+      "epoch": 0.727796447431589,
+      "grad_norm": 0.7081810984489154,
+      "learning_rate": 9.93865118995776e-06,
+      "loss": 1.2855,
+      "step": 379
+    },
+    {
+      "epoch": 0.7297167546807489,
+      "grad_norm": 0.721692280601312,
+      "learning_rate": 9.937600206782951e-06,
+      "loss": 1.2581,
+      "step": 380
+    },
+    {
+      "epoch": 0.7316370619299087,
+      "grad_norm": 0.7219716174107607,
+      "learning_rate": 9.93654035408261e-06,
+      "loss": 1.3989,
+      "step": 381
+    },
+    {
+      "epoch": 0.7335573691790687,
+      "grad_norm": 0.808844014227327,
+      "learning_rate": 9.935471633760572e-06,
+      "loss": 1.489,
+      "step": 382
+    },
+    {
+      "epoch": 0.7354776764282285,
+      "grad_norm": 0.6931394591191726,
+      "learning_rate": 9.934394047736608e-06,
+      "loss": 1.3596,
+      "step": 383
+    },
+    {
+      "epoch": 0.7373979836773884,
+      "grad_norm": 0.7255320444997646,
+      "learning_rate": 9.93330759794641e-06,
+      "loss": 1.2593,
+      "step": 384
+    },
+    {
+      "epoch": 0.7393182909265482,
+      "grad_norm": 0.6469865776133421,
+      "learning_rate": 9.932212286341591e-06,
+      "loss": 1.4305,
+      "step": 385
+    },
+    {
+      "epoch": 0.7412385981757081,
+      "grad_norm": 0.7276641692049547,
+      "learning_rate": 9.931108114889685e-06,
+      "loss": 1.531,
+      "step": 386
+    },
+    {
+      "epoch": 0.743158905424868,
+      "grad_norm": 0.7064363862608019,
+      "learning_rate": 9.929995085574142e-06,
+      "loss": 1.3905,
+      "step": 387
+    },
+    {
+      "epoch": 0.7450792126740279,
+      "grad_norm": 0.7331138015593877,
+      "learning_rate": 9.928873200394323e-06,
+      "loss": 1.3649,
+      "step": 388
+    },
+    {
+      "epoch": 0.7469995199231877,
+      "grad_norm": 0.7324112634125343,
+      "learning_rate": 9.927742461365493e-06,
+      "loss": 1.4049,
+      "step": 389
+    },
+    {
+      "epoch": 0.7489198271723476,
+      "grad_norm": 0.7448582260656762,
+      "learning_rate": 9.926602870518826e-06,
+      "loss": 1.3451,
+      "step": 390
+    },
+    {
+      "epoch": 0.7508401344215074,
+      "grad_norm": 0.7290925508892867,
+      "learning_rate": 9.925454429901397e-06,
+      "loss": 1.265,
+      "step": 391
+    },
+    {
+      "epoch": 0.7527604416706674,
+      "grad_norm": 0.8652107575311744,
+      "learning_rate": 9.924297141576176e-06,
+      "loss": 1.2601,
+      "step": 392
+    },
+    {
+      "epoch": 0.7546807489198272,
+      "grad_norm": 0.8112589543786329,
+      "learning_rate": 9.923131007622027e-06,
+      "loss": 1.3949,
+      "step": 393
+    },
+    {
+      "epoch": 0.756601056168987,
+      "grad_norm": 0.7564370059278668,
+      "learning_rate": 9.9219560301337e-06,
+      "loss": 1.5582,
+      "step": 394
+    },
+    {
+      "epoch": 0.7585213634181469,
+      "grad_norm": 0.7217323314506744,
+      "learning_rate": 9.920772211221841e-06,
+      "loss": 1.3385,
+      "step": 395
+    },
+    {
+      "epoch": 0.7604416706673067,
+      "grad_norm": 0.7276996906484348,
+      "learning_rate": 9.919579553012964e-06,
+      "loss": 1.3778,
+      "step": 396
+    },
+    {
+      "epoch": 0.7623619779164666,
+      "grad_norm": 0.7823149650395126,
+      "learning_rate": 9.918378057649474e-06,
+      "loss": 1.3767,
+      "step": 397
+    },
+    {
+      "epoch": 0.7642822851656265,
+      "grad_norm": 0.6630175095699308,
+      "learning_rate": 9.917167727289641e-06,
+      "loss": 1.4844,
+      "step": 398
+    },
+    {
+      "epoch": 0.7662025924147864,
+      "grad_norm": 0.7344567983746931,
+      "learning_rate": 9.915948564107611e-06,
+      "loss": 1.3379,
+      "step": 399
+    },
+    {
+      "epoch": 0.7681228996639462,
+      "grad_norm": 0.7466071399468908,
+      "learning_rate": 9.914720570293397e-06,
+      "loss": 1.3972,
+      "step": 400
+    },
+    {
+      "epoch": 0.7700432069131061,
+      "grad_norm": 0.6949526386517407,
+      "learning_rate": 9.913483748052871e-06,
+      "loss": 1.4014,
+      "step": 401
+    },
+    {
+      "epoch": 0.7719635141622659,
+      "grad_norm": 0.6613334975361209,
+      "learning_rate": 9.912238099607763e-06,
+      "loss": 1.2069,
+      "step": 402
+    },
+    {
+      "epoch": 0.7738838214114259,
+      "grad_norm": 0.6632714829710001,
+      "learning_rate": 9.910983627195665e-06,
+      "loss": 1.2427,
+      "step": 403
+    },
+    {
+      "epoch": 0.7758041286605857,
+      "grad_norm": 0.6899922310091848,
+      "learning_rate": 9.90972033307001e-06,
+      "loss": 1.4041,
+      "step": 404
+    },
+    {
+      "epoch": 0.7777244359097456,
+      "grad_norm": 0.6259530173512385,
+      "learning_rate": 9.908448219500087e-06,
+      "loss": 1.0889,
+      "step": 405
+    },
+    {
+      "epoch": 0.7796447431589054,
+      "grad_norm": 0.7856357851084043,
+      "learning_rate": 9.90716728877102e-06,
+      "loss": 1.4446,
+      "step": 406
+    },
+    {
+      "epoch": 0.7815650504080653,
+      "grad_norm": 0.6391414313005859,
+      "learning_rate": 9.905877543183776e-06,
+      "loss": 1.3569,
+      "step": 407
+    },
+    {
+      "epoch": 0.7834853576572252,
+      "grad_norm": 0.7168868905941123,
+      "learning_rate": 9.904578985055151e-06,
+      "loss": 1.3422,
+      "step": 408
+    },
+    {
+      "epoch": 0.785405664906385,
+      "grad_norm": 0.7244557046923163,
+      "learning_rate": 9.903271616717782e-06,
+      "loss": 1.4439,
+      "step": 409
+    },
+    {
+      "epoch": 0.7873259721555449,
+      "grad_norm": 0.7135679454347851,
+      "learning_rate": 9.901955440520121e-06,
+      "loss": 1.417,
+      "step": 410
+    },
+    {
+      "epoch": 0.7892462794047047,
+      "grad_norm": 0.6975305523738247,
+      "learning_rate": 9.900630458826443e-06,
+      "loss": 1.2732,
+      "step": 411
+    },
+    {
+      "epoch": 0.7911665866538646,
+      "grad_norm": 0.6844791536520864,
+      "learning_rate": 9.89929667401685e-06,
+      "loss": 1.1994,
+      "step": 412
+    },
+    {
+      "epoch": 0.7930868939030244,
+      "grad_norm": 0.6995208409953875,
+      "learning_rate": 9.897954088487245e-06,
+      "loss": 1.44,
+      "step": 413
+    },
+    {
+      "epoch": 0.7950072011521844,
+      "grad_norm": 0.6664160171541003,
+      "learning_rate": 9.896602704649348e-06,
+      "loss": 1.3604,
+      "step": 414
+    },
+    {
+      "epoch": 0.7969275084013442,
+      "grad_norm": 0.6554390013099766,
+      "learning_rate": 9.89524252493068e-06,
+      "loss": 1.3002,
+      "step": 415
+    },
+    {
+      "epoch": 0.7988478156505041,
+      "grad_norm": 0.7714848495681179,
+      "learning_rate": 9.893873551774561e-06,
+      "loss": 1.4559,
+      "step": 416
+    },
+    {
+      "epoch": 0.8007681228996639,
+      "grad_norm": 0.684531285979124,
+      "learning_rate": 9.892495787640117e-06,
+      "loss": 1.4116,
+      "step": 417
+    },
+    {
+      "epoch": 0.8026884301488239,
+      "grad_norm": 0.625351185162249,
+      "learning_rate": 9.891109235002248e-06,
+      "loss": 1.2968,
+      "step": 418
+    },
+    {
+      "epoch": 0.8046087373979837,
+      "grad_norm": 0.6072425157093212,
+      "learning_rate": 9.889713896351658e-06,
+      "loss": 1.2834,
+      "step": 419
+    },
+    {
+      "epoch": 0.8065290446471436,
+      "grad_norm": 0.7557266979543664,
+      "learning_rate": 9.888309774194822e-06,
+      "loss": 1.2581,
+      "step": 420
+    },
+    {
+      "epoch": 0.8084493518963034,
+      "grad_norm": 0.6948893024278046,
+      "learning_rate": 9.886896871053996e-06,
+      "loss": 1.3472,
+      "step": 421
+    },
+    {
+      "epoch": 0.8103696591454633,
+      "grad_norm": 0.782279217584078,
+      "learning_rate": 9.885475189467217e-06,
+      "loss": 1.3546,
+      "step": 422
+    },
+    {
+      "epoch": 0.8122899663946231,
+      "grad_norm": 0.7188760215338773,
+      "learning_rate": 9.884044731988278e-06,
+      "loss": 1.3683,
+      "step": 423
+    },
+    {
+      "epoch": 0.814210273643783,
+      "grad_norm": 0.7215722020243497,
+      "learning_rate": 9.882605501186747e-06,
+      "loss": 1.2629,
+      "step": 424
+    },
+    {
+      "epoch": 0.8161305808929429,
+      "grad_norm": 0.699083447265363,
+      "learning_rate": 9.881157499647944e-06,
+      "loss": 1.3218,
+      "step": 425
+    },
+    {
+      "epoch": 0.8180508881421027,
+      "grad_norm": 0.724984065762376,
+      "learning_rate": 9.87970072997295e-06,
+      "loss": 1.3034,
+      "step": 426
+    },
+    {
+      "epoch": 0.8199711953912626,
+      "grad_norm": 0.7049961948332424,
+      "learning_rate": 9.878235194778594e-06,
+      "loss": 1.4015,
+      "step": 427
+    },
+    {
+      "epoch": 0.8218915026404224,
+      "grad_norm": 0.6961558656551843,
+      "learning_rate": 9.87676089669745e-06,
+      "loss": 1.3292,
+      "step": 428
+    },
+    {
+      "epoch": 0.8238118098895824,
+      "grad_norm": 0.6755568609982437,
+      "learning_rate": 9.875277838377835e-06,
+      "loss": 1.2485,
+      "step": 429
+    },
+    {
+      "epoch": 0.8257321171387422,
+      "grad_norm": 0.731506928442002,
+      "learning_rate": 9.8737860224838e-06,
+      "loss": 1.371,
+      "step": 430
+    },
+    {
+      "epoch": 0.8276524243879021,
+      "grad_norm": 0.8321982922227138,
+      "learning_rate": 9.872285451695128e-06,
+      "loss": 1.3981,
+      "step": 431
+    },
+    {
+      "epoch": 0.8295727316370619,
+      "grad_norm": 0.7315030508325402,
+      "learning_rate": 9.87077612870733e-06,
+      "loss": 1.3311,
+      "step": 432
+    },
+    {
+      "epoch": 0.8314930388862218,
+      "grad_norm": 0.798773303498576,
+      "learning_rate": 9.869258056231638e-06,
+      "loss": 1.3727,
+      "step": 433
+    },
+    {
+      "epoch": 0.8334133461353816,
+      "grad_norm": 0.651844540018506,
+      "learning_rate": 9.867731236995e-06,
+      "loss": 1.3471,
+      "step": 434
+    },
+    {
+      "epoch": 0.8353336533845416,
+      "grad_norm": 0.6771670988304741,
+      "learning_rate": 9.866195673740076e-06,
+      "loss": 1.3032,
+      "step": 435
+    },
+    {
+      "epoch": 0.8372539606337014,
+      "grad_norm": 0.8611651792236157,
+      "learning_rate": 9.864651369225236e-06,
+      "loss": 1.2559,
+      "step": 436
+    },
+    {
+      "epoch": 0.8391742678828612,
+      "grad_norm": 0.7436061953882284,
+      "learning_rate": 9.863098326224546e-06,
+      "loss": 1.3166,
+      "step": 437
+    },
+    {
+      "epoch": 0.8410945751320211,
+      "grad_norm": 0.7409544469403813,
+      "learning_rate": 9.86153654752778e-06,
+      "loss": 1.4249,
+      "step": 438
+    },
+    {
+      "epoch": 0.8430148823811809,
+      "grad_norm": 0.7184845856939001,
+      "learning_rate": 9.859966035940391e-06,
+      "loss": 1.3899,
+      "step": 439
+    },
+    {
+      "epoch": 0.8449351896303409,
+      "grad_norm": 0.7569549756216626,
+      "learning_rate": 9.858386794283527e-06,
+      "loss": 1.4622,
+      "step": 440
+    },
+    {
+      "epoch": 0.8468554968795007,
+      "grad_norm": 0.7230751877755235,
+      "learning_rate": 9.856798825394017e-06,
+      "loss": 1.3834,
+      "step": 441
+    },
+    {
+      "epoch": 0.8487758041286606,
+      "grad_norm": 0.7861385241652444,
+      "learning_rate": 9.855202132124367e-06,
+      "loss": 1.4005,
+      "step": 442
+    },
+    {
+      "epoch": 0.8506961113778204,
+      "grad_norm": 0.654961628331442,
+      "learning_rate": 9.853596717342751e-06,
+      "loss": 1.2536,
+      "step": 443
+    },
+    {
+      "epoch": 0.8526164186269803,
+      "grad_norm": 0.7338664793979522,
+      "learning_rate": 9.851982583933015e-06,
+      "loss": 1.4289,
+      "step": 444
+    },
+    {
+      "epoch": 0.8545367258761402,
+      "grad_norm": 0.647102767057043,
+      "learning_rate": 9.850359734794664e-06,
+      "loss": 1.3167,
+      "step": 445
+    },
+    {
+      "epoch": 0.8564570331253001,
+      "grad_norm": 0.7940838292459027,
+      "learning_rate": 9.84872817284286e-06,
+      "loss": 1.4604,
+      "step": 446
+    },
+    {
+      "epoch": 0.8583773403744599,
+      "grad_norm": 0.7447428727236874,
+      "learning_rate": 9.847087901008415e-06,
+      "loss": 1.392,
+      "step": 447
+    },
+    {
+      "epoch": 0.8602976476236198,
+      "grad_norm": 0.7643508165383381,
+      "learning_rate": 9.845438922237787e-06,
+      "loss": 1.3017,
+      "step": 448
+    },
+    {
+      "epoch": 0.8622179548727796,
+      "grad_norm": 0.7617962346454569,
+      "learning_rate": 9.843781239493076e-06,
+      "loss": 1.6087,
+      "step": 449
+    },
+    {
+      "epoch": 0.8641382621219396,
+      "grad_norm": 0.7385042271210234,
+      "learning_rate": 9.842114855752013e-06,
+      "loss": 1.287,
+      "step": 450
+    },
+    {
+      "epoch": 0.8660585693710994,
+      "grad_norm": 0.7904769194673538,
+      "learning_rate": 9.840439774007963e-06,
+      "loss": 1.3847,
+      "step": 451
+    },
+    {
+      "epoch": 0.8679788766202592,
+      "grad_norm": 0.7796917727237478,
+      "learning_rate": 9.838755997269917e-06,
+      "loss": 1.4052,
+      "step": 452
+    },
+    {
+      "epoch": 0.8698991838694191,
+      "grad_norm": 0.6685557701211076,
+      "learning_rate": 9.837063528562479e-06,
+      "loss": 1.2191,
+      "step": 453
+    },
+    {
+      "epoch": 0.8718194911185789,
+      "grad_norm": 0.682007015556512,
+      "learning_rate": 9.835362370925868e-06,
+      "loss": 1.4041,
+      "step": 454
+    },
+    {
+      "epoch": 0.8737397983677389,
+      "grad_norm": 0.6809420760471568,
+      "learning_rate": 9.833652527415918e-06,
+      "loss": 1.1179,
+      "step": 455
+    },
+    {
+      "epoch": 0.8756601056168987,
+      "grad_norm": 0.7496451496119366,
+      "learning_rate": 9.831934001104056e-06,
+      "loss": 1.1863,
+      "step": 456
+    },
+    {
+      "epoch": 0.8775804128660586,
+      "grad_norm": 0.7958820882337777,
+      "learning_rate": 9.830206795077313e-06,
+      "loss": 1.4097,
+      "step": 457
+    },
+    {
+      "epoch": 0.8795007201152184,
+      "grad_norm": 0.7760893359826682,
+      "learning_rate": 9.828470912438308e-06,
+      "loss": 1.4765,
+      "step": 458
+    },
+    {
+      "epoch": 0.8814210273643783,
+      "grad_norm": 0.7181251163054755,
+      "learning_rate": 9.826726356305248e-06,
+      "loss": 1.2336,
+      "step": 459
+    },
+    {
+      "epoch": 0.8833413346135381,
+      "grad_norm": 0.7130113363618711,
+      "learning_rate": 9.824973129811919e-06,
+      "loss": 1.3356,
+      "step": 460
+    },
+    {
+      "epoch": 0.8852616418626981,
+      "grad_norm": 0.6976783021258673,
+      "learning_rate": 9.823211236107684e-06,
+      "loss": 1.5375,
+      "step": 461
+    },
+    {
+      "epoch": 0.8871819491118579,
+      "grad_norm": 0.6494482314881563,
+      "learning_rate": 9.82144067835747e-06,
+      "loss": 1.4216,
+      "step": 462
+    },
+    {
+      "epoch": 0.8891022563610178,
+      "grad_norm": 0.6799368531582445,
+      "learning_rate": 9.819661459741774e-06,
+      "loss": 1.2607,
+      "step": 463
+    },
+    {
+      "epoch": 0.8910225636101776,
+      "grad_norm": 0.7200924252996468,
+      "learning_rate": 9.817873583456646e-06,
+      "loss": 1.3954,
+      "step": 464
+    },
+    {
+      "epoch": 0.8929428708593375,
+      "grad_norm": 0.723830426362561,
+      "learning_rate": 9.816077052713689e-06,
+      "loss": 1.2634,
+      "step": 465
+    },
+    {
+      "epoch": 0.8948631781084974,
+      "grad_norm": 0.6795402172591166,
+      "learning_rate": 9.814271870740054e-06,
+      "loss": 1.288,
+      "step": 466
+    },
+    {
+      "epoch": 0.8967834853576572,
+      "grad_norm": 0.7798610485610813,
+      "learning_rate": 9.812458040778433e-06,
+      "loss": 1.3816,
+      "step": 467
+    },
+    {
+      "epoch": 0.8987037926068171,
+      "grad_norm": 0.6598552167407624,
+      "learning_rate": 9.810635566087046e-06,
+      "loss": 1.3142,
+      "step": 468
+    },
+    {
+      "epoch": 0.9006240998559769,
+      "grad_norm": 0.8267440234447042,
+      "learning_rate": 9.808804449939649e-06,
+      "loss": 1.3952,
+      "step": 469
+    },
+    {
+      "epoch": 0.9025444071051368,
+      "grad_norm": 0.6831251782753334,
+      "learning_rate": 9.806964695625521e-06,
+      "loss": 1.3507,
+      "step": 470
+    },
+    {
+      "epoch": 0.9044647143542967,
+      "grad_norm": 0.6913550521559437,
+      "learning_rate": 9.80511630644945e-06,
+      "loss": 1.3858,
+      "step": 471
+    },
+    {
+      "epoch": 0.9063850216034566,
+      "grad_norm": 0.651357091385134,
+      "learning_rate": 9.803259285731744e-06,
+      "loss": 1.3119,
+      "step": 472
+    },
+    {
+      "epoch": 0.9083053288526164,
+      "grad_norm": 0.6876771805281661,
+      "learning_rate": 9.801393636808213e-06,
+      "loss": 1.405,
+      "step": 473
+    },
+    {
+      "epoch": 0.9102256361017763,
+      "grad_norm": 0.6704161705700665,
+      "learning_rate": 9.79951936303016e-06,
+      "loss": 1.1645,
+      "step": 474
+    },
+    {
+      "epoch": 0.9121459433509361,
+      "grad_norm": 0.8403757549868232,
+      "learning_rate": 9.797636467764392e-06,
+      "loss": 1.3374,
+      "step": 475
+    },
+    {
+      "epoch": 0.914066250600096,
+      "grad_norm": 0.6976996239643926,
+      "learning_rate": 9.795744954393193e-06,
+      "loss": 1.2789,
+      "step": 476
+    },
+    {
+      "epoch": 0.9159865578492559,
+      "grad_norm": 0.7007105159698541,
+      "learning_rate": 9.793844826314338e-06,
+      "loss": 1.2513,
+      "step": 477
+    },
+    {
+      "epoch": 0.9179068650984158,
+      "grad_norm": 0.8046443692914484,
+      "learning_rate": 9.791936086941065e-06,
+      "loss": 1.4276,
+      "step": 478
+    },
+    {
+      "epoch": 0.9198271723475756,
+      "grad_norm": 0.6689604084398844,
+      "learning_rate": 9.790018739702091e-06,
+      "loss": 1.1329,
+      "step": 479
+    },
+    {
+      "epoch": 0.9217474795967355,
+      "grad_norm": 0.7418530237185074,
+      "learning_rate": 9.788092788041589e-06,
+      "loss": 1.2312,
+      "step": 480
+    },
+    {
+      "epoch": 0.9236677868458953,
+      "grad_norm": 0.7344158824384261,
+      "learning_rate": 9.78615823541919e-06,
+      "loss": 1.5327,
+      "step": 481
+    },
+    {
+      "epoch": 0.9255880940950552,
+      "grad_norm": 0.7029451657520306,
+      "learning_rate": 9.784215085309977e-06,
+      "loss": 1.3297,
+      "step": 482
+    },
+    {
+      "epoch": 0.9275084013442151,
+      "grad_norm": 0.8107964882673727,
+      "learning_rate": 9.782263341204477e-06,
+      "loss": 1.2561,
+      "step": 483
+    },
+    {
+      "epoch": 0.9294287085933749,
+      "grad_norm": 0.7114545290156662,
+      "learning_rate": 9.78030300660865e-06,
+      "loss": 1.3645,
+      "step": 484
+    },
+    {
+      "epoch": 0.9313490158425348,
+      "grad_norm": 0.7402003366232237,
+      "learning_rate": 9.77833408504389e-06,
+      "loss": 1.3642,
+      "step": 485
+    },
+    {
+      "epoch": 0.9332693230916946,
+      "grad_norm": 0.7109408774118504,
+      "learning_rate": 9.77635658004702e-06,
+      "loss": 1.3221,
+      "step": 486
+    },
+    {
+      "epoch": 0.9351896303408546,
+      "grad_norm": 0.7433358040489166,
+      "learning_rate": 9.774370495170276e-06,
+      "loss": 1.4449,
+      "step": 487
+    },
+    {
+      "epoch": 0.9371099375900144,
+      "grad_norm": 0.6785055655567391,
+      "learning_rate": 9.772375833981306e-06,
+      "loss": 1.3555,
+      "step": 488
+    },
+    {
+      "epoch": 0.9390302448391743,
+      "grad_norm": 0.76870851128488,
+      "learning_rate": 9.770372600063172e-06,
+      "loss": 1.284,
+      "step": 489
+    },
+    {
+      "epoch": 0.9409505520883341,
+      "grad_norm": 0.7344357851176699,
+      "learning_rate": 9.768360797014325e-06,
+      "loss": 1.2853,
+      "step": 490
+    },
+    {
+      "epoch": 0.942870859337494,
+      "grad_norm": 0.8470942465410728,
+      "learning_rate": 9.766340428448614e-06,
+      "loss": 1.3829,
+      "step": 491
+    },
+    {
+      "epoch": 0.9447911665866539,
+      "grad_norm": 0.7211389944931649,
+      "learning_rate": 9.764311497995272e-06,
+      "loss": 1.2677,
+      "step": 492
+    },
+    {
+      "epoch": 0.9467114738358138,
+      "grad_norm": 0.7084359929065828,
+      "learning_rate": 9.762274009298918e-06,
+      "loss": 1.2434,
+      "step": 493
+    },
+    {
+      "epoch": 0.9486317810849736,
+      "grad_norm": 0.7689934246592068,
+      "learning_rate": 9.760227966019537e-06,
+      "loss": 1.4095,
+      "step": 494
+    },
+    {
+      "epoch": 0.9505520883341335,
+      "grad_norm": 0.7773642092371199,
+      "learning_rate": 9.758173371832485e-06,
+      "loss": 1.3244,
+      "step": 495
+    },
+    {
+      "epoch": 0.9524723955832933,
+      "grad_norm": 0.6978701658115153,
+      "learning_rate": 9.756110230428476e-06,
+      "loss": 1.2787,
+      "step": 496
+    },
+    {
+      "epoch": 0.9543927028324531,
+      "grad_norm": 0.6910966359494893,
+      "learning_rate": 9.75403854551358e-06,
+      "loss": 1.3348,
+      "step": 497
+    },
+    {
+      "epoch": 0.9563130100816131,
+      "grad_norm": 0.732636720833676,
+      "learning_rate": 9.751958320809213e-06,
+      "loss": 1.2403,
+      "step": 498
+    },
+    {
+      "epoch": 0.9582333173307729,
+      "grad_norm": 0.7804889809056719,
+      "learning_rate": 9.749869560052128e-06,
+      "loss": 1.1905,
+      "step": 499
+    },
+    {
+      "epoch": 0.9601536245799328,
+      "grad_norm": 0.7286628977028098,
+      "learning_rate": 9.747772266994418e-06,
+      "loss": 1.3252,
+      "step": 500
+    },
+    {
+      "epoch": 0.9620739318290926,
+      "grad_norm": 0.730360159535326,
+      "learning_rate": 9.745666445403496e-06,
+      "loss": 1.4712,
+      "step": 501
+    },
+    {
+      "epoch": 0.9639942390782525,
+      "grad_norm": 0.6650898577066633,
+      "learning_rate": 9.7435520990621e-06,
+      "loss": 1.2945,
+      "step": 502
+    },
+    {
+      "epoch": 0.9659145463274124,
+      "grad_norm": 0.6813965501305161,
+      "learning_rate": 9.741429231768278e-06,
+      "loss": 1.3214,
+      "step": 503
+    },
+    {
+      "epoch": 0.9678348535765723,
+      "grad_norm": 0.8240692929170976,
+      "learning_rate": 9.739297847335387e-06,
+      "loss": 1.4367,
+      "step": 504
+    },
+    {
+      "epoch": 0.9697551608257321,
+      "grad_norm": 0.7664659172540482,
+      "learning_rate": 9.73715794959208e-06,
+      "loss": 1.2429,
+      "step": 505
+    },
+    {
+      "epoch": 0.971675468074892,
+      "grad_norm": 0.663273604561841,
+      "learning_rate": 9.735009542382308e-06,
+      "loss": 1.2678,
+      "step": 506
+    },
+    {
+      "epoch": 0.9735957753240518,
+      "grad_norm": 0.7286317666999398,
+      "learning_rate": 9.732852629565302e-06,
+      "loss": 1.351,
+      "step": 507
+    },
+    {
+      "epoch": 0.9755160825732118,
+      "grad_norm": 0.7222364516570275,
+      "learning_rate": 9.730687215015576e-06,
+      "loss": 1.3875,
+      "step": 508
+    },
+    {
+      "epoch": 0.9774363898223716,
+      "grad_norm": 0.792789498600007,
+      "learning_rate": 9.728513302622911e-06,
+      "loss": 1.4158,
+      "step": 509
+    },
+    {
+      "epoch": 0.9793566970715314,
+      "grad_norm": 0.6898048543889689,
+      "learning_rate": 9.72633089629236e-06,
+      "loss": 1.3018,
+      "step": 510
+    },
+    {
+      "epoch": 0.9812770043206913,
+      "grad_norm": 0.7080789810250435,
+      "learning_rate": 9.72413999994423e-06,
+      "loss": 1.2951,
+      "step": 511
+    },
+    {
+      "epoch": 0.9831973115698511,
+      "grad_norm": 0.6471793883594157,
+      "learning_rate": 9.721940617514076e-06,
+      "loss": 1.1768,
+      "step": 512
+    },
+    {
+      "epoch": 0.985117618819011,
+      "grad_norm": 0.6848848680839071,
+      "learning_rate": 9.719732752952702e-06,
+      "loss": 1.262,
+      "step": 513
+    },
+    {
+      "epoch": 0.9870379260681709,
+      "grad_norm": 0.7903965882462866,
+      "learning_rate": 9.717516410226144e-06,
+      "loss": 1.4717,
+      "step": 514
+    },
+    {
+      "epoch": 0.9889582333173308,
+      "grad_norm": 0.7404310033314039,
+      "learning_rate": 9.715291593315672e-06,
+      "loss": 1.3879,
+      "step": 515
+    },
+    {
+      "epoch": 0.9908785405664906,
+      "grad_norm": 0.735452133325044,
+      "learning_rate": 9.713058306217776e-06,
+      "loss": 1.3079,
+      "step": 516
+    },
+    {
+      "epoch": 0.9927988478156505,
+      "grad_norm": 0.8130352152653534,
+      "learning_rate": 9.710816552944157e-06,
+      "loss": 1.434,
+      "step": 517
+    },
+    {
+      "epoch": 0.9947191550648103,
+      "grad_norm": 0.7502971580652452,
+      "learning_rate": 9.708566337521736e-06,
+      "loss": 1.3013,
+      "step": 518
+    },
+    {
+      "epoch": 0.9966394623139703,
+      "grad_norm": 0.6582057806093718,
+      "learning_rate": 9.70630766399262e-06,
+      "loss": 1.2994,
+      "step": 519
+    },
+    {
+      "epoch": 0.9985597695631301,
+      "grad_norm": 0.7398007707770013,
+      "learning_rate": 9.70404053641412e-06,
+      "loss": 1.3135,
+      "step": 520
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 0.7398007707770013,
+      "learning_rate": 9.701764958858729e-06,
+      "loss": 1.1265,
+      "step": 521
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 2605,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 90367657066496.0,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-521/zero_to_fp32.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)

config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.1,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "4.52.4"
+}

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,208 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 5.0,
+    "total_flos": 451881746563072.0,
+    "train_loss": 1.2663256504714147,
+    "train_runtime": 105590.0324,
+    "train_samples_per_second": 0.394,
+    "train_steps_per_second": 0.025
+}