Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+checkpoint-1000/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoint-1500/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoint-500/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

adapter_config.json

@@ -0,0 +1,37 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-7B",
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "gate_proj",
+    "down_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1f1ce8c9922c384cf7af3046be04f4ed75efbc3066b8e3ee012e28782155f77
+size 80792880

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-1000/README.md

@@ -0,0 +1,202 @@
+---
+base_model: Qwen/Qwen2.5-7B
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.14.0

checkpoint-1000/adapter_config.json

@@ -0,0 +1,37 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-7B",
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "gate_proj",
+    "down_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-1000/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad117196d0adf2eb276e01fe5eda339cfca3825629b58df7e861daf6357260da
+size 80792880

checkpoint-1000/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-1000/latest

@@ -0,0 +1 @@
+global_step1001

checkpoint-1000/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6cb795a5cea0baa625c50007a6c9da09c6bbb5c16b560424070384a479e7d8a6
+size 14512

checkpoint-1000/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f19604377bd828eb366c68946ad997a4ff4d69beaeea93ee58915135768ec63
+size 14512

checkpoint-1000/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": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-1000/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896

checkpoint-1000/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,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\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>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\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\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-1000/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:754e4ae9d879853457ade197edaf9a4e09b3e2d2a23d8bffcf3f56a687083eb3
+size 6904

checkpoint-1000/zero_to_fp32.py

@@ -0,0 +1,674 @@
+#!/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 json
+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)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in 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}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_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)
+
+    # 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")
+        state_dict_split = split_torch_state_dict_into_shards(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
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard, output_path)
+
+    # 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-1500/README.md

@@ -0,0 +1,202 @@
+---
+base_model: Qwen/Qwen2.5-7B
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.14.0

checkpoint-1500/adapter_config.json

@@ -0,0 +1,37 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-7B",
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "gate_proj",
+    "down_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-1500/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1f1ce8c9922c384cf7af3046be04f4ed75efbc3066b8e3ee012e28782155f77
+size 80792880

checkpoint-1500/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-1500/latest

@@ -0,0 +1 @@
+global_step1501

checkpoint-1500/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34bcae41c589c7e4cab7b2ef263b878c90c2741404a6af11994dc31537b2319b
+size 14512

checkpoint-1500/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d05dc84075e8f7dd1191c36f3be9dda12073208e12f7d2cef433c38d6336774a
+size 14512

checkpoint-1500/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": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-1500/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896

checkpoint-1500/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,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\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>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\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\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-1500/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:754e4ae9d879853457ade197edaf9a4e09b3e2d2a23d8bffcf3f56a687083eb3
+size 6904

checkpoint-1500/zero_to_fp32.py

@@ -0,0 +1,674 @@
+#!/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 json
+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)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in 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}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_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)
+
+    # 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")
+        state_dict_split = split_torch_state_dict_into_shards(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
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard, output_path)
+
+    # 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-500/README.md

@@ -0,0 +1,202 @@
+---
+base_model: Qwen/Qwen2.5-7B
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.14.0

checkpoint-500/adapter_config.json

@@ -0,0 +1,37 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-7B",
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "gate_proj",
+    "down_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-500/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:888aea7ce26863acfdec85b6f9309e6828f7053858e3f88f6021c493e1e745c7
+size 80792880

checkpoint-500/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-500/latest

@@ -0,0 +1 @@
+global_step500

checkpoint-500/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8d6a959372d5e0c2ea025dd26c9d0ad2046fce19352056cae8074dcbd0a6fd4
+size 14512

checkpoint-500/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f68a37892a1b445d21bb35cc10bf7a058a6f9ec8c363f5ed156ff4f49d90fb6
+size 14512

checkpoint-500/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": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-500/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896

checkpoint-500/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,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\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>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\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\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-500/trainer_state.json

@@ -0,0 +1,3533 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.998377636340946,
+  "eval_steps": 500,
+  "global_step": 500,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.001996755272681892,
+      "grad_norm": 0.7412710189819336,
+      "learning_rate": 2.5000000000000004e-07,
+      "loss": 1.5868,
+      "step": 1
+    },
+    {
+      "epoch": 0.003993510545363784,
+      "grad_norm": 0.706846296787262,
+      "learning_rate": 5.000000000000001e-07,
+      "loss": 1.5439,
+      "step": 2
+    },
+    {
+      "epoch": 0.005990265818045676,
+      "grad_norm": 0.7040128707885742,
+      "learning_rate": 7.5e-07,
+      "loss": 1.5332,
+      "step": 3
+    },
+    {
+      "epoch": 0.007987021090727568,
+      "grad_norm": 0.7102969884872437,
+      "learning_rate": 1.0000000000000002e-06,
+      "loss": 1.5666,
+      "step": 4
+    },
+    {
+      "epoch": 0.00998377636340946,
+      "grad_norm": 0.7481162548065186,
+      "learning_rate": 1.25e-06,
+      "loss": 1.5532,
+      "step": 5
+    },
+    {
+      "epoch": 0.011980531636091352,
+      "grad_norm": 0.7318422198295593,
+      "learning_rate": 1.5e-06,
+      "loss": 1.5908,
+      "step": 6
+    },
+    {
+      "epoch": 0.013977286908773243,
+      "grad_norm": 0.6923463344573975,
+      "learning_rate": 1.75e-06,
+      "loss": 1.5761,
+      "step": 7
+    },
+    {
+      "epoch": 0.015974042181455136,
+      "grad_norm": 0.6836492419242859,
+      "learning_rate": 2.0000000000000003e-06,
+      "loss": 1.5314,
+      "step": 8
+    },
+    {
+      "epoch": 0.017970797454137027,
+      "grad_norm": 0.742129921913147,
+      "learning_rate": 2.25e-06,
+      "loss": 1.5662,
+      "step": 9
+    },
+    {
+      "epoch": 0.01996755272681892,
+      "grad_norm": 0.7376399636268616,
+      "learning_rate": 2.5e-06,
+      "loss": 1.5888,
+      "step": 10
+    },
+    {
+      "epoch": 0.02196430799950081,
+      "grad_norm": 0.7043939828872681,
+      "learning_rate": 2.7500000000000004e-06,
+      "loss": 1.5691,
+      "step": 11
+    },
+    {
+      "epoch": 0.023961063272182705,
+      "grad_norm": 0.8563888072967529,
+      "learning_rate": 3e-06,
+      "loss": 1.5439,
+      "step": 12
+    },
+    {
+      "epoch": 0.025957818544864595,
+      "grad_norm": 0.7369752526283264,
+      "learning_rate": 3.2500000000000002e-06,
+      "loss": 1.5735,
+      "step": 13
+    },
+    {
+      "epoch": 0.027954573817546485,
+      "grad_norm": 0.7235389351844788,
+      "learning_rate": 3.5e-06,
+      "loss": 1.5216,
+      "step": 14
+    },
+    {
+      "epoch": 0.02995132909022838,
+      "grad_norm": 0.7691176533699036,
+      "learning_rate": 3.7500000000000005e-06,
+      "loss": 1.5382,
+      "step": 15
+    },
+    {
+      "epoch": 0.03194808436291027,
+      "grad_norm": 1.0237406492233276,
+      "learning_rate": 4.000000000000001e-06,
+      "loss": 1.5465,
+      "step": 16
+    },
+    {
+      "epoch": 0.03394483963559216,
+      "grad_norm": 0.7893355488777161,
+      "learning_rate": 4.25e-06,
+      "loss": 1.5353,
+      "step": 17
+    },
+    {
+      "epoch": 0.035941594908274054,
+      "grad_norm": 1.089913010597229,
+      "learning_rate": 4.5e-06,
+      "loss": 1.5779,
+      "step": 18
+    },
+    {
+      "epoch": 0.03793835018095595,
+      "grad_norm": 0.7530987858772278,
+      "learning_rate": 4.75e-06,
+      "loss": 1.5248,
+      "step": 19
+    },
+    {
+      "epoch": 0.03993510545363784,
+      "grad_norm": 0.8638855814933777,
+      "learning_rate": 5e-06,
+      "loss": 1.5888,
+      "step": 20
+    },
+    {
+      "epoch": 0.04193186072631973,
+      "grad_norm": 0.9727922081947327,
+      "learning_rate": 5.2500000000000006e-06,
+      "loss": 1.5495,
+      "step": 21
+    },
+    {
+      "epoch": 0.04392861599900162,
+      "grad_norm": 0.9645998477935791,
+      "learning_rate": 5.500000000000001e-06,
+      "loss": 1.5365,
+      "step": 22
+    },
+    {
+      "epoch": 0.045925371271683516,
+      "grad_norm": 1.1736634969711304,
+      "learning_rate": 5.75e-06,
+      "loss": 1.5407,
+      "step": 23
+    },
+    {
+      "epoch": 0.04792212654436541,
+      "grad_norm": 1.091434121131897,
+      "learning_rate": 6e-06,
+      "loss": 1.5231,
+      "step": 24
+    },
+    {
+      "epoch": 0.049918881817047296,
+      "grad_norm": 1.1759644746780396,
+      "learning_rate": 6.25e-06,
+      "loss": 1.4901,
+      "step": 25
+    },
+    {
+      "epoch": 0.05191563708972919,
+      "grad_norm": 1.1277236938476562,
+      "learning_rate": 6.5000000000000004e-06,
+      "loss": 1.5326,
+      "step": 26
+    },
+    {
+      "epoch": 0.053912392362411084,
+      "grad_norm": 0.863665759563446,
+      "learning_rate": 6.750000000000001e-06,
+      "loss": 1.5185,
+      "step": 27
+    },
+    {
+      "epoch": 0.05590914763509297,
+      "grad_norm": 0.5838161706924438,
+      "learning_rate": 7e-06,
+      "loss": 1.5053,
+      "step": 28
+    },
+    {
+      "epoch": 0.057905902907774864,
+      "grad_norm": 0.541231095790863,
+      "learning_rate": 7.25e-06,
+      "loss": 1.4914,
+      "step": 29
+    },
+    {
+      "epoch": 0.05990265818045676,
+      "grad_norm": 0.5768315196037292,
+      "learning_rate": 7.500000000000001e-06,
+      "loss": 1.5208,
+      "step": 30
+    },
+    {
+      "epoch": 0.06189941345313865,
+      "grad_norm": 0.5609717965126038,
+      "learning_rate": 7.75e-06,
+      "loss": 1.4928,
+      "step": 31
+    },
+    {
+      "epoch": 0.06389616872582055,
+      "grad_norm": 0.5491043329238892,
+      "learning_rate": 8.000000000000001e-06,
+      "loss": 1.5208,
+      "step": 32
+    },
+    {
+      "epoch": 0.06589292399850244,
+      "grad_norm": 0.4885888695716858,
+      "learning_rate": 8.25e-06,
+      "loss": 1.4811,
+      "step": 33
+    },
+    {
+      "epoch": 0.06788967927118432,
+      "grad_norm": 0.5466718077659607,
+      "learning_rate": 8.5e-06,
+      "loss": 1.4608,
+      "step": 34
+    },
+    {
+      "epoch": 0.06988643454386621,
+      "grad_norm": 0.8744406700134277,
+      "learning_rate": 8.750000000000001e-06,
+      "loss": 1.465,
+      "step": 35
+    },
+    {
+      "epoch": 0.07188318981654811,
+      "grad_norm": 0.7382685542106628,
+      "learning_rate": 9e-06,
+      "loss": 1.454,
+      "step": 36
+    },
+    {
+      "epoch": 0.07387994508923,
+      "grad_norm": 0.6689203977584839,
+      "learning_rate": 9.250000000000001e-06,
+      "loss": 1.4707,
+      "step": 37
+    },
+    {
+      "epoch": 0.0758767003619119,
+      "grad_norm": 0.4725808799266815,
+      "learning_rate": 9.5e-06,
+      "loss": 1.4593,
+      "step": 38
+    },
+    {
+      "epoch": 0.07787345563459379,
+      "grad_norm": 0.4035033881664276,
+      "learning_rate": 9.75e-06,
+      "loss": 1.3955,
+      "step": 39
+    },
+    {
+      "epoch": 0.07987021090727568,
+      "grad_norm": 0.36325860023498535,
+      "learning_rate": 1e-05,
+      "loss": 1.4006,
+      "step": 40
+    },
+    {
+      "epoch": 0.08186696617995756,
+      "grad_norm": 0.33312714099884033,
+      "learning_rate": 9.993150684931508e-06,
+      "loss": 1.4212,
+      "step": 41
+    },
+    {
+      "epoch": 0.08386372145263946,
+      "grad_norm": 0.3342682719230652,
+      "learning_rate": 9.986301369863014e-06,
+      "loss": 1.4458,
+      "step": 42
+    },
+    {
+      "epoch": 0.08586047672532135,
+      "grad_norm": 0.4984380900859833,
+      "learning_rate": 9.979452054794521e-06,
+      "loss": 1.3925,
+      "step": 43
+    },
+    {
+      "epoch": 0.08785723199800324,
+      "grad_norm": 0.36669808626174927,
+      "learning_rate": 9.972602739726028e-06,
+      "loss": 1.3997,
+      "step": 44
+    },
+    {
+      "epoch": 0.08985398727068514,
+      "grad_norm": 0.3464302122592926,
+      "learning_rate": 9.965753424657536e-06,
+      "loss": 1.4019,
+      "step": 45
+    },
+    {
+      "epoch": 0.09185074254336703,
+      "grad_norm": 0.36931392550468445,
+      "learning_rate": 9.958904109589041e-06,
+      "loss": 1.3793,
+      "step": 46
+    },
+    {
+      "epoch": 0.09384749781604892,
+      "grad_norm": 0.35277390480041504,
+      "learning_rate": 9.952054794520548e-06,
+      "loss": 1.3854,
+      "step": 47
+    },
+    {
+      "epoch": 0.09584425308873082,
+      "grad_norm": 0.3395546078681946,
+      "learning_rate": 9.945205479452056e-06,
+      "loss": 1.3622,
+      "step": 48
+    },
+    {
+      "epoch": 0.0978410083614127,
+      "grad_norm": 0.37092125415802,
+      "learning_rate": 9.938356164383563e-06,
+      "loss": 1.347,
+      "step": 49
+    },
+    {
+      "epoch": 0.09983776363409459,
+      "grad_norm": 0.3518228232860565,
+      "learning_rate": 9.931506849315069e-06,
+      "loss": 1.3419,
+      "step": 50
+    },
+    {
+      "epoch": 0.10183451890677649,
+      "grad_norm": 0.3498891294002533,
+      "learning_rate": 9.924657534246576e-06,
+      "loss": 1.3374,
+      "step": 51
+    },
+    {
+      "epoch": 0.10383127417945838,
+      "grad_norm": 0.35601961612701416,
+      "learning_rate": 9.917808219178083e-06,
+      "loss": 1.3229,
+      "step": 52
+    },
+    {
+      "epoch": 0.10582802945214027,
+      "grad_norm": 0.3646593391895294,
+      "learning_rate": 9.91095890410959e-06,
+      "loss": 1.3361,
+      "step": 53
+    },
+    {
+      "epoch": 0.10782478472482217,
+      "grad_norm": 0.37503260374069214,
+      "learning_rate": 9.904109589041096e-06,
+      "loss": 1.3202,
+      "step": 54
+    },
+    {
+      "epoch": 0.10982153999750406,
+      "grad_norm": 0.36810755729675293,
+      "learning_rate": 9.897260273972603e-06,
+      "loss": 1.3012,
+      "step": 55
+    },
+    {
+      "epoch": 0.11181829527018594,
+      "grad_norm": 0.38846829533576965,
+      "learning_rate": 9.89041095890411e-06,
+      "loss": 1.2971,
+      "step": 56
+    },
+    {
+      "epoch": 0.11381505054286783,
+      "grad_norm": 0.3800796866416931,
+      "learning_rate": 9.883561643835618e-06,
+      "loss": 1.2568,
+      "step": 57
+    },
+    {
+      "epoch": 0.11581180581554973,
+      "grad_norm": 0.44192200899124146,
+      "learning_rate": 9.876712328767123e-06,
+      "loss": 1.2972,
+      "step": 58
+    },
+    {
+      "epoch": 0.11780856108823162,
+      "grad_norm": 0.4018876254558563,
+      "learning_rate": 9.86986301369863e-06,
+      "loss": 1.2712,
+      "step": 59
+    },
+    {
+      "epoch": 0.11980531636091352,
+      "grad_norm": 0.40182632207870483,
+      "learning_rate": 9.863013698630138e-06,
+      "loss": 1.2547,
+      "step": 60
+    },
+    {
+      "epoch": 0.12180207163359541,
+      "grad_norm": 0.40880608558654785,
+      "learning_rate": 9.856164383561645e-06,
+      "loss": 1.2776,
+      "step": 61
+    },
+    {
+      "epoch": 0.1237988269062773,
+      "grad_norm": 0.404816597700119,
+      "learning_rate": 9.849315068493151e-06,
+      "loss": 1.2258,
+      "step": 62
+    },
+    {
+      "epoch": 0.1257955821789592,
+      "grad_norm": 0.4100593030452728,
+      "learning_rate": 9.842465753424658e-06,
+      "loss": 1.228,
+      "step": 63
+    },
+    {
+      "epoch": 0.1277923374516411,
+      "grad_norm": 0.4166107773780823,
+      "learning_rate": 9.835616438356166e-06,
+      "loss": 1.1834,
+      "step": 64
+    },
+    {
+      "epoch": 0.12978909272432299,
+      "grad_norm": 0.44978272914886475,
+      "learning_rate": 9.828767123287673e-06,
+      "loss": 1.1874,
+      "step": 65
+    },
+    {
+      "epoch": 0.13178584799700488,
+      "grad_norm": 0.4174051582813263,
+      "learning_rate": 9.821917808219178e-06,
+      "loss": 1.1884,
+      "step": 66
+    },
+    {
+      "epoch": 0.13378260326968677,
+      "grad_norm": 0.453905314207077,
+      "learning_rate": 9.815068493150686e-06,
+      "loss": 1.1646,
+      "step": 67
+    },
+    {
+      "epoch": 0.13577935854236864,
+      "grad_norm": 0.5017035007476807,
+      "learning_rate": 9.808219178082193e-06,
+      "loss": 1.1491,
+      "step": 68
+    },
+    {
+      "epoch": 0.13777611381505053,
+      "grad_norm": 0.44494616985321045,
+      "learning_rate": 9.8013698630137e-06,
+      "loss": 1.1516,
+      "step": 69
+    },
+    {
+      "epoch": 0.13977286908773243,
+      "grad_norm": 0.43786314129829407,
+      "learning_rate": 9.794520547945206e-06,
+      "loss": 1.1131,
+      "step": 70
+    },
+    {
+      "epoch": 0.14176962436041432,
+      "grad_norm": 0.4679386615753174,
+      "learning_rate": 9.787671232876713e-06,
+      "loss": 1.1154,
+      "step": 71
+    },
+    {
+      "epoch": 0.14376637963309621,
+      "grad_norm": 0.4783805310726166,
+      "learning_rate": 9.78082191780822e-06,
+      "loss": 1.1102,
+      "step": 72
+    },
+    {
+      "epoch": 0.1457631349057781,
+      "grad_norm": 0.4883531332015991,
+      "learning_rate": 9.773972602739726e-06,
+      "loss": 1.0969,
+      "step": 73
+    },
+    {
+      "epoch": 0.14775989017846,
+      "grad_norm": 0.497222363948822,
+      "learning_rate": 9.767123287671233e-06,
+      "loss": 1.0894,
+      "step": 74
+    },
+    {
+      "epoch": 0.1497566454511419,
+      "grad_norm": 0.5368260741233826,
+      "learning_rate": 9.76027397260274e-06,
+      "loss": 1.0872,
+      "step": 75
+    },
+    {
+      "epoch": 0.1517534007238238,
+      "grad_norm": 0.5358818173408508,
+      "learning_rate": 9.753424657534248e-06,
+      "loss": 1.0771,
+      "step": 76
+    },
+    {
+      "epoch": 0.15375015599650568,
+      "grad_norm": 0.5469568371772766,
+      "learning_rate": 9.746575342465753e-06,
+      "loss": 1.0142,
+      "step": 77
+    },
+    {
+      "epoch": 0.15574691126918758,
+      "grad_norm": 0.5410879254341125,
+      "learning_rate": 9.73972602739726e-06,
+      "loss": 1.0019,
+      "step": 78
+    },
+    {
+      "epoch": 0.15774366654186947,
+      "grad_norm": 0.5405011177062988,
+      "learning_rate": 9.732876712328768e-06,
+      "loss": 0.9998,
+      "step": 79
+    },
+    {
+      "epoch": 0.15974042181455136,
+      "grad_norm": 0.532302737236023,
+      "learning_rate": 9.726027397260275e-06,
+      "loss": 0.9659,
+      "step": 80
+    },
+    {
+      "epoch": 0.16173717708723326,
+      "grad_norm": 0.5053435564041138,
+      "learning_rate": 9.719178082191781e-06,
+      "loss": 0.9679,
+      "step": 81
+    },
+    {
+      "epoch": 0.16373393235991512,
+      "grad_norm": 0.49565696716308594,
+      "learning_rate": 9.712328767123288e-06,
+      "loss": 0.9246,
+      "step": 82
+    },
+    {
+      "epoch": 0.16573068763259702,
+      "grad_norm": 0.5060250759124756,
+      "learning_rate": 9.705479452054795e-06,
+      "loss": 0.938,
+      "step": 83
+    },
+    {
+      "epoch": 0.1677274429052789,
+      "grad_norm": 0.4905273914337158,
+      "learning_rate": 9.698630136986303e-06,
+      "loss": 0.9137,
+      "step": 84
+    },
+    {
+      "epoch": 0.1697241981779608,
+      "grad_norm": 0.6356221437454224,
+      "learning_rate": 9.691780821917808e-06,
+      "loss": 0.8919,
+      "step": 85
+    },
+    {
+      "epoch": 0.1717209534506427,
+      "grad_norm": 0.5084801316261292,
+      "learning_rate": 9.684931506849316e-06,
+      "loss": 0.9097,
+      "step": 86
+    },
+    {
+      "epoch": 0.1737177087233246,
+      "grad_norm": 0.44582366943359375,
+      "learning_rate": 9.678082191780823e-06,
+      "loss": 0.849,
+      "step": 87
+    },
+    {
+      "epoch": 0.1757144639960065,
+      "grad_norm": 0.45911890268325806,
+      "learning_rate": 9.67123287671233e-06,
+      "loss": 0.8481,
+      "step": 88
+    },
+    {
+      "epoch": 0.17771121926868838,
+      "grad_norm": 0.44006654620170593,
+      "learning_rate": 9.664383561643836e-06,
+      "loss": 0.8248,
+      "step": 89
+    },
+    {
+      "epoch": 0.17970797454137027,
+      "grad_norm": 0.4369649589061737,
+      "learning_rate": 9.657534246575343e-06,
+      "loss": 0.8159,
+      "step": 90
+    },
+    {
+      "epoch": 0.18170472981405217,
+      "grad_norm": 0.407516747713089,
+      "learning_rate": 9.65068493150685e-06,
+      "loss": 0.8087,
+      "step": 91
+    },
+    {
+      "epoch": 0.18370148508673406,
+      "grad_norm": 0.403859943151474,
+      "learning_rate": 9.643835616438358e-06,
+      "loss": 0.8072,
+      "step": 92
+    },
+    {
+      "epoch": 0.18569824035941596,
+      "grad_norm": 0.39687636494636536,
+      "learning_rate": 9.636986301369863e-06,
+      "loss": 0.7441,
+      "step": 93
+    },
+    {
+      "epoch": 0.18769499563209785,
+      "grad_norm": 0.3767126202583313,
+      "learning_rate": 9.63013698630137e-06,
+      "loss": 0.7648,
+      "step": 94
+    },
+    {
+      "epoch": 0.18969175090477974,
+      "grad_norm": 0.361546128988266,
+      "learning_rate": 9.623287671232878e-06,
+      "loss": 0.7611,
+      "step": 95
+    },
+    {
+      "epoch": 0.19168850617746164,
+      "grad_norm": 0.34515851736068726,
+      "learning_rate": 9.616438356164385e-06,
+      "loss": 0.7625,
+      "step": 96
+    },
+    {
+      "epoch": 0.1936852614501435,
+      "grad_norm": 0.3361088037490845,
+      "learning_rate": 9.60958904109589e-06,
+      "loss": 0.7827,
+      "step": 97
+    },
+    {
+      "epoch": 0.1956820167228254,
+      "grad_norm": 0.3142891228199005,
+      "learning_rate": 9.602739726027398e-06,
+      "loss": 0.7339,
+      "step": 98
+    },
+    {
+      "epoch": 0.1976787719955073,
+      "grad_norm": 0.3227478265762329,
+      "learning_rate": 9.595890410958905e-06,
+      "loss": 0.7389,
+      "step": 99
+    },
+    {
+      "epoch": 0.19967552726818918,
+      "grad_norm": 0.3079572319984436,
+      "learning_rate": 9.589041095890411e-06,
+      "loss": 0.7268,
+      "step": 100
+    },
+    {
+      "epoch": 0.20167228254087108,
+      "grad_norm": 0.3069862127304077,
+      "learning_rate": 9.582191780821918e-06,
+      "loss": 0.7198,
+      "step": 101
+    },
+    {
+      "epoch": 0.20366903781355297,
+      "grad_norm": 0.29050567746162415,
+      "learning_rate": 9.575342465753425e-06,
+      "loss": 0.7549,
+      "step": 102
+    },
+    {
+      "epoch": 0.20566579308623487,
+      "grad_norm": 0.29106447100639343,
+      "learning_rate": 9.568493150684933e-06,
+      "loss": 0.7167,
+      "step": 103
+    },
+    {
+      "epoch": 0.20766254835891676,
+      "grad_norm": 0.2757619619369507,
+      "learning_rate": 9.561643835616438e-06,
+      "loss": 0.7239,
+      "step": 104
+    },
+    {
+      "epoch": 0.20965930363159865,
+      "grad_norm": 0.2644850015640259,
+      "learning_rate": 9.554794520547946e-06,
+      "loss": 0.7028,
+      "step": 105
+    },
+    {
+      "epoch": 0.21165605890428055,
+      "grad_norm": 0.2568458318710327,
+      "learning_rate": 9.547945205479453e-06,
+      "loss": 0.7312,
+      "step": 106
+    },
+    {
+      "epoch": 0.21365281417696244,
+      "grad_norm": 0.24979732930660248,
+      "learning_rate": 9.54109589041096e-06,
+      "loss": 0.6933,
+      "step": 107
+    },
+    {
+      "epoch": 0.21564956944964433,
+      "grad_norm": 0.2434425950050354,
+      "learning_rate": 9.534246575342466e-06,
+      "loss": 0.6824,
+      "step": 108
+    },
+    {
+      "epoch": 0.21764632472232623,
+      "grad_norm": 0.27341383695602417,
+      "learning_rate": 9.527397260273975e-06,
+      "loss": 0.6832,
+      "step": 109
+    },
+    {
+      "epoch": 0.21964307999500812,
+      "grad_norm": 0.22272831201553345,
+      "learning_rate": 9.52054794520548e-06,
+      "loss": 0.6889,
+      "step": 110
+    },
+    {
+      "epoch": 0.22163983526769002,
+      "grad_norm": 0.2114173024892807,
+      "learning_rate": 9.513698630136988e-06,
+      "loss": 0.7047,
+      "step": 111
+    },
+    {
+      "epoch": 0.22363659054037188,
+      "grad_norm": 0.20138710737228394,
+      "learning_rate": 9.506849315068493e-06,
+      "loss": 0.6926,
+      "step": 112
+    },
+    {
+      "epoch": 0.22563334581305378,
+      "grad_norm": 0.18181370198726654,
+      "learning_rate": 9.5e-06,
+      "loss": 0.7121,
+      "step": 113
+    },
+    {
+      "epoch": 0.22763010108573567,
+      "grad_norm": 0.16511252522468567,
+      "learning_rate": 9.493150684931508e-06,
+      "loss": 0.6879,
+      "step": 114
+    },
+    {
+      "epoch": 0.22962685635841756,
+      "grad_norm": 0.15935856103897095,
+      "learning_rate": 9.486301369863015e-06,
+      "loss": 0.692,
+      "step": 115
+    },
+    {
+      "epoch": 0.23162361163109946,
+      "grad_norm": 0.15437360107898712,
+      "learning_rate": 9.47945205479452e-06,
+      "loss": 0.6365,
+      "step": 116
+    },
+    {
+      "epoch": 0.23362036690378135,
+      "grad_norm": 0.15790362656116486,
+      "learning_rate": 9.472602739726028e-06,
+      "loss": 0.6734,
+      "step": 117
+    },
+    {
+      "epoch": 0.23561712217646325,
+      "grad_norm": 0.1540212631225586,
+      "learning_rate": 9.465753424657535e-06,
+      "loss": 0.6785,
+      "step": 118
+    },
+    {
+      "epoch": 0.23761387744914514,
+      "grad_norm": 0.14174145460128784,
+      "learning_rate": 9.458904109589043e-06,
+      "loss": 0.662,
+      "step": 119
+    },
+    {
+      "epoch": 0.23961063272182703,
+      "grad_norm": 0.13557200133800507,
+      "learning_rate": 9.452054794520548e-06,
+      "loss": 0.6576,
+      "step": 120
+    },
+    {
+      "epoch": 0.24160738799450893,
+      "grad_norm": 0.13709504902362823,
+      "learning_rate": 9.445205479452055e-06,
+      "loss": 0.6776,
+      "step": 121
+    },
+    {
+      "epoch": 0.24360414326719082,
+      "grad_norm": 0.13799452781677246,
+      "learning_rate": 9.438356164383563e-06,
+      "loss": 0.6678,
+      "step": 122
+    },
+    {
+      "epoch": 0.24560089853987271,
+      "grad_norm": 0.13457225263118744,
+      "learning_rate": 9.43150684931507e-06,
+      "loss": 0.678,
+      "step": 123
+    },
+    {
+      "epoch": 0.2475976538125546,
+      "grad_norm": 0.12947948276996613,
+      "learning_rate": 9.424657534246576e-06,
+      "loss": 0.6735,
+      "step": 124
+    },
+    {
+      "epoch": 0.2495944090852365,
+      "grad_norm": 0.12308862805366516,
+      "learning_rate": 9.417808219178083e-06,
+      "loss": 0.6912,
+      "step": 125
+    },
+    {
+      "epoch": 0.2515911643579184,
+      "grad_norm": 0.14674226939678192,
+      "learning_rate": 9.41095890410959e-06,
+      "loss": 0.6533,
+      "step": 126
+    },
+    {
+      "epoch": 0.2535879196306003,
+      "grad_norm": 0.12698639929294586,
+      "learning_rate": 9.404109589041097e-06,
+      "loss": 0.6876,
+      "step": 127
+    },
+    {
+      "epoch": 0.2555846749032822,
+      "grad_norm": 0.11984587460756302,
+      "learning_rate": 9.397260273972603e-06,
+      "loss": 0.6555,
+      "step": 128
+    },
+    {
+      "epoch": 0.2575814301759641,
+      "grad_norm": 0.11617796868085861,
+      "learning_rate": 9.39041095890411e-06,
+      "loss": 0.6641,
+      "step": 129
+    },
+    {
+      "epoch": 0.25957818544864597,
+      "grad_norm": 0.11317376047372818,
+      "learning_rate": 9.383561643835618e-06,
+      "loss": 0.6563,
+      "step": 130
+    },
+    {
+      "epoch": 0.26157494072132786,
+      "grad_norm": 0.12450871616601944,
+      "learning_rate": 9.376712328767123e-06,
+      "loss": 0.6778,
+      "step": 131
+    },
+    {
+      "epoch": 0.26357169599400976,
+      "grad_norm": 0.11382495611906052,
+      "learning_rate": 9.36986301369863e-06,
+      "loss": 0.6472,
+      "step": 132
+    },
+    {
+      "epoch": 0.26556845126669165,
+      "grad_norm": 0.11967656016349792,
+      "learning_rate": 9.363013698630138e-06,
+      "loss": 0.6584,
+      "step": 133
+    },
+    {
+      "epoch": 0.26756520653937355,
+      "grad_norm": 0.10746151208877563,
+      "learning_rate": 9.356164383561645e-06,
+      "loss": 0.6605,
+      "step": 134
+    },
+    {
+      "epoch": 0.2695619618120554,
+      "grad_norm": 0.11343806982040405,
+      "learning_rate": 9.34931506849315e-06,
+      "loss": 0.6699,
+      "step": 135
+    },
+    {
+      "epoch": 0.2715587170847373,
+      "grad_norm": 0.10874643176794052,
+      "learning_rate": 9.342465753424658e-06,
+      "loss": 0.6396,
+      "step": 136
+    },
+    {
+      "epoch": 0.27355547235741917,
+      "grad_norm": 0.11221782863140106,
+      "learning_rate": 9.335616438356165e-06,
+      "loss": 0.6517,
+      "step": 137
+    },
+    {
+      "epoch": 0.27555222763010107,
+      "grad_norm": 0.12015150487422943,
+      "learning_rate": 9.328767123287673e-06,
+      "loss": 0.6561,
+      "step": 138
+    },
+    {
+      "epoch": 0.27754898290278296,
+      "grad_norm": 0.10812759399414062,
+      "learning_rate": 9.321917808219178e-06,
+      "loss": 0.6505,
+      "step": 139
+    },
+    {
+      "epoch": 0.27954573817546485,
+      "grad_norm": 0.10647746920585632,
+      "learning_rate": 9.315068493150685e-06,
+      "loss": 0.6433,
+      "step": 140
+    },
+    {
+      "epoch": 0.28154249344814675,
+      "grad_norm": 0.10739017277956009,
+      "learning_rate": 9.308219178082193e-06,
+      "loss": 0.6504,
+      "step": 141
+    },
+    {
+      "epoch": 0.28353924872082864,
+      "grad_norm": 0.09796225279569626,
+      "learning_rate": 9.3013698630137e-06,
+      "loss": 0.6766,
+      "step": 142
+    },
+    {
+      "epoch": 0.28553600399351053,
+      "grad_norm": 0.10549402981996536,
+      "learning_rate": 9.294520547945206e-06,
+      "loss": 0.6401,
+      "step": 143
+    },
+    {
+      "epoch": 0.28753275926619243,
+      "grad_norm": 0.10363396257162094,
+      "learning_rate": 9.287671232876713e-06,
+      "loss": 0.6711,
+      "step": 144
+    },
+    {
+      "epoch": 0.2895295145388743,
+      "grad_norm": 0.10072599351406097,
+      "learning_rate": 9.28082191780822e-06,
+      "loss": 0.6294,
+      "step": 145
+    },
+    {
+      "epoch": 0.2915262698115562,
+      "grad_norm": 0.09450184553861618,
+      "learning_rate": 9.273972602739727e-06,
+      "loss": 0.6511,
+      "step": 146
+    },
+    {
+      "epoch": 0.2935230250842381,
+      "grad_norm": 0.09612888097763062,
+      "learning_rate": 9.267123287671233e-06,
+      "loss": 0.6441,
+      "step": 147
+    },
+    {
+      "epoch": 0.29551978035692,
+      "grad_norm": 0.09231892973184586,
+      "learning_rate": 9.26027397260274e-06,
+      "loss": 0.6474,
+      "step": 148
+    },
+    {
+      "epoch": 0.2975165356296019,
+      "grad_norm": 0.0949755609035492,
+      "learning_rate": 9.253424657534248e-06,
+      "loss": 0.633,
+      "step": 149
+    },
+    {
+      "epoch": 0.2995132909022838,
+      "grad_norm": 0.09723283350467682,
+      "learning_rate": 9.246575342465755e-06,
+      "loss": 0.641,
+      "step": 150
+    },
+    {
+      "epoch": 0.3015100461749657,
+      "grad_norm": 0.09414070099592209,
+      "learning_rate": 9.23972602739726e-06,
+      "loss": 0.6646,
+      "step": 151
+    },
+    {
+      "epoch": 0.3035068014476476,
+      "grad_norm": 0.0909348651766777,
+      "learning_rate": 9.232876712328768e-06,
+      "loss": 0.6629,
+      "step": 152
+    },
+    {
+      "epoch": 0.30550355672032947,
+      "grad_norm": 0.09564705193042755,
+      "learning_rate": 9.226027397260275e-06,
+      "loss": 0.6727,
+      "step": 153
+    },
+    {
+      "epoch": 0.30750031199301137,
+      "grad_norm": 0.09434914588928223,
+      "learning_rate": 9.219178082191782e-06,
+      "loss": 0.6885,
+      "step": 154
+    },
+    {
+      "epoch": 0.30949706726569326,
+      "grad_norm": 0.1040414348244667,
+      "learning_rate": 9.212328767123288e-06,
+      "loss": 0.6516,
+      "step": 155
+    },
+    {
+      "epoch": 0.31149382253837515,
+      "grad_norm": 0.08882972598075867,
+      "learning_rate": 9.205479452054795e-06,
+      "loss": 0.6564,
+      "step": 156
+    },
+    {
+      "epoch": 0.31349057781105705,
+      "grad_norm": 0.09209997206926346,
+      "learning_rate": 9.198630136986302e-06,
+      "loss": 0.6756,
+      "step": 157
+    },
+    {
+      "epoch": 0.31548733308373894,
+      "grad_norm": 0.0916871652007103,
+      "learning_rate": 9.19178082191781e-06,
+      "loss": 0.6568,
+      "step": 158
+    },
+    {
+      "epoch": 0.31748408835642083,
+      "grad_norm": 0.08956406265497208,
+      "learning_rate": 9.184931506849315e-06,
+      "loss": 0.6464,
+      "step": 159
+    },
+    {
+      "epoch": 0.31948084362910273,
+      "grad_norm": 0.08891318738460541,
+      "learning_rate": 9.178082191780823e-06,
+      "loss": 0.6665,
+      "step": 160
+    },
+    {
+      "epoch": 0.3214775989017846,
+      "grad_norm": 0.08891268819570541,
+      "learning_rate": 9.17123287671233e-06,
+      "loss": 0.6491,
+      "step": 161
+    },
+    {
+      "epoch": 0.3234743541744665,
+      "grad_norm": 0.08686690032482147,
+      "learning_rate": 9.164383561643836e-06,
+      "loss": 0.6221,
+      "step": 162
+    },
+    {
+      "epoch": 0.3254711094471484,
+      "grad_norm": 0.0884203389286995,
+      "learning_rate": 9.157534246575343e-06,
+      "loss": 0.6676,
+      "step": 163
+    },
+    {
+      "epoch": 0.32746786471983025,
+      "grad_norm": 0.09307505935430527,
+      "learning_rate": 9.15068493150685e-06,
+      "loss": 0.6568,
+      "step": 164
+    },
+    {
+      "epoch": 0.32946461999251214,
+      "grad_norm": 0.08617405593395233,
+      "learning_rate": 9.143835616438357e-06,
+      "loss": 0.6641,
+      "step": 165
+    },
+    {
+      "epoch": 0.33146137526519404,
+      "grad_norm": 0.09388285130262375,
+      "learning_rate": 9.136986301369863e-06,
+      "loss": 0.6617,
+      "step": 166
+    },
+    {
+      "epoch": 0.33345813053787593,
+      "grad_norm": 0.09413463622331619,
+      "learning_rate": 9.130136986301372e-06,
+      "loss": 0.6669,
+      "step": 167
+    },
+    {
+      "epoch": 0.3354548858105578,
+      "grad_norm": 0.09485959261655807,
+      "learning_rate": 9.123287671232878e-06,
+      "loss": 0.6408,
+      "step": 168
+    },
+    {
+      "epoch": 0.3374516410832397,
+      "grad_norm": 0.0874333456158638,
+      "learning_rate": 9.116438356164385e-06,
+      "loss": 0.656,
+      "step": 169
+    },
+    {
+      "epoch": 0.3394483963559216,
+      "grad_norm": 0.08628030866384506,
+      "learning_rate": 9.10958904109589e-06,
+      "loss": 0.6534,
+      "step": 170
+    },
+    {
+      "epoch": 0.3414451516286035,
+      "grad_norm": 0.0860527902841568,
+      "learning_rate": 9.102739726027398e-06,
+      "loss": 0.6417,
+      "step": 171
+    },
+    {
+      "epoch": 0.3434419069012854,
+      "grad_norm": 0.08399222046136856,
+      "learning_rate": 9.095890410958905e-06,
+      "loss": 0.6233,
+      "step": 172
+    },
+    {
+      "epoch": 0.3454386621739673,
+      "grad_norm": 0.08794861286878586,
+      "learning_rate": 9.089041095890412e-06,
+      "loss": 0.6612,
+      "step": 173
+    },
+    {
+      "epoch": 0.3474354174466492,
+      "grad_norm": 0.08727024495601654,
+      "learning_rate": 9.082191780821918e-06,
+      "loss": 0.6725,
+      "step": 174
+    },
+    {
+      "epoch": 0.3494321727193311,
+      "grad_norm": 0.08527853339910507,
+      "learning_rate": 9.075342465753425e-06,
+      "loss": 0.6381,
+      "step": 175
+    },
+    {
+      "epoch": 0.351428927992013,
+      "grad_norm": 0.08497577905654907,
+      "learning_rate": 9.068493150684932e-06,
+      "loss": 0.6353,
+      "step": 176
+    },
+    {
+      "epoch": 0.35342568326469487,
+      "grad_norm": 0.08887533843517303,
+      "learning_rate": 9.06164383561644e-06,
+      "loss": 0.6392,
+      "step": 177
+    },
+    {
+      "epoch": 0.35542243853737676,
+      "grad_norm": 0.08368115872144699,
+      "learning_rate": 9.054794520547945e-06,
+      "loss": 0.6041,
+      "step": 178
+    },
+    {
+      "epoch": 0.35741919381005866,
+      "grad_norm": 0.08194791525602341,
+      "learning_rate": 9.047945205479453e-06,
+      "loss": 0.6507,
+      "step": 179
+    },
+    {
+      "epoch": 0.35941594908274055,
+      "grad_norm": 0.07991446554660797,
+      "learning_rate": 9.04109589041096e-06,
+      "loss": 0.6309,
+      "step": 180
+    },
+    {
+      "epoch": 0.36141270435542244,
+      "grad_norm": 0.08230883628129959,
+      "learning_rate": 9.034246575342467e-06,
+      "loss": 0.6348,
+      "step": 181
+    },
+    {
+      "epoch": 0.36340945962810434,
+      "grad_norm": 0.08559627085924149,
+      "learning_rate": 9.027397260273973e-06,
+      "loss": 0.6358,
+      "step": 182
+    },
+    {
+      "epoch": 0.36540621490078623,
+      "grad_norm": 0.08570659905672073,
+      "learning_rate": 9.02054794520548e-06,
+      "loss": 0.5993,
+      "step": 183
+    },
+    {
+      "epoch": 0.3674029701734681,
+      "grad_norm": 0.08620599657297134,
+      "learning_rate": 9.013698630136987e-06,
+      "loss": 0.6587,
+      "step": 184
+    },
+    {
+      "epoch": 0.36939972544615,
+      "grad_norm": 0.0859847143292427,
+      "learning_rate": 9.006849315068495e-06,
+      "loss": 0.6533,
+      "step": 185
+    },
+    {
+      "epoch": 0.3713964807188319,
+      "grad_norm": 0.08272071182727814,
+      "learning_rate": 9e-06,
+      "loss": 0.6121,
+      "step": 186
+    },
+    {
+      "epoch": 0.3733932359915138,
+      "grad_norm": 0.08778800070285797,
+      "learning_rate": 8.993150684931508e-06,
+      "loss": 0.6396,
+      "step": 187
+    },
+    {
+      "epoch": 0.3753899912641957,
+      "grad_norm": 0.08992988616228104,
+      "learning_rate": 8.986301369863015e-06,
+      "loss": 0.6381,
+      "step": 188
+    },
+    {
+      "epoch": 0.3773867465368776,
+      "grad_norm": 0.08496169745922089,
+      "learning_rate": 8.97945205479452e-06,
+      "loss": 0.6475,
+      "step": 189
+    },
+    {
+      "epoch": 0.3793835018095595,
+      "grad_norm": 0.08212533593177795,
+      "learning_rate": 8.972602739726028e-06,
+      "loss": 0.6709,
+      "step": 190
+    },
+    {
+      "epoch": 0.3813802570822414,
+      "grad_norm": 0.08200209587812424,
+      "learning_rate": 8.965753424657535e-06,
+      "loss": 0.6338,
+      "step": 191
+    },
+    {
+      "epoch": 0.3833770123549233,
+      "grad_norm": 0.09153922647237778,
+      "learning_rate": 8.958904109589042e-06,
+      "loss": 0.6158,
+      "step": 192
+    },
+    {
+      "epoch": 0.38537376762760517,
+      "grad_norm": 0.08760355412960052,
+      "learning_rate": 8.952054794520548e-06,
+      "loss": 0.6041,
+      "step": 193
+    },
+    {
+      "epoch": 0.387370522900287,
+      "grad_norm": 0.08381730318069458,
+      "learning_rate": 8.945205479452055e-06,
+      "loss": 0.6337,
+      "step": 194
+    },
+    {
+      "epoch": 0.3893672781729689,
+      "grad_norm": 0.08099182695150375,
+      "learning_rate": 8.938356164383562e-06,
+      "loss": 0.6073,
+      "step": 195
+    },
+    {
+      "epoch": 0.3913640334456508,
+      "grad_norm": 0.08702554553747177,
+      "learning_rate": 8.93150684931507e-06,
+      "loss": 0.6497,
+      "step": 196
+    },
+    {
+      "epoch": 0.3933607887183327,
+      "grad_norm": 0.08087313920259476,
+      "learning_rate": 8.924657534246575e-06,
+      "loss": 0.6036,
+      "step": 197
+    },
+    {
+      "epoch": 0.3953575439910146,
+      "grad_norm": 0.08589271456003189,
+      "learning_rate": 8.917808219178083e-06,
+      "loss": 0.6498,
+      "step": 198
+    },
+    {
+      "epoch": 0.3973542992636965,
+      "grad_norm": 0.0862244963645935,
+      "learning_rate": 8.91095890410959e-06,
+      "loss": 0.6331,
+      "step": 199
+    },
+    {
+      "epoch": 0.39935105453637837,
+      "grad_norm": 0.0803682953119278,
+      "learning_rate": 8.904109589041097e-06,
+      "loss": 0.6508,
+      "step": 200
+    },
+    {
+      "epoch": 0.40134780980906026,
+      "grad_norm": 0.07755322754383087,
+      "learning_rate": 8.897260273972603e-06,
+      "loss": 0.6409,
+      "step": 201
+    },
+    {
+      "epoch": 0.40334456508174216,
+      "grad_norm": 0.08488644659519196,
+      "learning_rate": 8.89041095890411e-06,
+      "loss": 0.6293,
+      "step": 202
+    },
+    {
+      "epoch": 0.40534132035442405,
+      "grad_norm": 0.08399825543165207,
+      "learning_rate": 8.883561643835617e-06,
+      "loss": 0.6517,
+      "step": 203
+    },
+    {
+      "epoch": 0.40733807562710594,
+      "grad_norm": 0.093985915184021,
+      "learning_rate": 8.876712328767125e-06,
+      "loss": 0.6482,
+      "step": 204
+    },
+    {
+      "epoch": 0.40933483089978784,
+      "grad_norm": 0.08805140852928162,
+      "learning_rate": 8.86986301369863e-06,
+      "loss": 0.6415,
+      "step": 205
+    },
+    {
+      "epoch": 0.41133158617246973,
+      "grad_norm": 0.08228921890258789,
+      "learning_rate": 8.863013698630137e-06,
+      "loss": 0.658,
+      "step": 206
+    },
+    {
+      "epoch": 0.4133283414451516,
+      "grad_norm": 0.08415067940950394,
+      "learning_rate": 8.856164383561645e-06,
+      "loss": 0.6492,
+      "step": 207
+    },
+    {
+      "epoch": 0.4153250967178335,
+      "grad_norm": 0.08831077814102173,
+      "learning_rate": 8.849315068493152e-06,
+      "loss": 0.6421,
+      "step": 208
+    },
+    {
+      "epoch": 0.4173218519905154,
+      "grad_norm": 0.07943451404571533,
+      "learning_rate": 8.842465753424658e-06,
+      "loss": 0.6445,
+      "step": 209
+    },
+    {
+      "epoch": 0.4193186072631973,
+      "grad_norm": 0.08044181019067764,
+      "learning_rate": 8.835616438356165e-06,
+      "loss": 0.6753,
+      "step": 210
+    },
+    {
+      "epoch": 0.4213153625358792,
+      "grad_norm": 0.08346639573574066,
+      "learning_rate": 8.828767123287672e-06,
+      "loss": 0.6702,
+      "step": 211
+    },
+    {
+      "epoch": 0.4233121178085611,
+      "grad_norm": 0.08326448500156403,
+      "learning_rate": 8.82191780821918e-06,
+      "loss": 0.6206,
+      "step": 212
+    },
+    {
+      "epoch": 0.425308873081243,
+      "grad_norm": 0.08394528925418854,
+      "learning_rate": 8.815068493150685e-06,
+      "loss": 0.6387,
+      "step": 213
+    },
+    {
+      "epoch": 0.4273056283539249,
+      "grad_norm": 0.07964486628770828,
+      "learning_rate": 8.808219178082192e-06,
+      "loss": 0.6416,
+      "step": 214
+    },
+    {
+      "epoch": 0.4293023836266068,
+      "grad_norm": 0.08593829721212387,
+      "learning_rate": 8.8013698630137e-06,
+      "loss": 0.6467,
+      "step": 215
+    },
+    {
+      "epoch": 0.43129913889928867,
+      "grad_norm": 0.08312725275754929,
+      "learning_rate": 8.794520547945207e-06,
+      "loss": 0.6337,
+      "step": 216
+    },
+    {
+      "epoch": 0.43329589417197056,
+      "grad_norm": 0.07907330989837646,
+      "learning_rate": 8.787671232876713e-06,
+      "loss": 0.6182,
+      "step": 217
+    },
+    {
+      "epoch": 0.43529264944465246,
+      "grad_norm": 0.0807236060500145,
+      "learning_rate": 8.78082191780822e-06,
+      "loss": 0.648,
+      "step": 218
+    },
+    {
+      "epoch": 0.43728940471733435,
+      "grad_norm": 0.08214370906352997,
+      "learning_rate": 8.773972602739727e-06,
+      "loss": 0.6159,
+      "step": 219
+    },
+    {
+      "epoch": 0.43928615999001625,
+      "grad_norm": 0.07902273535728455,
+      "learning_rate": 8.767123287671233e-06,
+      "loss": 0.6209,
+      "step": 220
+    },
+    {
+      "epoch": 0.44128291526269814,
+      "grad_norm": 0.08463626354932785,
+      "learning_rate": 8.76027397260274e-06,
+      "loss": 0.6388,
+      "step": 221
+    },
+    {
+      "epoch": 0.44327967053538003,
+      "grad_norm": 0.08723177760839462,
+      "learning_rate": 8.753424657534247e-06,
+      "loss": 0.6639,
+      "step": 222
+    },
+    {
+      "epoch": 0.4452764258080619,
+      "grad_norm": 0.08998522907495499,
+      "learning_rate": 8.746575342465755e-06,
+      "loss": 0.6491,
+      "step": 223
+    },
+    {
+      "epoch": 0.44727318108074376,
+      "grad_norm": 0.08876420557498932,
+      "learning_rate": 8.73972602739726e-06,
+      "loss": 0.6434,
+      "step": 224
+    },
+    {
+      "epoch": 0.44926993635342566,
+      "grad_norm": 0.08207367360591888,
+      "learning_rate": 8.732876712328769e-06,
+      "loss": 0.6159,
+      "step": 225
+    },
+    {
+      "epoch": 0.45126669162610755,
+      "grad_norm": 0.08738122880458832,
+      "learning_rate": 8.726027397260275e-06,
+      "loss": 0.6383,
+      "step": 226
+    },
+    {
+      "epoch": 0.45326344689878945,
+      "grad_norm": 0.08093910664319992,
+      "learning_rate": 8.719178082191782e-06,
+      "loss": 0.6863,
+      "step": 227
+    },
+    {
+      "epoch": 0.45526020217147134,
+      "grad_norm": 0.0844852477312088,
+      "learning_rate": 8.712328767123288e-06,
+      "loss": 0.6308,
+      "step": 228
+    },
+    {
+      "epoch": 0.45725695744415323,
+      "grad_norm": 0.0834694430232048,
+      "learning_rate": 8.705479452054795e-06,
+      "loss": 0.647,
+      "step": 229
+    },
+    {
+      "epoch": 0.4592537127168351,
+      "grad_norm": 0.08448044210672379,
+      "learning_rate": 8.698630136986302e-06,
+      "loss": 0.6408,
+      "step": 230
+    },
+    {
+      "epoch": 0.461250467989517,
+      "grad_norm": 0.0820637121796608,
+      "learning_rate": 8.69178082191781e-06,
+      "loss": 0.6268,
+      "step": 231
+    },
+    {
+      "epoch": 0.4632472232621989,
+      "grad_norm": 0.08034293353557587,
+      "learning_rate": 8.684931506849315e-06,
+      "loss": 0.6308,
+      "step": 232
+    },
+    {
+      "epoch": 0.4652439785348808,
+      "grad_norm": 0.08095631003379822,
+      "learning_rate": 8.678082191780822e-06,
+      "loss": 0.6236,
+      "step": 233
+    },
+    {
+      "epoch": 0.4672407338075627,
+      "grad_norm": 0.08462950587272644,
+      "learning_rate": 8.67123287671233e-06,
+      "loss": 0.6479,
+      "step": 234
+    },
+    {
+      "epoch": 0.4692374890802446,
+      "grad_norm": 0.07999172806739807,
+      "learning_rate": 8.664383561643837e-06,
+      "loss": 0.6523,
+      "step": 235
+    },
+    {
+      "epoch": 0.4712342443529265,
+      "grad_norm": 0.08923690766096115,
+      "learning_rate": 8.657534246575343e-06,
+      "loss": 0.634,
+      "step": 236
+    },
+    {
+      "epoch": 0.4732309996256084,
+      "grad_norm": 0.08821967989206314,
+      "learning_rate": 8.65068493150685e-06,
+      "loss": 0.6412,
+      "step": 237
+    },
+    {
+      "epoch": 0.4752277548982903,
+      "grad_norm": 0.08308009803295135,
+      "learning_rate": 8.643835616438357e-06,
+      "loss": 0.5984,
+      "step": 238
+    },
+    {
+      "epoch": 0.47722451017097217,
+      "grad_norm": 0.08100554347038269,
+      "learning_rate": 8.636986301369864e-06,
+      "loss": 0.6221,
+      "step": 239
+    },
+    {
+      "epoch": 0.47922126544365407,
+      "grad_norm": 0.08808869123458862,
+      "learning_rate": 8.63013698630137e-06,
+      "loss": 0.6387,
+      "step": 240
+    },
+    {
+      "epoch": 0.48121802071633596,
+      "grad_norm": 0.08174411207437515,
+      "learning_rate": 8.623287671232877e-06,
+      "loss": 0.6384,
+      "step": 241
+    },
+    {
+      "epoch": 0.48321477598901785,
+      "grad_norm": 0.08261357247829437,
+      "learning_rate": 8.616438356164385e-06,
+      "loss": 0.6512,
+      "step": 242
+    },
+    {
+      "epoch": 0.48521153126169975,
+      "grad_norm": 0.0868501365184784,
+      "learning_rate": 8.609589041095892e-06,
+      "loss": 0.6355,
+      "step": 243
+    },
+    {
+      "epoch": 0.48720828653438164,
+      "grad_norm": 0.07903306931257248,
+      "learning_rate": 8.602739726027397e-06,
+      "loss": 0.6758,
+      "step": 244
+    },
+    {
+      "epoch": 0.48920504180706353,
+      "grad_norm": 0.0908120647072792,
+      "learning_rate": 8.595890410958905e-06,
+      "loss": 0.6165,
+      "step": 245
+    },
+    {
+      "epoch": 0.49120179707974543,
+      "grad_norm": 0.08266864717006683,
+      "learning_rate": 8.589041095890412e-06,
+      "loss": 0.6642,
+      "step": 246
+    },
+    {
+      "epoch": 0.4931985523524273,
+      "grad_norm": 0.08055173605680466,
+      "learning_rate": 8.58219178082192e-06,
+      "loss": 0.617,
+      "step": 247
+    },
+    {
+      "epoch": 0.4951953076251092,
+      "grad_norm": 0.08168390393257141,
+      "learning_rate": 8.575342465753425e-06,
+      "loss": 0.6268,
+      "step": 248
+    },
+    {
+      "epoch": 0.4971920628977911,
+      "grad_norm": 0.0787629634141922,
+      "learning_rate": 8.568493150684932e-06,
+      "loss": 0.6398,
+      "step": 249
+    },
+    {
+      "epoch": 0.499188818170473,
+      "grad_norm": 0.09043371677398682,
+      "learning_rate": 8.56164383561644e-06,
+      "loss": 0.6239,
+      "step": 250
+    },
+    {
+      "epoch": 0.5011855734431548,
+      "grad_norm": 0.08357525616884232,
+      "learning_rate": 8.554794520547945e-06,
+      "loss": 0.659,
+      "step": 251
+    },
+    {
+      "epoch": 0.5031823287158368,
+      "grad_norm": 0.09504301100969315,
+      "learning_rate": 8.547945205479454e-06,
+      "loss": 0.6239,
+      "step": 252
+    },
+    {
+      "epoch": 0.5051790839885186,
+      "grad_norm": 0.08067186176776886,
+      "learning_rate": 8.54109589041096e-06,
+      "loss": 0.6158,
+      "step": 253
+    },
+    {
+      "epoch": 0.5071758392612006,
+      "grad_norm": 0.08868859708309174,
+      "learning_rate": 8.534246575342467e-06,
+      "loss": 0.6434,
+      "step": 254
+    },
+    {
+      "epoch": 0.5091725945338824,
+      "grad_norm": 0.08489713072776794,
+      "learning_rate": 8.527397260273972e-06,
+      "loss": 0.6417,
+      "step": 255
+    },
+    {
+      "epoch": 0.5111693498065644,
+      "grad_norm": 0.08246369659900665,
+      "learning_rate": 8.520547945205481e-06,
+      "loss": 0.6137,
+      "step": 256
+    },
+    {
+      "epoch": 0.5131661050792462,
+      "grad_norm": 0.08800794929265976,
+      "learning_rate": 8.513698630136987e-06,
+      "loss": 0.622,
+      "step": 257
+    },
+    {
+      "epoch": 0.5151628603519282,
+      "grad_norm": 0.0991496592760086,
+      "learning_rate": 8.506849315068494e-06,
+      "loss": 0.6392,
+      "step": 258
+    },
+    {
+      "epoch": 0.51715961562461,
+      "grad_norm": 0.09626424312591553,
+      "learning_rate": 8.5e-06,
+      "loss": 0.6226,
+      "step": 259
+    },
+    {
+      "epoch": 0.5191563708972919,
+      "grad_norm": 0.08874102681875229,
+      "learning_rate": 8.493150684931507e-06,
+      "loss": 0.6269,
+      "step": 260
+    },
+    {
+      "epoch": 0.5211531261699738,
+      "grad_norm": 0.0876912996172905,
+      "learning_rate": 8.486301369863015e-06,
+      "loss": 0.6441,
+      "step": 261
+    },
+    {
+      "epoch": 0.5231498814426557,
+      "grad_norm": 0.08492390066385269,
+      "learning_rate": 8.479452054794522e-06,
+      "loss": 0.6558,
+      "step": 262
+    },
+    {
+      "epoch": 0.5251466367153376,
+      "grad_norm": 0.08091188222169876,
+      "learning_rate": 8.472602739726027e-06,
+      "loss": 0.6603,
+      "step": 263
+    },
+    {
+      "epoch": 0.5271433919880195,
+      "grad_norm": 0.08319351822137833,
+      "learning_rate": 8.465753424657535e-06,
+      "loss": 0.6495,
+      "step": 264
+    },
+    {
+      "epoch": 0.5291401472607014,
+      "grad_norm": 0.08336567878723145,
+      "learning_rate": 8.458904109589042e-06,
+      "loss": 0.6361,
+      "step": 265
+    },
+    {
+      "epoch": 0.5311369025333833,
+      "grad_norm": 0.0845741406083107,
+      "learning_rate": 8.45205479452055e-06,
+      "loss": 0.6435,
+      "step": 266
+    },
+    {
+      "epoch": 0.5331336578060651,
+      "grad_norm": 0.08287670463323593,
+      "learning_rate": 8.445205479452055e-06,
+      "loss": 0.6368,
+      "step": 267
+    },
+    {
+      "epoch": 0.5351304130787471,
+      "grad_norm": 0.17075520753860474,
+      "learning_rate": 8.438356164383562e-06,
+      "loss": 0.6565,
+      "step": 268
+    },
+    {
+      "epoch": 0.5371271683514289,
+      "grad_norm": 0.08212152868509293,
+      "learning_rate": 8.43150684931507e-06,
+      "loss": 0.6207,
+      "step": 269
+    },
+    {
+      "epoch": 0.5391239236241108,
+      "grad_norm": 0.08784583956003189,
+      "learning_rate": 8.424657534246577e-06,
+      "loss": 0.633,
+      "step": 270
+    },
+    {
+      "epoch": 0.5411206788967927,
+      "grad_norm": 0.08179745078086853,
+      "learning_rate": 8.417808219178082e-06,
+      "loss": 0.6232,
+      "step": 271
+    },
+    {
+      "epoch": 0.5431174341694746,
+      "grad_norm": 0.08889183402061462,
+      "learning_rate": 8.41095890410959e-06,
+      "loss": 0.6284,
+      "step": 272
+    },
+    {
+      "epoch": 0.5451141894421565,
+      "grad_norm": 0.0841706246137619,
+      "learning_rate": 8.404109589041097e-06,
+      "loss": 0.621,
+      "step": 273
+    },
+    {
+      "epoch": 0.5471109447148383,
+      "grad_norm": 0.0807892307639122,
+      "learning_rate": 8.397260273972604e-06,
+      "loss": 0.629,
+      "step": 274
+    },
+    {
+      "epoch": 0.5491076999875203,
+      "grad_norm": 0.09147186577320099,
+      "learning_rate": 8.39041095890411e-06,
+      "loss": 0.6723,
+      "step": 275
+    },
+    {
+      "epoch": 0.5511044552602021,
+      "grad_norm": 0.10155276209115982,
+      "learning_rate": 8.383561643835617e-06,
+      "loss": 0.6366,
+      "step": 276
+    },
+    {
+      "epoch": 0.5531012105328841,
+      "grad_norm": 0.08878958970308304,
+      "learning_rate": 8.376712328767124e-06,
+      "loss": 0.6082,
+      "step": 277
+    },
+    {
+      "epoch": 0.5550979658055659,
+      "grad_norm": 0.08273486793041229,
+      "learning_rate": 8.36986301369863e-06,
+      "loss": 0.6477,
+      "step": 278
+    },
+    {
+      "epoch": 0.5570947210782479,
+      "grad_norm": 0.08911009132862091,
+      "learning_rate": 8.363013698630137e-06,
+      "loss": 0.641,
+      "step": 279
+    },
+    {
+      "epoch": 0.5590914763509297,
+      "grad_norm": 0.08673524856567383,
+      "learning_rate": 8.356164383561644e-06,
+      "loss": 0.6658,
+      "step": 280
+    },
+    {
+      "epoch": 0.5610882316236117,
+      "grad_norm": 0.0841827318072319,
+      "learning_rate": 8.349315068493152e-06,
+      "loss": 0.6732,
+      "step": 281
+    },
+    {
+      "epoch": 0.5630849868962935,
+      "grad_norm": 0.09424469619989395,
+      "learning_rate": 8.342465753424657e-06,
+      "loss": 0.6609,
+      "step": 282
+    },
+    {
+      "epoch": 0.5650817421689754,
+      "grad_norm": 0.08211454004049301,
+      "learning_rate": 8.335616438356166e-06,
+      "loss": 0.6376,
+      "step": 283
+    },
+    {
+      "epoch": 0.5670784974416573,
+      "grad_norm": 0.07968860119581223,
+      "learning_rate": 8.328767123287672e-06,
+      "loss": 0.6386,
+      "step": 284
+    },
+    {
+      "epoch": 0.5690752527143392,
+      "grad_norm": 0.08694186806678772,
+      "learning_rate": 8.32191780821918e-06,
+      "loss": 0.6348,
+      "step": 285
+    },
+    {
+      "epoch": 0.5710720079870211,
+      "grad_norm": 0.08228432387113571,
+      "learning_rate": 8.315068493150685e-06,
+      "loss": 0.6388,
+      "step": 286
+    },
+    {
+      "epoch": 0.573068763259703,
+      "grad_norm": 0.08904524892568588,
+      "learning_rate": 8.308219178082192e-06,
+      "loss": 0.6369,
+      "step": 287
+    },
+    {
+      "epoch": 0.5750655185323849,
+      "grad_norm": 0.07731188088655472,
+      "learning_rate": 8.3013698630137e-06,
+      "loss": 0.5778,
+      "step": 288
+    },
+    {
+      "epoch": 0.5770622738050668,
+      "grad_norm": 0.12828634679317474,
+      "learning_rate": 8.294520547945207e-06,
+      "loss": 0.6044,
+      "step": 289
+    },
+    {
+      "epoch": 0.5790590290777486,
+      "grad_norm": 0.09142670780420303,
+      "learning_rate": 8.287671232876712e-06,
+      "loss": 0.612,
+      "step": 290
+    },
+    {
+      "epoch": 0.5810557843504306,
+      "grad_norm": 0.08378866314888,
+      "learning_rate": 8.28082191780822e-06,
+      "loss": 0.5937,
+      "step": 291
+    },
+    {
+      "epoch": 0.5830525396231124,
+      "grad_norm": 0.08074430376291275,
+      "learning_rate": 8.273972602739727e-06,
+      "loss": 0.626,
+      "step": 292
+    },
+    {
+      "epoch": 0.5850492948957944,
+      "grad_norm": 0.1017611101269722,
+      "learning_rate": 8.267123287671234e-06,
+      "loss": 0.5893,
+      "step": 293
+    },
+    {
+      "epoch": 0.5870460501684762,
+      "grad_norm": 0.08989594131708145,
+      "learning_rate": 8.26027397260274e-06,
+      "loss": 0.6521,
+      "step": 294
+    },
+    {
+      "epoch": 0.5890428054411582,
+      "grad_norm": 0.08201949298381805,
+      "learning_rate": 8.253424657534247e-06,
+      "loss": 0.6315,
+      "step": 295
+    },
+    {
+      "epoch": 0.59103956071384,
+      "grad_norm": 0.08401259034872055,
+      "learning_rate": 8.246575342465754e-06,
+      "loss": 0.6177,
+      "step": 296
+    },
+    {
+      "epoch": 0.593036315986522,
+      "grad_norm": 0.10680308938026428,
+      "learning_rate": 8.239726027397262e-06,
+      "loss": 0.646,
+      "step": 297
+    },
+    {
+      "epoch": 0.5950330712592038,
+      "grad_norm": 0.09091546386480331,
+      "learning_rate": 8.232876712328767e-06,
+      "loss": 0.6334,
+      "step": 298
+    },
+    {
+      "epoch": 0.5970298265318856,
+      "grad_norm": 0.08410380780696869,
+      "learning_rate": 8.226027397260274e-06,
+      "loss": 0.6465,
+      "step": 299
+    },
+    {
+      "epoch": 0.5990265818045676,
+      "grad_norm": 0.08529913425445557,
+      "learning_rate": 8.219178082191782e-06,
+      "loss": 0.6545,
+      "step": 300
+    },
+    {
+      "epoch": 0.6010233370772494,
+      "grad_norm": 0.08057110011577606,
+      "learning_rate": 8.212328767123289e-06,
+      "loss": 0.6415,
+      "step": 301
+    },
+    {
+      "epoch": 0.6030200923499314,
+      "grad_norm": 0.08814297616481781,
+      "learning_rate": 8.205479452054795e-06,
+      "loss": 0.6522,
+      "step": 302
+    },
+    {
+      "epoch": 0.6050168476226132,
+      "grad_norm": 0.08834853768348694,
+      "learning_rate": 8.198630136986302e-06,
+      "loss": 0.6369,
+      "step": 303
+    },
+    {
+      "epoch": 0.6070136028952952,
+      "grad_norm": 0.08998329937458038,
+      "learning_rate": 8.19178082191781e-06,
+      "loss": 0.6136,
+      "step": 304
+    },
+    {
+      "epoch": 0.609010358167977,
+      "grad_norm": 0.08033844828605652,
+      "learning_rate": 8.184931506849316e-06,
+      "loss": 0.6479,
+      "step": 305
+    },
+    {
+      "epoch": 0.6110071134406589,
+      "grad_norm": 0.08053375780582428,
+      "learning_rate": 8.178082191780822e-06,
+      "loss": 0.6177,
+      "step": 306
+    },
+    {
+      "epoch": 0.6130038687133408,
+      "grad_norm": 0.08807919919490814,
+      "learning_rate": 8.17123287671233e-06,
+      "loss": 0.6227,
+      "step": 307
+    },
+    {
+      "epoch": 0.6150006239860227,
+      "grad_norm": 0.09442495554685593,
+      "learning_rate": 8.164383561643837e-06,
+      "loss": 0.6372,
+      "step": 308
+    },
+    {
+      "epoch": 0.6169973792587046,
+      "grad_norm": 0.08191791921854019,
+      "learning_rate": 8.157534246575342e-06,
+      "loss": 0.6146,
+      "step": 309
+    },
+    {
+      "epoch": 0.6189941345313865,
+      "grad_norm": 0.08474911749362946,
+      "learning_rate": 8.150684931506851e-06,
+      "loss": 0.6334,
+      "step": 310
+    },
+    {
+      "epoch": 0.6209908898040684,
+      "grad_norm": 0.0809718444943428,
+      "learning_rate": 8.143835616438357e-06,
+      "loss": 0.6417,
+      "step": 311
+    },
+    {
+      "epoch": 0.6229876450767503,
+      "grad_norm": 0.09069975465536118,
+      "learning_rate": 8.136986301369864e-06,
+      "loss": 0.6224,
+      "step": 312
+    },
+    {
+      "epoch": 0.6249844003494321,
+      "grad_norm": 0.08229870349168777,
+      "learning_rate": 8.13013698630137e-06,
+      "loss": 0.6477,
+      "step": 313
+    },
+    {
+      "epoch": 0.6269811556221141,
+      "grad_norm": 0.0842418447136879,
+      "learning_rate": 8.123287671232879e-06,
+      "loss": 0.6155,
+      "step": 314
+    },
+    {
+      "epoch": 0.6289779108947959,
+      "grad_norm": 0.09184005856513977,
+      "learning_rate": 8.116438356164384e-06,
+      "loss": 0.6256,
+      "step": 315
+    },
+    {
+      "epoch": 0.6309746661674779,
+      "grad_norm": 0.08636561781167984,
+      "learning_rate": 8.109589041095892e-06,
+      "loss": 0.6178,
+      "step": 316
+    },
+    {
+      "epoch": 0.6329714214401597,
+      "grad_norm": 0.08534813672304153,
+      "learning_rate": 8.102739726027397e-06,
+      "loss": 0.6478,
+      "step": 317
+    },
+    {
+      "epoch": 0.6349681767128417,
+      "grad_norm": 0.07706432789564133,
+      "learning_rate": 8.095890410958904e-06,
+      "loss": 0.6434,
+      "step": 318
+    },
+    {
+      "epoch": 0.6369649319855235,
+      "grad_norm": 0.09776600450277328,
+      "learning_rate": 8.089041095890412e-06,
+      "loss": 0.6417,
+      "step": 319
+    },
+    {
+      "epoch": 0.6389616872582055,
+      "grad_norm": 0.08862747251987457,
+      "learning_rate": 8.082191780821919e-06,
+      "loss": 0.6475,
+      "step": 320
+    },
+    {
+      "epoch": 0.6409584425308873,
+      "grad_norm": 0.08816659450531006,
+      "learning_rate": 8.075342465753425e-06,
+      "loss": 0.6326,
+      "step": 321
+    },
+    {
+      "epoch": 0.6429551978035692,
+      "grad_norm": 0.1188221126794815,
+      "learning_rate": 8.068493150684932e-06,
+      "loss": 0.6336,
+      "step": 322
+    },
+    {
+      "epoch": 0.6449519530762511,
+      "grad_norm": 0.09697777777910233,
+      "learning_rate": 8.061643835616439e-06,
+      "loss": 0.5971,
+      "step": 323
+    },
+    {
+      "epoch": 0.646948708348933,
+      "grad_norm": 0.08567334711551666,
+      "learning_rate": 8.054794520547946e-06,
+      "loss": 0.6417,
+      "step": 324
+    },
+    {
+      "epoch": 0.6489454636216149,
+      "grad_norm": 0.0873376801609993,
+      "learning_rate": 8.047945205479452e-06,
+      "loss": 0.613,
+      "step": 325
+    },
+    {
+      "epoch": 0.6509422188942968,
+      "grad_norm": 0.08814659714698792,
+      "learning_rate": 8.04109589041096e-06,
+      "loss": 0.6378,
+      "step": 326
+    },
+    {
+      "epoch": 0.6529389741669787,
+      "grad_norm": 0.10195590555667877,
+      "learning_rate": 8.034246575342467e-06,
+      "loss": 0.616,
+      "step": 327
+    },
+    {
+      "epoch": 0.6549357294396605,
+      "grad_norm": 0.09628842771053314,
+      "learning_rate": 8.027397260273974e-06,
+      "loss": 0.6555,
+      "step": 328
+    },
+    {
+      "epoch": 0.6569324847123424,
+      "grad_norm": 0.09534582495689392,
+      "learning_rate": 8.02054794520548e-06,
+      "loss": 0.6322,
+      "step": 329
+    },
+    {
+      "epoch": 0.6589292399850243,
+      "grad_norm": 0.08554988354444504,
+      "learning_rate": 8.013698630136987e-06,
+      "loss": 0.6686,
+      "step": 330
+    },
+    {
+      "epoch": 0.6609259952577062,
+      "grad_norm": 0.08279106765985489,
+      "learning_rate": 8.006849315068494e-06,
+      "loss": 0.6272,
+      "step": 331
+    },
+    {
+      "epoch": 0.6629227505303881,
+      "grad_norm": 0.0819801315665245,
+      "learning_rate": 8.000000000000001e-06,
+      "loss": 0.6284,
+      "step": 332
+    },
+    {
+      "epoch": 0.66491950580307,
+      "grad_norm": 0.10195945203304291,
+      "learning_rate": 7.993150684931507e-06,
+      "loss": 0.6218,
+      "step": 333
+    },
+    {
+      "epoch": 0.6669162610757519,
+      "grad_norm": 0.08146601915359497,
+      "learning_rate": 7.986301369863014e-06,
+      "loss": 0.631,
+      "step": 334
+    },
+    {
+      "epoch": 0.6689130163484338,
+      "grad_norm": 0.08729390054941177,
+      "learning_rate": 7.979452054794521e-06,
+      "loss": 0.6534,
+      "step": 335
+    },
+    {
+      "epoch": 0.6709097716211156,
+      "grad_norm": 0.09326784312725067,
+      "learning_rate": 7.972602739726027e-06,
+      "loss": 0.6107,
+      "step": 336
+    },
+    {
+      "epoch": 0.6729065268937976,
+      "grad_norm": 0.08763978630304337,
+      "learning_rate": 7.965753424657534e-06,
+      "loss": 0.6302,
+      "step": 337
+    },
+    {
+      "epoch": 0.6749032821664794,
+      "grad_norm": 0.0776083841919899,
+      "learning_rate": 7.958904109589042e-06,
+      "loss": 0.6259,
+      "step": 338
+    },
+    {
+      "epoch": 0.6769000374391614,
+      "grad_norm": 0.08602346479892731,
+      "learning_rate": 7.952054794520549e-06,
+      "loss": 0.6129,
+      "step": 339
+    },
+    {
+      "epoch": 0.6788967927118432,
+      "grad_norm": 0.09022011607885361,
+      "learning_rate": 7.945205479452055e-06,
+      "loss": 0.6292,
+      "step": 340
+    },
+    {
+      "epoch": 0.6808935479845252,
+      "grad_norm": 0.08988498151302338,
+      "learning_rate": 7.938356164383564e-06,
+      "loss": 0.6245,
+      "step": 341
+    },
+    {
+      "epoch": 0.682890303257207,
+      "grad_norm": 0.10817273706197739,
+      "learning_rate": 7.931506849315069e-06,
+      "loss": 0.6399,
+      "step": 342
+    },
+    {
+      "epoch": 0.684887058529889,
+      "grad_norm": 0.0887124165892601,
+      "learning_rate": 7.924657534246576e-06,
+      "loss": 0.6235,
+      "step": 343
+    },
+    {
+      "epoch": 0.6868838138025708,
+      "grad_norm": 0.09099099785089493,
+      "learning_rate": 7.917808219178082e-06,
+      "loss": 0.6446,
+      "step": 344
+    },
+    {
+      "epoch": 0.6888805690752527,
+      "grad_norm": 0.08027100563049316,
+      "learning_rate": 7.910958904109591e-06,
+      "loss": 0.6331,
+      "step": 345
+    },
+    {
+      "epoch": 0.6908773243479346,
+      "grad_norm": 0.08706498146057129,
+      "learning_rate": 7.904109589041097e-06,
+      "loss": 0.6098,
+      "step": 346
+    },
+    {
+      "epoch": 0.6928740796206165,
+      "grad_norm": 0.08696519583463669,
+      "learning_rate": 7.897260273972604e-06,
+      "loss": 0.621,
+      "step": 347
+    },
+    {
+      "epoch": 0.6948708348932984,
+      "grad_norm": 0.09188802540302277,
+      "learning_rate": 7.89041095890411e-06,
+      "loss": 0.6342,
+      "step": 348
+    },
+    {
+      "epoch": 0.6968675901659803,
+      "grad_norm": 0.09153212606906891,
+      "learning_rate": 7.883561643835617e-06,
+      "loss": 0.6569,
+      "step": 349
+    },
+    {
+      "epoch": 0.6988643454386622,
+      "grad_norm": 0.10127896070480347,
+      "learning_rate": 7.876712328767124e-06,
+      "loss": 0.6281,
+      "step": 350
+    },
+    {
+      "epoch": 0.7008611007113441,
+      "grad_norm": 0.10001371800899506,
+      "learning_rate": 7.869863013698631e-06,
+      "loss": 0.6669,
+      "step": 351
+    },
+    {
+      "epoch": 0.702857855984026,
+      "grad_norm": 0.08944805711507797,
+      "learning_rate": 7.863013698630137e-06,
+      "loss": 0.6561,
+      "step": 352
+    },
+    {
+      "epoch": 0.7048546112567079,
+      "grad_norm": 0.1012006402015686,
+      "learning_rate": 7.856164383561644e-06,
+      "loss": 0.6303,
+      "step": 353
+    },
+    {
+      "epoch": 0.7068513665293897,
+      "grad_norm": 0.08384621888399124,
+      "learning_rate": 7.849315068493151e-06,
+      "loss": 0.6044,
+      "step": 354
+    },
+    {
+      "epoch": 0.7088481218020717,
+      "grad_norm": 0.10328217595815659,
+      "learning_rate": 7.842465753424659e-06,
+      "loss": 0.6311,
+      "step": 355
+    },
+    {
+      "epoch": 0.7108448770747535,
+      "grad_norm": 0.09059418737888336,
+      "learning_rate": 7.835616438356164e-06,
+      "loss": 0.6575,
+      "step": 356
+    },
+    {
+      "epoch": 0.7128416323474355,
+      "grad_norm": 0.09650541841983795,
+      "learning_rate": 7.828767123287672e-06,
+      "loss": 0.6305,
+      "step": 357
+    },
+    {
+      "epoch": 0.7148383876201173,
+      "grad_norm": 0.08358507603406906,
+      "learning_rate": 7.821917808219179e-06,
+      "loss": 0.6182,
+      "step": 358
+    },
+    {
+      "epoch": 0.7168351428927991,
+      "grad_norm": 0.10657758265733719,
+      "learning_rate": 7.815068493150686e-06,
+      "loss": 0.6341,
+      "step": 359
+    },
+    {
+      "epoch": 0.7188318981654811,
+      "grad_norm": 0.09342513978481293,
+      "learning_rate": 7.808219178082192e-06,
+      "loss": 0.6292,
+      "step": 360
+    },
+    {
+      "epoch": 0.7208286534381629,
+      "grad_norm": 0.08470805734395981,
+      "learning_rate": 7.801369863013699e-06,
+      "loss": 0.6188,
+      "step": 361
+    },
+    {
+      "epoch": 0.7228254087108449,
+      "grad_norm": 0.0813937783241272,
+      "learning_rate": 7.794520547945206e-06,
+      "loss": 0.6133,
+      "step": 362
+    },
+    {
+      "epoch": 0.7248221639835267,
+      "grad_norm": 0.08287567645311356,
+      "learning_rate": 7.787671232876714e-06,
+      "loss": 0.6281,
+      "step": 363
+    },
+    {
+      "epoch": 0.7268189192562087,
+      "grad_norm": 0.09619621187448502,
+      "learning_rate": 7.78082191780822e-06,
+      "loss": 0.6386,
+      "step": 364
+    },
+    {
+      "epoch": 0.7288156745288905,
+      "grad_norm": 0.10158401727676392,
+      "learning_rate": 7.773972602739727e-06,
+      "loss": 0.6355,
+      "step": 365
+    },
+    {
+      "epoch": 0.7308124298015725,
+      "grad_norm": 0.09512092918157578,
+      "learning_rate": 7.767123287671234e-06,
+      "loss": 0.6112,
+      "step": 366
+    },
+    {
+      "epoch": 0.7328091850742543,
+      "grad_norm": 0.08900922536849976,
+      "learning_rate": 7.76027397260274e-06,
+      "loss": 0.6219,
+      "step": 367
+    },
+    {
+      "epoch": 0.7348059403469362,
+      "grad_norm": 0.08434482663869858,
+      "learning_rate": 7.753424657534248e-06,
+      "loss": 0.6343,
+      "step": 368
+    },
+    {
+      "epoch": 0.7368026956196181,
+      "grad_norm": 0.08434830605983734,
+      "learning_rate": 7.746575342465754e-06,
+      "loss": 0.6218,
+      "step": 369
+    },
+    {
+      "epoch": 0.7387994508923,
+      "grad_norm": 0.08528321981430054,
+      "learning_rate": 7.739726027397261e-06,
+      "loss": 0.6285,
+      "step": 370
+    },
+    {
+      "epoch": 0.7407962061649819,
+      "grad_norm": 0.0943007618188858,
+      "learning_rate": 7.732876712328767e-06,
+      "loss": 0.611,
+      "step": 371
+    },
+    {
+      "epoch": 0.7427929614376638,
+      "grad_norm": 0.09442902356386185,
+      "learning_rate": 7.726027397260276e-06,
+      "loss": 0.6212,
+      "step": 372
+    },
+    {
+      "epoch": 0.7447897167103457,
+      "grad_norm": 0.09973379969596863,
+      "learning_rate": 7.719178082191781e-06,
+      "loss": 0.6493,
+      "step": 373
+    },
+    {
+      "epoch": 0.7467864719830276,
+      "grad_norm": 0.09333740174770355,
+      "learning_rate": 7.712328767123289e-06,
+      "loss": 0.6189,
+      "step": 374
+    },
+    {
+      "epoch": 0.7487832272557094,
+      "grad_norm": 0.07975912094116211,
+      "learning_rate": 7.705479452054794e-06,
+      "loss": 0.5907,
+      "step": 375
+    },
+    {
+      "epoch": 0.7507799825283914,
+      "grad_norm": 0.09251334518194199,
+      "learning_rate": 7.698630136986302e-06,
+      "loss": 0.6249,
+      "step": 376
+    },
+    {
+      "epoch": 0.7527767378010732,
+      "grad_norm": 0.0966658890247345,
+      "learning_rate": 7.691780821917809e-06,
+      "loss": 0.6074,
+      "step": 377
+    },
+    {
+      "epoch": 0.7547734930737552,
+      "grad_norm": 0.09172848612070084,
+      "learning_rate": 7.684931506849316e-06,
+      "loss": 0.6623,
+      "step": 378
+    },
+    {
+      "epoch": 0.756770248346437,
+      "grad_norm": 0.0874529704451561,
+      "learning_rate": 7.678082191780822e-06,
+      "loss": 0.6219,
+      "step": 379
+    },
+    {
+      "epoch": 0.758767003619119,
+      "grad_norm": 0.08857685327529907,
+      "learning_rate": 7.671232876712329e-06,
+      "loss": 0.6153,
+      "step": 380
+    },
+    {
+      "epoch": 0.7607637588918008,
+      "grad_norm": 0.08565176278352737,
+      "learning_rate": 7.664383561643836e-06,
+      "loss": 0.6378,
+      "step": 381
+    },
+    {
+      "epoch": 0.7627605141644828,
+      "grad_norm": 0.09114093333482742,
+      "learning_rate": 7.657534246575344e-06,
+      "loss": 0.6344,
+      "step": 382
+    },
+    {
+      "epoch": 0.7647572694371646,
+      "grad_norm": 0.0963209792971611,
+      "learning_rate": 7.65068493150685e-06,
+      "loss": 0.6565,
+      "step": 383
+    },
+    {
+      "epoch": 0.7667540247098465,
+      "grad_norm": 0.08320071548223495,
+      "learning_rate": 7.643835616438356e-06,
+      "loss": 0.6533,
+      "step": 384
+    },
+    {
+      "epoch": 0.7687507799825284,
+      "grad_norm": 0.08294820040464401,
+      "learning_rate": 7.636986301369864e-06,
+      "loss": 0.6681,
+      "step": 385
+    },
+    {
+      "epoch": 0.7707475352552103,
+      "grad_norm": 0.08448363840579987,
+      "learning_rate": 7.630136986301371e-06,
+      "loss": 0.6294,
+      "step": 386
+    },
+    {
+      "epoch": 0.7727442905278922,
+      "grad_norm": 0.09567107260227203,
+      "learning_rate": 7.6232876712328775e-06,
+      "loss": 0.6301,
+      "step": 387
+    },
+    {
+      "epoch": 0.774741045800574,
+      "grad_norm": 0.08810196071863174,
+      "learning_rate": 7.616438356164384e-06,
+      "loss": 0.633,
+      "step": 388
+    },
+    {
+      "epoch": 0.776737801073256,
+      "grad_norm": 0.08176866173744202,
+      "learning_rate": 7.60958904109589e-06,
+      "loss": 0.6035,
+      "step": 389
+    },
+    {
+      "epoch": 0.7787345563459378,
+      "grad_norm": 0.09339761734008789,
+      "learning_rate": 7.6027397260273985e-06,
+      "loss": 0.6663,
+      "step": 390
+    },
+    {
+      "epoch": 0.7807313116186197,
+      "grad_norm": 0.08789702504873276,
+      "learning_rate": 7.595890410958905e-06,
+      "loss": 0.6,
+      "step": 391
+    },
+    {
+      "epoch": 0.7827280668913016,
+      "grad_norm": 0.08306291699409485,
+      "learning_rate": 7.589041095890411e-06,
+      "loss": 0.6306,
+      "step": 392
+    },
+    {
+      "epoch": 0.7847248221639835,
+      "grad_norm": 0.09348531812429428,
+      "learning_rate": 7.582191780821918e-06,
+      "loss": 0.6443,
+      "step": 393
+    },
+    {
+      "epoch": 0.7867215774366654,
+      "grad_norm": 0.08978314697742462,
+      "learning_rate": 7.575342465753426e-06,
+      "loss": 0.6573,
+      "step": 394
+    },
+    {
+      "epoch": 0.7887183327093473,
+      "grad_norm": 0.08130855113267899,
+      "learning_rate": 7.568493150684932e-06,
+      "loss": 0.5953,
+      "step": 395
+    },
+    {
+      "epoch": 0.7907150879820292,
+      "grad_norm": 0.0940231904387474,
+      "learning_rate": 7.561643835616439e-06,
+      "loss": 0.6101,
+      "step": 396
+    },
+    {
+      "epoch": 0.7927118432547111,
+      "grad_norm": 0.08444412052631378,
+      "learning_rate": 7.554794520547945e-06,
+      "loss": 0.6373,
+      "step": 397
+    },
+    {
+      "epoch": 0.794708598527393,
+      "grad_norm": 0.08339810371398926,
+      "learning_rate": 7.5479452054794526e-06,
+      "loss": 0.5962,
+      "step": 398
+    },
+    {
+      "epoch": 0.7967053538000749,
+      "grad_norm": 0.09607964754104614,
+      "learning_rate": 7.54109589041096e-06,
+      "loss": 0.6595,
+      "step": 399
+    },
+    {
+      "epoch": 0.7987021090727567,
+      "grad_norm": 0.0899866372346878,
+      "learning_rate": 7.534246575342466e-06,
+      "loss": 0.6422,
+      "step": 400
+    },
+    {
+      "epoch": 0.8006988643454387,
+      "grad_norm": 0.09215114265680313,
+      "learning_rate": 7.527397260273973e-06,
+      "loss": 0.648,
+      "step": 401
+    },
+    {
+      "epoch": 0.8026956196181205,
+      "grad_norm": 0.09480249136686325,
+      "learning_rate": 7.52054794520548e-06,
+      "loss": 0.6101,
+      "step": 402
+    },
+    {
+      "epoch": 0.8046923748908025,
+      "grad_norm": 0.08409575372934341,
+      "learning_rate": 7.513698630136987e-06,
+      "loss": 0.6149,
+      "step": 403
+    },
+    {
+      "epoch": 0.8066891301634843,
+      "grad_norm": 0.09122111648321152,
+      "learning_rate": 7.506849315068494e-06,
+      "loss": 0.6078,
+      "step": 404
+    },
+    {
+      "epoch": 0.8086858854361663,
+      "grad_norm": 0.08759909123182297,
+      "learning_rate": 7.500000000000001e-06,
+      "loss": 0.664,
+      "step": 405
+    },
+    {
+      "epoch": 0.8106826407088481,
+      "grad_norm": 0.08290824294090271,
+      "learning_rate": 7.4931506849315075e-06,
+      "loss": 0.6229,
+      "step": 406
+    },
+    {
+      "epoch": 0.81267939598153,
+      "grad_norm": 0.10028953105211258,
+      "learning_rate": 7.486301369863014e-06,
+      "loss": 0.6266,
+      "step": 407
+    },
+    {
+      "epoch": 0.8146761512542119,
+      "grad_norm": 0.10031448304653168,
+      "learning_rate": 7.479452054794521e-06,
+      "loss": 0.6356,
+      "step": 408
+    },
+    {
+      "epoch": 0.8166729065268938,
+      "grad_norm": 0.08541914820671082,
+      "learning_rate": 7.4726027397260285e-06,
+      "loss": 0.6375,
+      "step": 409
+    },
+    {
+      "epoch": 0.8186696617995757,
+      "grad_norm": 0.08713380247354507,
+      "learning_rate": 7.465753424657535e-06,
+      "loss": 0.6394,
+      "step": 410
+    },
+    {
+      "epoch": 0.8206664170722576,
+      "grad_norm": 0.09178909659385681,
+      "learning_rate": 7.458904109589041e-06,
+      "loss": 0.5901,
+      "step": 411
+    },
+    {
+      "epoch": 0.8226631723449395,
+      "grad_norm": 0.08403053879737854,
+      "learning_rate": 7.452054794520549e-06,
+      "loss": 0.6128,
+      "step": 412
+    },
+    {
+      "epoch": 0.8246599276176214,
+      "grad_norm": 0.09222221374511719,
+      "learning_rate": 7.445205479452056e-06,
+      "loss": 0.6172,
+      "step": 413
+    },
+    {
+      "epoch": 0.8266566828903033,
+      "grad_norm": 0.07988160848617554,
+      "learning_rate": 7.438356164383562e-06,
+      "loss": 0.6301,
+      "step": 414
+    },
+    {
+      "epoch": 0.8286534381629852,
+      "grad_norm": 0.08793946355581284,
+      "learning_rate": 7.431506849315069e-06,
+      "loss": 0.6482,
+      "step": 415
+    },
+    {
+      "epoch": 0.830650193435667,
+      "grad_norm": 0.08366432040929794,
+      "learning_rate": 7.424657534246575e-06,
+      "loss": 0.63,
+      "step": 416
+    },
+    {
+      "epoch": 0.832646948708349,
+      "grad_norm": 0.08675362914800644,
+      "learning_rate": 7.417808219178083e-06,
+      "loss": 0.6145,
+      "step": 417
+    },
+    {
+      "epoch": 0.8346437039810308,
+      "grad_norm": 0.08647222071886063,
+      "learning_rate": 7.41095890410959e-06,
+      "loss": 0.6525,
+      "step": 418
+    },
+    {
+      "epoch": 0.8366404592537127,
+      "grad_norm": 0.08656927198171616,
+      "learning_rate": 7.404109589041096e-06,
+      "loss": 0.6316,
+      "step": 419
+    },
+    {
+      "epoch": 0.8386372145263946,
+      "grad_norm": 0.10375358909368515,
+      "learning_rate": 7.397260273972603e-06,
+      "loss": 0.632,
+      "step": 420
+    },
+    {
+      "epoch": 0.8406339697990765,
+      "grad_norm": 0.08459735661745071,
+      "learning_rate": 7.390410958904111e-06,
+      "loss": 0.6264,
+      "step": 421
+    },
+    {
+      "epoch": 0.8426307250717584,
+      "grad_norm": 0.08126968145370483,
+      "learning_rate": 7.383561643835617e-06,
+      "loss": 0.6052,
+      "step": 422
+    },
+    {
+      "epoch": 0.8446274803444402,
+      "grad_norm": 0.08688778430223465,
+      "learning_rate": 7.376712328767124e-06,
+      "loss": 0.6554,
+      "step": 423
+    },
+    {
+      "epoch": 0.8466242356171222,
+      "grad_norm": 0.08540894836187363,
+      "learning_rate": 7.36986301369863e-06,
+      "loss": 0.6349,
+      "step": 424
+    },
+    {
+      "epoch": 0.848620990889804,
+      "grad_norm": 0.0881495550274849,
+      "learning_rate": 7.3630136986301374e-06,
+      "loss": 0.6212,
+      "step": 425
+    },
+    {
+      "epoch": 0.850617746162486,
+      "grad_norm": 0.09146378934383392,
+      "learning_rate": 7.356164383561645e-06,
+      "loss": 0.6251,
+      "step": 426
+    },
+    {
+      "epoch": 0.8526145014351678,
+      "grad_norm": 0.08816662430763245,
+      "learning_rate": 7.349315068493151e-06,
+      "loss": 0.6403,
+      "step": 427
+    },
+    {
+      "epoch": 0.8546112567078498,
+      "grad_norm": 0.08613763749599457,
+      "learning_rate": 7.342465753424658e-06,
+      "loss": 0.6314,
+      "step": 428
+    },
+    {
+      "epoch": 0.8566080119805316,
+      "grad_norm": 0.08418456465005875,
+      "learning_rate": 7.335616438356165e-06,
+      "loss": 0.6524,
+      "step": 429
+    },
+    {
+      "epoch": 0.8586047672532136,
+      "grad_norm": 0.08402597904205322,
+      "learning_rate": 7.328767123287672e-06,
+      "loss": 0.6229,
+      "step": 430
+    },
+    {
+      "epoch": 0.8606015225258954,
+      "grad_norm": 0.08308534324169159,
+      "learning_rate": 7.321917808219179e-06,
+      "loss": 0.6203,
+      "step": 431
+    },
+    {
+      "epoch": 0.8625982777985773,
+      "grad_norm": 0.10041052848100662,
+      "learning_rate": 7.315068493150685e-06,
+      "loss": 0.6021,
+      "step": 432
+    },
+    {
+      "epoch": 0.8645950330712592,
+      "grad_norm": 0.10795553028583527,
+      "learning_rate": 7.308219178082192e-06,
+      "loss": 0.6234,
+      "step": 433
+    },
+    {
+      "epoch": 0.8665917883439411,
+      "grad_norm": 0.08625102043151855,
+      "learning_rate": 7.3013698630137e-06,
+      "loss": 0.6344,
+      "step": 434
+    },
+    {
+      "epoch": 0.868588543616623,
+      "grad_norm": 0.08903323858976364,
+      "learning_rate": 7.294520547945206e-06,
+      "loss": 0.6404,
+      "step": 435
+    },
+    {
+      "epoch": 0.8705852988893049,
+      "grad_norm": 0.08824340999126434,
+      "learning_rate": 7.287671232876713e-06,
+      "loss": 0.6542,
+      "step": 436
+    },
+    {
+      "epoch": 0.8725820541619868,
+      "grad_norm": 0.09185468405485153,
+      "learning_rate": 7.28082191780822e-06,
+      "loss": 0.6218,
+      "step": 437
+    },
+    {
+      "epoch": 0.8745788094346687,
+      "grad_norm": 0.08986788243055344,
+      "learning_rate": 7.273972602739726e-06,
+      "loss": 0.6411,
+      "step": 438
+    },
+    {
+      "epoch": 0.8765755647073505,
+      "grad_norm": 0.08936543017625809,
+      "learning_rate": 7.2671232876712335e-06,
+      "loss": 0.614,
+      "step": 439
+    },
+    {
+      "epoch": 0.8785723199800325,
+      "grad_norm": 0.09536322206258774,
+      "learning_rate": 7.260273972602741e-06,
+      "loss": 0.6505,
+      "step": 440
+    },
+    {
+      "epoch": 0.8805690752527143,
+      "grad_norm": 0.08595971018075943,
+      "learning_rate": 7.253424657534247e-06,
+      "loss": 0.6027,
+      "step": 441
+    },
+    {
+      "epoch": 0.8825658305253963,
+      "grad_norm": 0.08551914989948273,
+      "learning_rate": 7.246575342465754e-06,
+      "loss": 0.5974,
+      "step": 442
+    },
+    {
+      "epoch": 0.8845625857980781,
+      "grad_norm": 0.08609256148338318,
+      "learning_rate": 7.239726027397261e-06,
+      "loss": 0.6302,
+      "step": 443
+    },
+    {
+      "epoch": 0.8865593410707601,
+      "grad_norm": 0.09731276333332062,
+      "learning_rate": 7.232876712328768e-06,
+      "loss": 0.6385,
+      "step": 444
+    },
+    {
+      "epoch": 0.8885560963434419,
+      "grad_norm": 0.11011891067028046,
+      "learning_rate": 7.226027397260275e-06,
+      "loss": 0.6109,
+      "step": 445
+    },
+    {
+      "epoch": 0.8905528516161239,
+      "grad_norm": 0.08579560369253159,
+      "learning_rate": 7.219178082191781e-06,
+      "loss": 0.6168,
+      "step": 446
+    },
+    {
+      "epoch": 0.8925496068888057,
+      "grad_norm": 0.08741603791713715,
+      "learning_rate": 7.2123287671232876e-06,
+      "loss": 0.6375,
+      "step": 447
+    },
+    {
+      "epoch": 0.8945463621614875,
+      "grad_norm": 0.08839676529169083,
+      "learning_rate": 7.205479452054796e-06,
+      "loss": 0.6221,
+      "step": 448
+    },
+    {
+      "epoch": 0.8965431174341695,
+      "grad_norm": 0.09800486266613007,
+      "learning_rate": 7.198630136986302e-06,
+      "loss": 0.6525,
+      "step": 449
+    },
+    {
+      "epoch": 0.8985398727068513,
+      "grad_norm": 0.08749247342348099,
+      "learning_rate": 7.191780821917809e-06,
+      "loss": 0.6122,
+      "step": 450
+    },
+    {
+      "epoch": 0.9005366279795333,
+      "grad_norm": 0.08808398246765137,
+      "learning_rate": 7.184931506849315e-06,
+      "loss": 0.6203,
+      "step": 451
+    },
+    {
+      "epoch": 0.9025333832522151,
+      "grad_norm": 0.0888322964310646,
+      "learning_rate": 7.178082191780823e-06,
+      "loss": 0.6496,
+      "step": 452
+    },
+    {
+      "epoch": 0.904530138524897,
+      "grad_norm": 0.08242765069007874,
+      "learning_rate": 7.17123287671233e-06,
+      "loss": 0.6271,
+      "step": 453
+    },
+    {
+      "epoch": 0.9065268937975789,
+      "grad_norm": 0.08684896677732468,
+      "learning_rate": 7.164383561643836e-06,
+      "loss": 0.6436,
+      "step": 454
+    },
+    {
+      "epoch": 0.9085236490702608,
+      "grad_norm": 0.10484038293361664,
+      "learning_rate": 7.1575342465753425e-06,
+      "loss": 0.6502,
+      "step": 455
+    },
+    {
+      "epoch": 0.9105204043429427,
+      "grad_norm": 0.08097124844789505,
+      "learning_rate": 7.15068493150685e-06,
+      "loss": 0.6401,
+      "step": 456
+    },
+    {
+      "epoch": 0.9125171596156246,
+      "grad_norm": 0.08406626433134079,
+      "learning_rate": 7.143835616438357e-06,
+      "loss": 0.6399,
+      "step": 457
+    },
+    {
+      "epoch": 0.9145139148883065,
+      "grad_norm": 0.08640296757221222,
+      "learning_rate": 7.1369863013698635e-06,
+      "loss": 0.6613,
+      "step": 458
+    },
+    {
+      "epoch": 0.9165106701609884,
+      "grad_norm": 0.08418799191713333,
+      "learning_rate": 7.13013698630137e-06,
+      "loss": 0.6334,
+      "step": 459
+    },
+    {
+      "epoch": 0.9185074254336703,
+      "grad_norm": 0.08541527390480042,
+      "learning_rate": 7.123287671232877e-06,
+      "loss": 0.6263,
+      "step": 460
+    },
+    {
+      "epoch": 0.9205041807063522,
+      "grad_norm": 0.09964878857135773,
+      "learning_rate": 7.1164383561643845e-06,
+      "loss": 0.6313,
+      "step": 461
+    },
+    {
+      "epoch": 0.922500935979034,
+      "grad_norm": 0.08737082034349442,
+      "learning_rate": 7.109589041095891e-06,
+      "loss": 0.6295,
+      "step": 462
+    },
+    {
+      "epoch": 0.924497691251716,
+      "grad_norm": 0.08089511096477509,
+      "learning_rate": 7.102739726027398e-06,
+      "loss": 0.6327,
+      "step": 463
+    },
+    {
+      "epoch": 0.9264944465243978,
+      "grad_norm": 0.08780577033758163,
+      "learning_rate": 7.095890410958905e-06,
+      "loss": 0.6457,
+      "step": 464
+    },
+    {
+      "epoch": 0.9284912017970798,
+      "grad_norm": 0.08980534970760345,
+      "learning_rate": 7.089041095890411e-06,
+      "loss": 0.6208,
+      "step": 465
+    },
+    {
+      "epoch": 0.9304879570697616,
+      "grad_norm": 0.08758597820997238,
+      "learning_rate": 7.082191780821918e-06,
+      "loss": 0.6306,
+      "step": 466
+    },
+    {
+      "epoch": 0.9324847123424436,
+      "grad_norm": 0.09077585488557816,
+      "learning_rate": 7.075342465753426e-06,
+      "loss": 0.6196,
+      "step": 467
+    },
+    {
+      "epoch": 0.9344814676151254,
+      "grad_norm": 0.08882813900709152,
+      "learning_rate": 7.068493150684932e-06,
+      "loss": 0.6564,
+      "step": 468
+    },
+    {
+      "epoch": 0.9364782228878074,
+      "grad_norm": 0.08457037806510925,
+      "learning_rate": 7.0616438356164386e-06,
+      "loss": 0.6163,
+      "step": 469
+    },
+    {
+      "epoch": 0.9384749781604892,
+      "grad_norm": 0.08448978513479233,
+      "learning_rate": 7.054794520547946e-06,
+      "loss": 0.6185,
+      "step": 470
+    },
+    {
+      "epoch": 0.9404717334331711,
+      "grad_norm": 0.09043850004673004,
+      "learning_rate": 7.047945205479453e-06,
+      "loss": 0.6508,
+      "step": 471
+    },
+    {
+      "epoch": 0.942468488705853,
+      "grad_norm": 0.0870198905467987,
+      "learning_rate": 7.0410958904109596e-06,
+      "loss": 0.6457,
+      "step": 472
+    },
+    {
+      "epoch": 0.9444652439785349,
+      "grad_norm": 0.09303230792284012,
+      "learning_rate": 7.034246575342466e-06,
+      "loss": 0.6437,
+      "step": 473
+    },
+    {
+      "epoch": 0.9464619992512168,
+      "grad_norm": 0.09160307794809341,
+      "learning_rate": 7.027397260273974e-06,
+      "loss": 0.6648,
+      "step": 474
+    },
+    {
+      "epoch": 0.9484587545238987,
+      "grad_norm": 0.10001704841852188,
+      "learning_rate": 7.020547945205481e-06,
+      "loss": 0.622,
+      "step": 475
+    },
+    {
+      "epoch": 0.9504555097965806,
+      "grad_norm": 0.08349056541919708,
+      "learning_rate": 7.013698630136987e-06,
+      "loss": 0.6618,
+      "step": 476
+    },
+    {
+      "epoch": 0.9524522650692624,
+      "grad_norm": 0.08756298571825027,
+      "learning_rate": 7.0068493150684935e-06,
+      "loss": 0.6304,
+      "step": 477
+    },
+    {
+      "epoch": 0.9544490203419443,
+      "grad_norm": 0.09010770916938782,
+      "learning_rate": 7e-06,
+      "loss": 0.6269,
+      "step": 478
+    },
+    {
+      "epoch": 0.9564457756146262,
+      "grad_norm": 0.09259214997291565,
+      "learning_rate": 6.993150684931508e-06,
+      "loss": 0.6567,
+      "step": 479
+    },
+    {
+      "epoch": 0.9584425308873081,
+      "grad_norm": 0.08488563448190689,
+      "learning_rate": 6.9863013698630145e-06,
+      "loss": 0.5926,
+      "step": 480
+    },
+    {
+      "epoch": 0.96043928615999,
+      "grad_norm": 0.08800887316465378,
+      "learning_rate": 6.979452054794521e-06,
+      "loss": 0.6667,
+      "step": 481
+    },
+    {
+      "epoch": 0.9624360414326719,
+      "grad_norm": 0.09012701362371445,
+      "learning_rate": 6.972602739726027e-06,
+      "loss": 0.6063,
+      "step": 482
+    },
+    {
+      "epoch": 0.9644327967053538,
+      "grad_norm": 0.08592038601636887,
+      "learning_rate": 6.9657534246575355e-06,
+      "loss": 0.6396,
+      "step": 483
+    },
+    {
+      "epoch": 0.9664295519780357,
+      "grad_norm": 0.09730264544487,
+      "learning_rate": 6.958904109589042e-06,
+      "loss": 0.6169,
+      "step": 484
+    },
+    {
+      "epoch": 0.9684263072507175,
+      "grad_norm": 0.08748991042375565,
+      "learning_rate": 6.952054794520548e-06,
+      "loss": 0.6209,
+      "step": 485
+    },
+    {
+      "epoch": 0.9704230625233995,
+      "grad_norm": 0.08917925506830215,
+      "learning_rate": 6.945205479452055e-06,
+      "loss": 0.6248,
+      "step": 486
+    },
+    {
+      "epoch": 0.9724198177960813,
+      "grad_norm": 0.08646436780691147,
+      "learning_rate": 6.938356164383562e-06,
+      "loss": 0.595,
+      "step": 487
+    },
+    {
+      "epoch": 0.9744165730687633,
+      "grad_norm": 0.08982618153095245,
+      "learning_rate": 6.931506849315069e-06,
+      "loss": 0.6173,
+      "step": 488
+    },
+    {
+      "epoch": 0.9764133283414451,
+      "grad_norm": 0.091037318110466,
+      "learning_rate": 6.924657534246576e-06,
+      "loss": 0.6465,
+      "step": 489
+    },
+    {
+      "epoch": 0.9784100836141271,
+      "grad_norm": 0.08660898357629776,
+      "learning_rate": 6.917808219178082e-06,
+      "loss": 0.5961,
+      "step": 490
+    },
+    {
+      "epoch": 0.9804068388868089,
+      "grad_norm": 0.09286137670278549,
+      "learning_rate": 6.9109589041095895e-06,
+      "loss": 0.6001,
+      "step": 491
+    },
+    {
+      "epoch": 0.9824035941594909,
+      "grad_norm": 0.09190206974744797,
+      "learning_rate": 6.904109589041097e-06,
+      "loss": 0.652,
+      "step": 492
+    },
+    {
+      "epoch": 0.9844003494321727,
+      "grad_norm": 0.09255805611610413,
+      "learning_rate": 6.897260273972603e-06,
+      "loss": 0.6105,
+      "step": 493
+    },
+    {
+      "epoch": 0.9863971047048546,
+      "grad_norm": 0.08385660499334335,
+      "learning_rate": 6.8904109589041105e-06,
+      "loss": 0.6054,
+      "step": 494
+    },
+    {
+      "epoch": 0.9883938599775365,
+      "grad_norm": 0.08500760793685913,
+      "learning_rate": 6.883561643835617e-06,
+      "loss": 0.6044,
+      "step": 495
+    },
+    {
+      "epoch": 0.9903906152502184,
+      "grad_norm": 0.08311589807271957,
+      "learning_rate": 6.876712328767123e-06,
+      "loss": 0.6161,
+      "step": 496
+    },
+    {
+      "epoch": 0.9923873705229003,
+      "grad_norm": 0.08439301699399948,
+      "learning_rate": 6.869863013698631e-06,
+      "loss": 0.6253,
+      "step": 497
+    },
+    {
+      "epoch": 0.9943841257955822,
+      "grad_norm": 0.08830545842647552,
+      "learning_rate": 6.863013698630138e-06,
+      "loss": 0.6301,
+      "step": 498
+    },
+    {
+      "epoch": 0.9963808810682641,
+      "grad_norm": 0.08998648822307587,
+      "learning_rate": 6.8561643835616444e-06,
+      "loss": 0.6576,
+      "step": 499
+    },
+    {
+      "epoch": 0.998377636340946,
+      "grad_norm": 0.08443831652402878,
+      "learning_rate": 6.849315068493151e-06,
+      "loss": 0.5919,
+      "step": 500
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 1500,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "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": 1.1536860580434412e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-500/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:754e4ae9d879853457ade197edaf9a4e09b3e2d2a23d8bffcf3f56a687083eb3
+size 6904

checkpoint-500/zero_to_fp32.py

@@ -0,0 +1,674 @@
+#!/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 json
+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)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in 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}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_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)
+
+    # 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")
+        state_dict_split = split_torch_state_dict_into_shards(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
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard, output_path)
+
+    # 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)