Add files using upload-large-folder tool

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-1000/README.md

@@ -0,0 +1,209 @@
+---
+base_model: Qwen/Qwen2.5-Coder-3B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen2.5-Coder-3B
+- lora
+- sft
+- transformers
+- trl
+---
+
+# 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 0.18.0

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-10000/README.md

@@ -0,0 +1,209 @@
+---
+base_model: Qwen/Qwen2.5-Coder-3B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen2.5-Coder-3B
+- lora
+- sft
+- transformers
+- trl
+---
+
+# 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 0.18.0

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-10000/adapter_config.json

@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-Coder-3B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "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": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "down_proj",
+    "k_proj",
+    "o_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-10000/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
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-10000/chat_template.jinja

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

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-10000/latest

@@ -0,0 +1 @@
+global_step10000

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-10000/tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-10000/zero_to_fp32.py

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

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-2000/adapter_config.json

@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-Coder-3B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "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": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "down_proj",
+    "k_proj",
+    "o_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-2000/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
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-2000/chat_template.jinja

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

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-2000/zero_to_fp32.py

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

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/README.md

@@ -0,0 +1,209 @@
+---
+base_model: Qwen/Qwen2.5-Coder-3B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen2.5-Coder-3B
+- lora
+- sft
+- transformers
+- trl
+---
+
+# 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 0.18.0

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/adapter_config.json

@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-Coder-3B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "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": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "down_proj",
+    "k_proj",
+    "o_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/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
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/chat_template.jinja

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

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/latest

@@ -0,0 +1 @@
+global_step3000

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/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
+  }
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/trainer_state.json

@@ -0,0 +1,634 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.6,
+  "eval_steps": 500,
+  "global_step": 3000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "entropy": 0.13992431640625,
+      "epoch": 0.01,
+      "grad_norm": 0.07415792346000671,
+      "learning_rate": 3.266666666666667e-05,
+      "loss": 0.1267,
+      "mean_token_accuracy": 0.9717467153072357,
+      "num_tokens": 973322.0,
+      "step": 50
+    },
+    {
+      "entropy": 0.034481201171875,
+      "epoch": 0.02,
+      "grad_norm": 0.04046332463622093,
+      "learning_rate": 6.6e-05,
+      "loss": 0.0304,
+      "mean_token_accuracy": 0.9923338234424591,
+      "num_tokens": 1965908.0,
+      "step": 100
+    },
+    {
+      "entropy": 0.026558837890625,
+      "epoch": 0.03,
+      "grad_norm": 0.05207290127873421,
+      "learning_rate": 9.933333333333334e-05,
+      "loss": 0.0222,
+      "mean_token_accuracy": 0.9942584156990051,
+      "num_tokens": 2931943.0,
+      "step": 150
+    },
+    {
+      "entropy": 0.0237799072265625,
+      "epoch": 0.04,
+      "grad_norm": 0.05697847530245781,
+      "learning_rate": 0.00013266666666666667,
+      "loss": 0.0205,
+      "mean_token_accuracy": 0.9945518994331359,
+      "num_tokens": 3906979.0,
+      "step": 200
+    },
+    {
+      "entropy": 0.021153564453125,
+      "epoch": 0.05,
+      "grad_norm": 0.018519293516874313,
+      "learning_rate": 0.000166,
+      "loss": 0.0181,
+      "mean_token_accuracy": 0.9951600968837738,
+      "num_tokens": 4891791.0,
+      "step": 250
+    },
+    {
+      "entropy": 0.0194329833984375,
+      "epoch": 0.06,
+      "grad_norm": 0.04633281007409096,
+      "learning_rate": 0.00019933333333333334,
+      "loss": 0.0173,
+      "mean_token_accuracy": 0.9954203897714615,
+      "num_tokens": 5923811.0,
+      "step": 300
+    },
+    {
+      "entropy": 0.0206915283203125,
+      "epoch": 0.07,
+      "grad_norm": 0.028644593432545662,
+      "learning_rate": 0.00019998740757634997,
+      "loss": 0.0178,
+      "mean_token_accuracy": 0.9951675140857696,
+      "num_tokens": 6923169.0,
+      "step": 350
+    },
+    {
+      "entropy": 0.0210107421875,
+      "epoch": 0.08,
+      "grad_norm": 0.03093401901423931,
+      "learning_rate": 0.00019994860043273915,
+      "loss": 0.0169,
+      "mean_token_accuracy": 0.9953666085004806,
+      "num_tokens": 7901360.0,
+      "step": 400
+    },
+    {
+      "entropy": 0.0185345458984375,
+      "epoch": 0.09,
+      "grad_norm": 0.0355745367705822,
+      "learning_rate": 0.00019988358350090047,
+      "loss": 0.0159,
+      "mean_token_accuracy": 0.9957600867748261,
+      "num_tokens": 8899074.0,
+      "step": 450
+    },
+    {
+      "entropy": 0.0172314453125,
+      "epoch": 0.1,
+      "grad_norm": 0.01078465860337019,
+      "learning_rate": 0.00019979237383039745,
+      "loss": 0.0149,
+      "mean_token_accuracy": 0.9960266047716141,
+      "num_tokens": 9909340.0,
+      "step": 500
+    },
+    {
+      "entropy": 0.018392333984375,
+      "epoch": 0.11,
+      "grad_norm": 0.02546997182071209,
+      "learning_rate": 0.00019967499533938543,
+      "loss": 0.0154,
+      "mean_token_accuracy": 0.9959233921766281,
+      "num_tokens": 10894801.0,
+      "step": 550
+    },
+    {
+      "entropy": 0.020382080078125,
+      "epoch": 0.12,
+      "grad_norm": 0.014133265241980553,
+      "learning_rate": 0.00019953147880833933,
+      "loss": 0.0155,
+      "mean_token_accuracy": 0.9958612543344497,
+      "num_tokens": 11862148.0,
+      "step": 600
+    },
+    {
+      "entropy": 0.0170843505859375,
+      "epoch": 0.13,
+      "grad_norm": 0.013547766953706741,
+      "learning_rate": 0.00019936186187198213,
+      "loss": 0.0149,
+      "mean_token_accuracy": 0.9959749585390091,
+      "num_tokens": 12851463.0,
+      "step": 650
+    },
+    {
+      "entropy": 0.018056640625,
+      "epoch": 0.14,
+      "grad_norm": 0.013980485498905182,
+      "learning_rate": 0.00019916618900941567,
+      "loss": 0.0146,
+      "mean_token_accuracy": 0.9961372888088227,
+      "num_tokens": 13849228.0,
+      "step": 700
+    },
+    {
+      "entropy": 0.01843017578125,
+      "epoch": 0.15,
+      "grad_norm": 0.019236328080296516,
+      "learning_rate": 0.00019894451153245693,
+      "loss": 0.0155,
+      "mean_token_accuracy": 0.995870366692543,
+      "num_tokens": 14812166.0,
+      "step": 750
+    },
+    {
+      "entropy": 0.01776611328125,
+      "epoch": 0.16,
+      "grad_norm": 0.01584303192794323,
+      "learning_rate": 0.00019869688757218233,
+      "loss": 0.0148,
+      "mean_token_accuracy": 0.9960641515254974,
+      "num_tokens": 15789120.0,
+      "step": 800
+    },
+    {
+      "entropy": 0.018167724609375,
+      "epoch": 0.17,
+      "grad_norm": 0.007502980995923281,
+      "learning_rate": 0.00019842338206368373,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.9961654812097549,
+      "num_tokens": 16764645.0,
+      "step": 850
+    },
+    {
+      "entropy": 0.0162567138671875,
+      "epoch": 0.18,
+      "grad_norm": 0.025812745094299316,
+      "learning_rate": 0.00019812406672904058,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9961814099550247,
+      "num_tokens": 17730003.0,
+      "step": 900
+    },
+    {
+      "entropy": 0.0166925048828125,
+      "epoch": 0.19,
+      "grad_norm": 0.01129944622516632,
+      "learning_rate": 0.00019779902005851187,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962814956903457,
+      "num_tokens": 18723861.0,
+      "step": 950
+    },
+    {
+      "entropy": 0.016802978515625,
+      "epoch": 0.2,
+      "grad_norm": 0.010111505165696144,
+      "learning_rate": 0.0001974483272899535,
+      "loss": 0.0146,
+      "mean_token_accuracy": 0.9961066269874572,
+      "num_tokens": 19695091.0,
+      "step": 1000
+    },
+    {
+      "entropy": 0.0173785400390625,
+      "epoch": 0.21,
+      "grad_norm": 0.014308207668364048,
+      "learning_rate": 0.00019707208038646602,
+      "loss": 0.0143,
+      "mean_token_accuracy": 0.9962161642313003,
+      "num_tokens": 20662255.0,
+      "step": 1050
+    },
+    {
+      "entropy": 0.016761474609375,
+      "epoch": 0.22,
+      "grad_norm": 0.008552557788789272,
+      "learning_rate": 0.0001966703780122791,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962283843755722,
+      "num_tokens": 21642092.0,
+      "step": 1100
+    },
+    {
+      "entropy": 0.0194580078125,
+      "epoch": 0.23,
+      "grad_norm": 0.017318639904260635,
+      "learning_rate": 0.00019624332550687834,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.996121798157692,
+      "num_tokens": 22607505.0,
+      "step": 1150
+    },
+    {
+      "entropy": 0.0168377685546875,
+      "epoch": 0.24,
+      "grad_norm": 0.023380566388368607,
+      "learning_rate": 0.0001957910348573819,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962751477956772,
+      "num_tokens": 23584579.0,
+      "step": 1200
+    },
+    {
+      "entropy": 0.0168145751953125,
+      "epoch": 0.25,
+      "grad_norm": 0.03495066240429878,
+      "learning_rate": 0.00019531362466917386,
+      "loss": 0.0143,
+      "mean_token_accuracy": 0.9962228804826736,
+      "num_tokens": 24580285.0,
+      "step": 1250
+    },
+    {
+      "entropy": 0.0200946044921875,
+      "epoch": 0.26,
+      "grad_norm": 0.056234363466501236,
+      "learning_rate": 0.00019481122013480196,
+      "loss": 0.0149,
+      "mean_token_accuracy": 0.9959270429611206,
+      "num_tokens": 25525513.0,
+      "step": 1300
+    },
+    {
+      "entropy": 0.226702880859375,
+      "epoch": 0.27,
+      "grad_norm": 0.5915283560752869,
+      "learning_rate": 0.000194283953001148,
+      "loss": 0.3575,
+      "mean_token_accuracy": 0.9630927595775575,
+      "num_tokens": 26523426.0,
+      "step": 1350
+    },
+    {
+      "entropy": 0.3616259765625,
+      "epoch": 0.28,
+      "grad_norm": 0.05585862696170807,
+      "learning_rate": 0.00019373196153487962,
+      "loss": 0.4834,
+      "mean_token_accuracy": 0.922457065274939,
+      "num_tokens": 27493091.0,
+      "step": 1400
+    },
+    {
+      "entropy": 0.032896728515625,
+      "epoch": 0.29,
+      "grad_norm": 0.015423770062625408,
+      "learning_rate": 0.0001931553904861921,
+      "loss": 0.0361,
+      "mean_token_accuracy": 0.9919799596071244,
+      "num_tokens": 28458553.0,
+      "step": 1450
+    },
+    {
+      "entropy": 0.017860107421875,
+      "epoch": 0.3,
+      "grad_norm": 0.008802004158496857,
+      "learning_rate": 0.0001925543910508503,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.9960576659440994,
+      "num_tokens": 29430497.0,
+      "step": 1500
+    },
+    {
+      "entropy": 0.016207275390625,
+      "epoch": 0.31,
+      "grad_norm": 0.01611204445362091,
+      "learning_rate": 0.00019192912083054002,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.996290831565857,
+      "num_tokens": 30414152.0,
+      "step": 1550
+    },
+    {
+      "entropy": 0.016380615234375,
+      "epoch": 0.32,
+      "grad_norm": 0.026660647243261337,
+      "learning_rate": 0.00019127974379153982,
+      "loss": 0.0135,
+      "mean_token_accuracy": 0.9963940876722336,
+      "num_tokens": 31396417.0,
+      "step": 1600
+    },
+    {
+      "entropy": 0.0166357421875,
+      "epoch": 0.33,
+      "grad_norm": 0.007346960715949535,
+      "learning_rate": 0.00019060643022172365,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9963644134998322,
+      "num_tokens": 32364443.0,
+      "step": 1650
+    },
+    {
+      "entropy": 0.0140234375,
+      "epoch": 0.34,
+      "grad_norm": 0.007081303279846907,
+      "learning_rate": 0.00018990935668590581,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965172445774079,
+      "num_tokens": 33374821.0,
+      "step": 1700
+    },
+    {
+      "entropy": 0.0175,
+      "epoch": 0.35,
+      "grad_norm": 0.01651921682059765,
+      "learning_rate": 0.00018918870597953996,
+      "loss": 0.0135,
+      "mean_token_accuracy": 0.9964259153604508,
+      "num_tokens": 34338570.0,
+      "step": 1750
+    },
+    {
+      "entropy": 0.0165570068359375,
+      "epoch": 0.36,
+      "grad_norm": 0.06110697612166405,
+      "learning_rate": 0.00018844466708078396,
+      "loss": 0.021,
+      "mean_token_accuracy": 0.995423709154129,
+      "num_tokens": 35333253.0,
+      "step": 1800
+    },
+    {
+      "entropy": 0.018714599609375,
+      "epoch": 0.37,
+      "grad_norm": 0.014176812954246998,
+      "learning_rate": 0.00018767743510094396,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.99625692486763,
+      "num_tokens": 36316255.0,
+      "step": 1850
+    },
+    {
+      "entropy": 0.0153717041015625,
+      "epoch": 0.38,
+      "grad_norm": 0.03263692930340767,
+      "learning_rate": 0.00018688721123330952,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9964441621303558,
+      "num_tokens": 37303247.0,
+      "step": 1900
+    },
+    {
+      "entropy": 0.0149493408203125,
+      "epoch": 0.39,
+      "grad_norm": 0.006613259203732014,
+      "learning_rate": 0.0001860742027003944,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9965205788612366,
+      "num_tokens": 38287463.0,
+      "step": 1950
+    },
+    {
+      "entropy": 0.0151849365234375,
+      "epoch": 0.4,
+      "grad_norm": 0.013149562291800976,
+      "learning_rate": 0.00018523862269959583,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965434390306472,
+      "num_tokens": 39273383.0,
+      "step": 2000
+    },
+    {
+      "entropy": 0.0161175537109375,
+      "epoch": 0.41,
+      "grad_norm": 0.014459224417805672,
+      "learning_rate": 0.00018438069034728739,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965963387489318,
+      "num_tokens": 40251665.0,
+      "step": 2050
+    },
+    {
+      "entropy": 0.0144482421875,
+      "epoch": 0.42,
+      "grad_norm": 0.006077172700315714,
+      "learning_rate": 0.0001835006306213594,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965920424461365,
+      "num_tokens": 41243731.0,
+      "step": 2100
+    },
+    {
+      "entropy": 0.0148297119140625,
+      "epoch": 0.43,
+      "grad_norm": 0.0054248967207968235,
+      "learning_rate": 0.0001825986743022225,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965362721681594,
+      "num_tokens": 42247346.0,
+      "step": 2150
+    },
+    {
+      "entropy": 0.0157464599609375,
+      "epoch": 0.44,
+      "grad_norm": 0.007285145577043295,
+      "learning_rate": 0.0001816750579122893,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964449483156205,
+      "num_tokens": 43220001.0,
+      "step": 2200
+    },
+    {
+      "entropy": 0.0147015380859375,
+      "epoch": 0.45,
+      "grad_norm": 0.01308107003569603,
+      "learning_rate": 0.0001807300236539506,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965626937150955,
+      "num_tokens": 44205197.0,
+      "step": 2250
+    },
+    {
+      "entropy": 0.0149774169921875,
+      "epoch": 0.46,
+      "grad_norm": 0.008238991722464561,
+      "learning_rate": 0.00017976381934606202,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965445536375046,
+      "num_tokens": 45185181.0,
+      "step": 2300
+    },
+    {
+      "entropy": 0.0153643798828125,
+      "epoch": 0.47,
+      "grad_norm": 0.006222889292985201,
+      "learning_rate": 0.00017877669835895778,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965794909000397,
+      "num_tokens": 46170545.0,
+      "step": 2350
+    },
+    {
+      "entropy": 0.0146527099609375,
+      "epoch": 0.48,
+      "grad_norm": 0.0058044977486133575,
+      "learning_rate": 0.0001777689195480087,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965272891521454,
+      "num_tokens": 47152165.0,
+      "step": 2400
+    },
+    {
+      "entropy": 0.0149383544921875,
+      "epoch": 0.49,
+      "grad_norm": 0.0066731953993439674,
+      "learning_rate": 0.00017674074718574185,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965432262420655,
+      "num_tokens": 48122023.0,
+      "step": 2450
+    },
+    {
+      "entropy": 0.0154571533203125,
+      "epoch": 0.5,
+      "grad_norm": 0.00943196564912796,
+      "learning_rate": 0.0001756924508925397,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965236240625381,
+      "num_tokens": 49094218.0,
+      "step": 2500
+    },
+    {
+      "entropy": 0.014625244140625,
+      "epoch": 0.51,
+      "grad_norm": 0.01604033261537552,
+      "learning_rate": 0.00017462430556593685,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965505039691925,
+      "num_tokens": 50090248.0,
+      "step": 2550
+    },
+    {
+      "entropy": 0.0170452880859375,
+      "epoch": 0.52,
+      "grad_norm": 0.006134673021733761,
+      "learning_rate": 0.0001735365913085329,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964285892248154,
+      "num_tokens": 51055771.0,
+      "step": 2600
+    },
+    {
+      "entropy": 0.0144781494140625,
+      "epoch": 0.53,
+      "grad_norm": 0.006857218220829964,
+      "learning_rate": 0.0001724295933545404,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9966075003147126,
+      "num_tokens": 52051029.0,
+      "step": 2650
+    },
+    {
+      "entropy": 0.0141400146484375,
+      "epoch": 0.54,
+      "grad_norm": 0.00432049622759223,
+      "learning_rate": 0.0001713036019949871,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966482019424439,
+      "num_tokens": 53064519.0,
+      "step": 2700
+    },
+    {
+      "entropy": 0.0157073974609375,
+      "epoch": 0.55,
+      "grad_norm": 0.011981231160461903,
+      "learning_rate": 0.0001701589125015922,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965185868740082,
+      "num_tokens": 54043844.0,
+      "step": 2750
+    },
+    {
+      "entropy": 0.015498046875,
+      "epoch": 0.56,
+      "grad_norm": 0.004674599040299654,
+      "learning_rate": 0.00016899582504933638,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965594351291657,
+      "num_tokens": 55026422.0,
+      "step": 2800
+    },
+    {
+      "entropy": 0.01620849609375,
+      "epoch": 0.57,
+      "grad_norm": 0.007724366616457701,
+      "learning_rate": 0.00016781464463774628,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965313857793808,
+      "num_tokens": 56009908.0,
+      "step": 2850
+    },
+    {
+      "entropy": 0.0163592529296875,
+      "epoch": 0.58,
+      "grad_norm": 0.01893342286348343,
+      "learning_rate": 0.00016661568101091344,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9964901131391525,
+      "num_tokens": 56948904.0,
+      "step": 2900
+    },
+    {
+      "entropy": 0.0152777099609375,
+      "epoch": 0.59,
+      "grad_norm": 0.004910571966320276,
+      "learning_rate": 0.00016539924857626948,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965794110298156,
+      "num_tokens": 57925745.0,
+      "step": 2950
+    },
+    {
+      "entropy": 0.01528564453125,
+      "epoch": 0.6,
+      "grad_norm": 0.005053428467363119,
+      "learning_rate": 0.00016416566632213803,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966023367643356,
+      "num_tokens": 58902798.0,
+      "step": 3000
+    }
+  ],
+  "logging_steps": 50,
+  "max_steps": 10000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "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": 1.0229677374601953e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-3000/zero_to_fp32.py

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

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-8500/adapter_config.json

@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-Coder-3B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "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": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "down_proj",
+    "k_proj",
+    "o_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-8500/latest

@@ -0,0 +1 @@
+global_step8500

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-8500/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
+  }
+}

qwen2.5-coder-3B_r64_lr2e-4/checkpoint-8500/tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-10000/adapter_config.json

@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-Coder-3B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "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": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-10000/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
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-10000/latest

@@ -0,0 +1 @@
+global_step10000

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-10000/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
+  }
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-10000/tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-10000/trainer_state.json

@@ -0,0 +1,2034 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.0,
+  "eval_steps": 500,
+  "global_step": 10000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "entropy": 0.115654296875,
+      "epoch": 0.01,
+      "grad_norm": 0.061536818742752075,
+      "learning_rate": 6.533333333333334e-05,
+      "loss": 0.1051,
+      "mean_token_accuracy": 0.9763539999723434,
+      "num_tokens": 973322.0,
+      "step": 50
+    },
+    {
+      "entropy": 0.028941650390625,
+      "epoch": 0.02,
+      "grad_norm": 0.0290920902043581,
+      "learning_rate": 0.000132,
+      "loss": 0.0254,
+      "mean_token_accuracy": 0.9933820539712905,
+      "num_tokens": 1965908.0,
+      "step": 100
+    },
+    {
+      "entropy": 0.02515380859375,
+      "epoch": 0.03,
+      "grad_norm": 0.03164820373058319,
+      "learning_rate": 0.00019866666666666668,
+      "loss": 0.0212,
+      "mean_token_accuracy": 0.9944199597835541,
+      "num_tokens": 2931943.0,
+      "step": 150
+    },
+    {
+      "entropy": 0.022684326171875,
+      "epoch": 0.04,
+      "grad_norm": 0.2901318073272705,
+      "learning_rate": 0.00026533333333333335,
+      "loss": 0.0197,
+      "mean_token_accuracy": 0.9946700716018677,
+      "num_tokens": 3906979.0,
+      "step": 200
+    },
+    {
+      "entropy": 0.038095703125,
+      "epoch": 0.05,
+      "grad_norm": 0.4409797191619873,
+      "learning_rate": 0.000332,
+      "loss": 0.0463,
+      "mean_token_accuracy": 0.9893747383356094,
+      "num_tokens": 4891791.0,
+      "step": 250
+    },
+    {
+      "entropy": 0.168302001953125,
+      "epoch": 0.06,
+      "grad_norm": 0.2795103192329407,
+      "learning_rate": 0.0003986666666666667,
+      "loss": 0.2955,
+      "mean_token_accuracy": 0.953380196094513,
+      "num_tokens": 5923811.0,
+      "step": 300
+    },
+    {
+      "entropy": 0.021402587890625,
+      "epoch": 0.07,
+      "grad_norm": 0.030290283262729645,
+      "learning_rate": 0.00039997481515269994,
+      "loss": 0.0188,
+      "mean_token_accuracy": 0.9949713909626007,
+      "num_tokens": 6923169.0,
+      "step": 350
+    },
+    {
+      "entropy": 0.01998291015625,
+      "epoch": 0.08,
+      "grad_norm": 0.016790853813290596,
+      "learning_rate": 0.0003998972008654783,
+      "loss": 0.0165,
+      "mean_token_accuracy": 0.9954974794387818,
+      "num_tokens": 7901360.0,
+      "step": 400
+    },
+    {
+      "entropy": 0.01788330078125,
+      "epoch": 0.09,
+      "grad_norm": 0.0331241637468338,
+      "learning_rate": 0.00039976716700180093,
+      "loss": 0.0158,
+      "mean_token_accuracy": 0.9958017998933792,
+      "num_tokens": 8899074.0,
+      "step": 450
+    },
+    {
+      "entropy": 0.0165899658203125,
+      "epoch": 0.1,
+      "grad_norm": 0.007617511786520481,
+      "learning_rate": 0.0003995847476607949,
+      "loss": 0.0148,
+      "mean_token_accuracy": 0.9960355496406555,
+      "num_tokens": 9909340.0,
+      "step": 500
+    },
+    {
+      "entropy": 0.0178546142578125,
+      "epoch": 0.11,
+      "grad_norm": 0.015544206835329533,
+      "learning_rate": 0.00039934999067877086,
+      "loss": 0.0153,
+      "mean_token_accuracy": 0.9959514647722244,
+      "num_tokens": 10894801.0,
+      "step": 550
+    },
+    {
+      "entropy": 0.019884033203125,
+      "epoch": 0.12,
+      "grad_norm": 0.010751371271908283,
+      "learning_rate": 0.00039906295761667866,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958136093616485,
+      "num_tokens": 11862148.0,
+      "step": 600
+    },
+    {
+      "entropy": 0.01641845703125,
+      "epoch": 0.13,
+      "grad_norm": 0.019799184054136276,
+      "learning_rate": 0.00039872372374396425,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.9960724139213561,
+      "num_tokens": 12851463.0,
+      "step": 650
+    },
+    {
+      "entropy": 0.0172430419921875,
+      "epoch": 0.14,
+      "grad_norm": 0.010893420316278934,
+      "learning_rate": 0.00039833237801883135,
+      "loss": 0.0142,
+      "mean_token_accuracy": 0.9962661373615265,
+      "num_tokens": 13849228.0,
+      "step": 700
+    },
+    {
+      "entropy": 0.0172943115234375,
+      "epoch": 0.15,
+      "grad_norm": 0.010155708529055119,
+      "learning_rate": 0.00039788902306491387,
+      "loss": 0.0148,
+      "mean_token_accuracy": 0.9960399854183197,
+      "num_tokens": 14812166.0,
+      "step": 750
+    },
+    {
+      "entropy": 0.0169000244140625,
+      "epoch": 0.16,
+      "grad_norm": 0.018495460972189903,
+      "learning_rate": 0.00039739377514436467,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.9961151689291,
+      "num_tokens": 15789120.0,
+      "step": 800
+    },
+    {
+      "entropy": 0.0177178955078125,
+      "epoch": 0.17,
+      "grad_norm": 0.022594161331653595,
+      "learning_rate": 0.00039684676412736746,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9962309730052948,
+      "num_tokens": 16764645.0,
+      "step": 850
+    },
+    {
+      "entropy": 0.0165716552734375,
+      "epoch": 0.18,
+      "grad_norm": 0.028149377554655075,
+      "learning_rate": 0.00039624813345808115,
+      "loss": 0.0148,
+      "mean_token_accuracy": 0.996055343747139,
+      "num_tokens": 17730003.0,
+      "step": 900
+    },
+    {
+      "entropy": 0.0167034912109375,
+      "epoch": 0.19,
+      "grad_norm": 0.021027036011219025,
+      "learning_rate": 0.00039559804011702373,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.99612539768219,
+      "num_tokens": 18723861.0,
+      "step": 950
+    },
+    {
+      "entropy": 0.016591796875,
+      "epoch": 0.2,
+      "grad_norm": 0.007501603104174137,
+      "learning_rate": 0.000394896654579907,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.9960941064357758,
+      "num_tokens": 19695091.0,
+      "step": 1000
+    },
+    {
+      "entropy": 0.0202716064453125,
+      "epoch": 0.21,
+      "grad_norm": 0.037511009722948074,
+      "learning_rate": 0.00039414416077293205,
+      "loss": 0.0179,
+      "mean_token_accuracy": 0.9952600795030594,
+      "num_tokens": 20662255.0,
+      "step": 1050
+    },
+    {
+      "entropy": 0.0213037109375,
+      "epoch": 0.22,
+      "grad_norm": 0.3134332299232483,
+      "learning_rate": 0.0003933407560245582,
+      "loss": 0.0184,
+      "mean_token_accuracy": 0.9950169324874878,
+      "num_tokens": 21642092.0,
+      "step": 1100
+    },
+    {
+      "entropy": 1.808961181640625,
+      "epoch": 0.23,
+      "grad_norm": 5.116763591766357,
+      "learning_rate": 0.0003924866510137567,
+      "loss": 3.7666,
+      "mean_token_accuracy": 0.5819910534285009,
+      "num_tokens": 22607505.0,
+      "step": 1150
+    },
+    {
+      "entropy": 3.881328125,
+      "epoch": 0.24,
+      "grad_norm": 2.4985592365264893,
+      "learning_rate": 0.0003915820697147638,
+      "loss": 4.7109,
+      "mean_token_accuracy": 0.07561949249356985,
+      "num_tokens": 23584579.0,
+      "step": 1200
+    },
+    {
+      "entropy": 3.67046875,
+      "epoch": 0.25,
+      "grad_norm": 0.8699260950088501,
+      "learning_rate": 0.0003906272493383477,
+      "loss": 3.7535,
+      "mean_token_accuracy": 0.1315259176492691,
+      "num_tokens": 24580285.0,
+      "step": 1250
+    },
+    {
+      "entropy": 2.81453125,
+      "epoch": 0.26,
+      "grad_norm": 0.5681622624397278,
+      "learning_rate": 0.00038962244026960393,
+      "loss": 2.8209,
+      "mean_token_accuracy": 0.27949289306998254,
+      "num_tokens": 25525513.0,
+      "step": 1300
+    },
+    {
+      "entropy": 1.83828125,
+      "epoch": 0.27,
+      "grad_norm": 2.5825417041778564,
+      "learning_rate": 0.000388567906002296,
+      "loss": 1.8737,
+      "mean_token_accuracy": 0.45127583771944046,
+      "num_tokens": 26523426.0,
+      "step": 1350
+    },
+    {
+      "entropy": 1.411328125,
+      "epoch": 0.28,
+      "grad_norm": 1.3839943408966064,
+      "learning_rate": 0.00038746392306975925,
+      "loss": 1.4483,
+      "mean_token_accuracy": 0.5659240281581879,
+      "num_tokens": 27493091.0,
+      "step": 1400
+    },
+    {
+      "entropy": 1.2139453125,
+      "epoch": 0.29,
+      "grad_norm": 6.3498921394348145,
+      "learning_rate": 0.0003863107809723842,
+      "loss": 1.2179,
+      "mean_token_accuracy": 0.6492460906505585,
+      "num_tokens": 28458553.0,
+      "step": 1450
+    },
+    {
+      "entropy": 1.1167578125,
+      "epoch": 0.3,
+      "grad_norm": 0.5617316961288452,
+      "learning_rate": 0.0003851087821017006,
+      "loss": 1.1004,
+      "mean_token_accuracy": 0.6774312198162079,
+      "num_tokens": 29430497.0,
+      "step": 1500
+    },
+    {
+      "entropy": 1.032578125,
+      "epoch": 0.31,
+      "grad_norm": 1.0182623863220215,
+      "learning_rate": 0.00038385824166108003,
+      "loss": 1.0325,
+      "mean_token_accuracy": 0.6877904099225998,
+      "num_tokens": 30414152.0,
+      "step": 1550
+    },
+    {
+      "entropy": 0.960859375,
+      "epoch": 0.32,
+      "grad_norm": 1.647451400756836,
+      "learning_rate": 0.00038255948758307963,
+      "loss": 0.9682,
+      "mean_token_accuracy": 0.7042217415571213,
+      "num_tokens": 31396417.0,
+      "step": 1600
+    },
+    {
+      "entropy": 0.9691796875,
+      "epoch": 0.33,
+      "grad_norm": 1.280979871749878,
+      "learning_rate": 0.0003812128604434473,
+      "loss": 1.0185,
+      "mean_token_accuracy": 0.6871261638402939,
+      "num_tokens": 32364443.0,
+      "step": 1650
+    },
+    {
+      "entropy": 0.9367578125,
+      "epoch": 0.34,
+      "grad_norm": 1.2345616817474365,
+      "learning_rate": 0.00037981871337181163,
+      "loss": 0.9316,
+      "mean_token_accuracy": 0.7088124883174897,
+      "num_tokens": 33374821.0,
+      "step": 1700
+    },
+    {
+      "entropy": 0.93875,
+      "epoch": 0.35,
+      "grad_norm": 0.8552027940750122,
+      "learning_rate": 0.0003783774119590799,
+      "loss": 0.9486,
+      "mean_token_accuracy": 0.7063367128372192,
+      "num_tokens": 34338570.0,
+      "step": 1750
+    },
+    {
+      "entropy": 0.9719921875,
+      "epoch": 0.36,
+      "grad_norm": 1.9897940158843994,
+      "learning_rate": 0.0003768893341615679,
+      "loss": 1.0009,
+      "mean_token_accuracy": 0.688478217124939,
+      "num_tokens": 35333253.0,
+      "step": 1800
+    },
+    {
+      "entropy": 1.390546875,
+      "epoch": 0.37,
+      "grad_norm": 3.598830223083496,
+      "learning_rate": 0.0003753548702018879,
+      "loss": 2.1995,
+      "mean_token_accuracy": 0.5542165184020996,
+      "num_tokens": 36316255.0,
+      "step": 1850
+    },
+    {
+      "entropy": 1.2059375,
+      "epoch": 0.38,
+      "grad_norm": 1.703786849975586,
+      "learning_rate": 0.00037377442246661904,
+      "loss": 1.2266,
+      "mean_token_accuracy": 0.6345959466695785,
+      "num_tokens": 37303247.0,
+      "step": 1900
+    },
+    {
+      "entropy": 1.03515625,
+      "epoch": 0.39,
+      "grad_norm": 1.381039023399353,
+      "learning_rate": 0.0003721484054007888,
+      "loss": 1.0468,
+      "mean_token_accuracy": 0.686733016371727,
+      "num_tokens": 38287463.0,
+      "step": 1950
+    },
+    {
+      "entropy": 0.9496484375,
+      "epoch": 0.4,
+      "grad_norm": 0.3751803934574127,
+      "learning_rate": 0.00037047724539919166,
+      "loss": 0.9586,
+      "mean_token_accuracy": 0.702388773560524,
+      "num_tokens": 39273383.0,
+      "step": 2000
+    },
+    {
+      "entropy": 0.9304296875,
+      "epoch": 0.41,
+      "grad_norm": 0.8749915361404419,
+      "learning_rate": 0.00036876138069457477,
+      "loss": 0.9364,
+      "mean_token_accuracy": 0.703901634812355,
+      "num_tokens": 40251665.0,
+      "step": 2050
+    },
+    {
+      "entropy": 0.9848828125,
+      "epoch": 0.42,
+      "grad_norm": 1.0551183223724365,
+      "learning_rate": 0.0003670012612427188,
+      "loss": 1.0145,
+      "mean_token_accuracy": 0.6920664620399475,
+      "num_tokens": 41243731.0,
+      "step": 2100
+    },
+    {
+      "entropy": 0.9798046875,
+      "epoch": 0.43,
+      "grad_norm": 2.2353274822235107,
+      "learning_rate": 0.000365197348604445,
+      "loss": 1.0119,
+      "mean_token_accuracy": 0.6892273032665253,
+      "num_tokens": 42247346.0,
+      "step": 2150
+    },
+    {
+      "entropy": 1.095859375,
+      "epoch": 0.44,
+      "grad_norm": 22.283979415893555,
+      "learning_rate": 0.0003633501158245786,
+      "loss": 1.262,
+      "mean_token_accuracy": 0.6328810971975326,
+      "num_tokens": 43220001.0,
+      "step": 2200
+    },
+    {
+      "entropy": 1.37078125,
+      "epoch": 0.45,
+      "grad_norm": 35.408103942871094,
+      "learning_rate": 0.0003614600473079012,
+      "loss": 2.2774,
+      "mean_token_accuracy": 0.5015113458037377,
+      "num_tokens": 44205197.0,
+      "step": 2250
+    },
+    {
+      "entropy": 1.51609375,
+      "epoch": 0.46,
+      "grad_norm": 16.596914291381836,
+      "learning_rate": 0.00035952763869212404,
+      "loss": 1.6605,
+      "mean_token_accuracy": 0.5452959597110748,
+      "num_tokens": 45185181.0,
+      "step": 2300
+    },
+    {
+      "entropy": 1.799609375,
+      "epoch": 0.47,
+      "grad_norm": 11.63796329498291,
+      "learning_rate": 0.00035755339671791555,
+      "loss": 2.8005,
+      "mean_token_accuracy": 0.3781645102798939,
+      "num_tokens": 46170545.0,
+      "step": 2350
+    },
+    {
+      "entropy": 1.87328125,
+      "epoch": 0.48,
+      "grad_norm": 113.29953002929688,
+      "learning_rate": 0.0003555378390960174,
+      "loss": 1.8434,
+      "mean_token_accuracy": 0.49910520017147064,
+      "num_tokens": 47152165.0,
+      "step": 2400
+    },
+    {
+      "entropy": 1.745390625,
+      "epoch": 0.49,
+      "grad_norm": 944.27197265625,
+      "learning_rate": 0.0003534814943714837,
+      "loss": 1.9971,
+      "mean_token_accuracy": 0.47288999617099764,
+      "num_tokens": 48122023.0,
+      "step": 2450
+    },
+    {
+      "entropy": 2.0996875,
+      "epoch": 0.5,
+      "grad_norm": 1222.7352294921875,
+      "learning_rate": 0.0003513849017850794,
+      "loss": 3.7296,
+      "mean_token_accuracy": 0.2439292798936367,
+      "num_tokens": 49094218.0,
+      "step": 2500
+    },
+    {
+      "entropy": 2.57125,
+      "epoch": 0.51,
+      "grad_norm": 11.778553009033203,
+      "learning_rate": 0.0003492486111318737,
+      "loss": 2.6946,
+      "mean_token_accuracy": 0.31728788137435915,
+      "num_tokens": 50090248.0,
+      "step": 2550
+    },
+    {
+      "entropy": 2.065703125,
+      "epoch": 0.52,
+      "grad_norm": 15.745863914489746,
+      "learning_rate": 0.0003470731826170658,
+      "loss": 2.3795,
+      "mean_token_accuracy": 0.4028047382831573,
+      "num_tokens": 51055771.0,
+      "step": 2600
+    },
+    {
+      "entropy": 2.0690625,
+      "epoch": 0.53,
+      "grad_norm": 8.075424194335938,
+      "learning_rate": 0.0003448591867090808,
+      "loss": 2.1452,
+      "mean_token_accuracy": 0.4462559966742992,
+      "num_tokens": 52051029.0,
+      "step": 2650
+    },
+    {
+      "entropy": 1.932109375,
+      "epoch": 0.54,
+      "grad_norm": 16.22416114807129,
+      "learning_rate": 0.0003426072039899742,
+      "loss": 2.0108,
+      "mean_token_accuracy": 0.4872449347376823,
+      "num_tokens": 53064519.0,
+      "step": 2700
+    },
+    {
+      "entropy": 2.450859375,
+      "epoch": 0.55,
+      "grad_norm": 67.30685424804688,
+      "learning_rate": 0.0003403178250031844,
+      "loss": 2.8275,
+      "mean_token_accuracy": 0.3445167612656951,
+      "num_tokens": 54043844.0,
+      "step": 2750
+    },
+    {
+      "entropy": 2.9103125,
+      "epoch": 0.56,
+      "grad_norm": 6.917114734649658,
+      "learning_rate": 0.00033799165009867275,
+      "loss": 7.6596,
+      "mean_token_accuracy": 0.09394362942315639,
+      "num_tokens": 55026422.0,
+      "step": 2800
+    },
+    {
+      "entropy": 3.7703125,
+      "epoch": 0.57,
+      "grad_norm": 2.8941445350646973,
+      "learning_rate": 0.00033562928927549255,
+      "loss": 6.1955,
+      "mean_token_accuracy": 0.08376286070793867,
+      "num_tokens": 56009908.0,
+      "step": 2850
+    },
+    {
+      "entropy": 3.155,
+      "epoch": 0.58,
+      "grad_norm": 38.47834396362305,
+      "learning_rate": 0.00033323136202182687,
+      "loss": 3.4801,
+      "mean_token_accuracy": 0.18894144058227538,
+      "num_tokens": 56948904.0,
+      "step": 2900
+    },
+    {
+      "entropy": 3.15828125,
+      "epoch": 0.59,
+      "grad_norm": 4.531956672668457,
+      "learning_rate": 0.00033079849715253896,
+      "loss": 3.468,
+      "mean_token_accuracy": 0.18886703417636455,
+      "num_tokens": 57925745.0,
+      "step": 2950
+    },
+    {
+      "entropy": 3.0265625,
+      "epoch": 0.6,
+      "grad_norm": 17.36482810974121,
+      "learning_rate": 0.00032833133264427607,
+      "loss": 3.7218,
+      "mean_token_accuracy": 0.17754558715969324,
+      "num_tokens": 58902798.0,
+      "step": 3000
+    },
+    {
+      "entropy": 2.55078125,
+      "epoch": 0.61,
+      "grad_norm": 986.7384033203125,
+      "learning_rate": 0.00032583051546817145,
+      "loss": 6.3096,
+      "mean_token_accuracy": 0.09470455711008981,
+      "num_tokens": 59860890.0,
+      "step": 3050
+    },
+    {
+      "entropy": 3.13359375,
+      "epoch": 0.62,
+      "grad_norm": 256.65728759765625,
+      "learning_rate": 0.00032329670142018733,
+      "loss": 4.9533,
+      "mean_token_accuracy": 0.16578667119145393,
+      "num_tokens": 60845473.0,
+      "step": 3100
+    },
+    {
+      "entropy": 3.57515625,
+      "epoch": 0.63,
+      "grad_norm": 87.86058807373047,
+      "learning_rate": 0.0003207305549491434,
+      "loss": 4.4078,
+      "mean_token_accuracy": 0.14233185030519963,
+      "num_tokens": 61784192.0,
+      "step": 3150
+    },
+    {
+      "entropy": 3.73765625,
+      "epoch": 0.64,
+      "grad_norm": 54.086082458496094,
+      "learning_rate": 0.00031813274898247727,
+      "loss": 4.0546,
+      "mean_token_accuracy": 0.10980181120336056,
+      "num_tokens": 62766207.0,
+      "step": 3200
+    },
+    {
+      "entropy": 3.78609375,
+      "epoch": 0.65,
+      "grad_norm": 87.82706451416016,
+      "learning_rate": 0.00031550396474978035,
+      "loss": 4.0411,
+      "mean_token_accuracy": 0.09306481648236513,
+      "num_tokens": 63762073.0,
+      "step": 3250
+    },
+    {
+      "entropy": 3.8453125,
+      "epoch": 0.66,
+      "grad_norm": 28.116188049316406,
+      "learning_rate": 0.00031284489160415753,
+      "loss": 4.0527,
+      "mean_token_accuracy": 0.09482485845685006,
+      "num_tokens": 64710323.0,
+      "step": 3300
+    },
+    {
+      "entropy": 3.75515625,
+      "epoch": 0.67,
+      "grad_norm": 23.53553009033203,
+      "learning_rate": 0.0003101562268414561,
+      "loss": 4.0821,
+      "mean_token_accuracy": 0.1023617285490036,
+      "num_tokens": 65703355.0,
+      "step": 3350
+    },
+    {
+      "entropy": 3.8,
+      "epoch": 0.68,
+      "grad_norm": 114.70352172851562,
+      "learning_rate": 0.00030743867551741235,
+      "loss": 4.678,
+      "mean_token_accuracy": 0.07633953114971519,
+      "num_tokens": 66693270.0,
+      "step": 3400
+    },
+    {
+      "entropy": 3.8234375,
+      "epoch": 0.69,
+      "grad_norm": 370.61993408203125,
+      "learning_rate": 0.0003046929502627621,
+      "loss": 4.7489,
+      "mean_token_accuracy": 0.06648592198267579,
+      "num_tokens": 67662491.0,
+      "step": 3450
+    },
+    {
+      "entropy": 3.86265625,
+      "epoch": 0.7,
+      "grad_norm": 128.62081909179688,
+      "learning_rate": 0.0003019197710963667,
+      "loss": 4.0626,
+      "mean_token_accuracy": 0.10887656301259994,
+      "num_tokens": 68643197.0,
+      "step": 3500
+    },
+    {
+      "entropy": 3.91328125,
+      "epoch": 0.71,
+      "grad_norm": 108.9946060180664,
+      "learning_rate": 0.00029911986523639976,
+      "loss": 3.9516,
+      "mean_token_accuracy": 0.10265581667423249,
+      "num_tokens": 69638359.0,
+      "step": 3550
+    },
+    {
+      "entropy": 3.98296875,
+      "epoch": 0.72,
+      "grad_norm": 44.63246154785156,
+      "learning_rate": 0.0002962939669096472,
+      "loss": 3.9583,
+      "mean_token_accuracy": 0.09375975500792265,
+      "num_tokens": 70621331.0,
+      "step": 3600
+    },
+    {
+      "entropy": 3.879375,
+      "epoch": 0.73,
+      "grad_norm": 7.976827621459961,
+      "learning_rate": 0.00029344281715896873,
+      "loss": 3.8868,
+      "mean_token_accuracy": 0.09895332656800747,
+      "num_tokens": 71609172.0,
+      "step": 3650
+    },
+    {
+      "entropy": 3.8896875,
+      "epoch": 0.74,
+      "grad_norm": 68.43403625488281,
+      "learning_rate": 0.0002905671636489724,
+      "loss": 3.8063,
+      "mean_token_accuracy": 0.11953793436288834,
+      "num_tokens": 72589978.0,
+      "step": 3700
+    },
+    {
+      "entropy": 2.573828125,
+      "epoch": 0.75,
+      "grad_norm": 74.48558807373047,
+      "learning_rate": 0.0002876677604699527,
+      "loss": 6.8601,
+      "mean_token_accuracy": 0.1135245168209076,
+      "num_tokens": 73606622.0,
+      "step": 3750
+    },
+    {
+      "entropy": 3.7953125,
+      "epoch": 0.76,
+      "grad_norm": 107.41791534423828,
+      "learning_rate": 0.00028474536794014286,
+      "loss": 3.6967,
+      "mean_token_accuracy": 0.13533812060952186,
+      "num_tokens": 74626982.0,
+      "step": 3800
+    },
+    {
+      "entropy": 3.79109375,
+      "epoch": 0.77,
+      "grad_norm": 73.80934143066406,
+      "learning_rate": 0.000281800752406335,
+      "loss": 3.7537,
+      "mean_token_accuracy": 0.12222132451832295,
+      "num_tokens": 75622857.0,
+      "step": 3850
+    },
+    {
+      "entropy": 3.83546875,
+      "epoch": 0.78,
+      "grad_norm": 112.15492248535156,
+      "learning_rate": 0.0002788346860429192,
+      "loss": 3.7633,
+      "mean_token_accuracy": 0.11777356296777725,
+      "num_tokens": 76601614.0,
+      "step": 3900
+    },
+    {
+      "entropy": 3.94125,
+      "epoch": 0.79,
+      "grad_norm": 6.969588279724121,
+      "learning_rate": 0.00027584794664939333,
+      "loss": 4.12,
+      "mean_token_accuracy": 0.08590238390490412,
+      "num_tokens": 77598059.0,
+      "step": 3950
+    },
+    {
+      "entropy": 3.18484375,
+      "epoch": 0.8,
+      "grad_norm": 95.96894073486328,
+      "learning_rate": 0.0002728413174463994,
+      "loss": 4.4326,
+      "mean_token_accuracy": 0.09238502183929086,
+      "num_tokens": 78570180.0,
+      "step": 4000
+    },
+    {
+      "entropy": 3.66625,
+      "epoch": 0.81,
+      "grad_norm": 23.154451370239258,
+      "learning_rate": 0.0002698155868703374,
+      "loss": 3.9934,
+      "mean_token_accuracy": 0.11277755055576563,
+      "num_tokens": 79561943.0,
+      "step": 4050
+    },
+    {
+      "entropy": 3.78109375,
+      "epoch": 0.82,
+      "grad_norm": 173.46826171875,
+      "learning_rate": 0.00026677154836661103,
+      "loss": 3.7989,
+      "mean_token_accuracy": 0.11544285997748376,
+      "num_tokens": 80576757.0,
+      "step": 4100
+    },
+    {
+      "entropy": 3.30609375,
+      "epoch": 0.83,
+      "grad_norm": 843.54248046875,
+      "learning_rate": 0.0002637100001815608,
+      "loss": 3.949,
+      "mean_token_accuracy": 0.12134717077016831,
+      "num_tokens": 81549255.0,
+      "step": 4150
+    },
+    {
+      "entropy": 3.7515625,
+      "epoch": 0.84,
+      "grad_norm": 285.3513488769531,
+      "learning_rate": 0.00026063174515313786,
+      "loss": 3.9247,
+      "mean_token_accuracy": 0.09360126469284297,
+      "num_tokens": 82499683.0,
+      "step": 4200
+    },
+    {
+      "entropy": 3.825625,
+      "epoch": 0.85,
+      "grad_norm": 5339.5615234375,
+      "learning_rate": 0.0002575375905003733,
+      "loss": 4.0126,
+      "mean_token_accuracy": 0.10103015761822462,
+      "num_tokens": 83475810.0,
+      "step": 4250
+    },
+    {
+      "entropy": 3.73765625,
+      "epoch": 0.86,
+      "grad_norm": 37.455440521240234,
+      "learning_rate": 0.00025442834761169915,
+      "loss": 4.5628,
+      "mean_token_accuracy": 0.10651325285434723,
+      "num_tokens": 84432102.0,
+      "step": 4300
+    },
+    {
+      "entropy": 3.46921875,
+      "epoch": 0.87,
+      "grad_norm": 238.13558959960938,
+      "learning_rate": 0.000251304831832176,
+      "loss": 3.8017,
+      "mean_token_accuracy": 0.15232711791992187,
+      "num_tokens": 85429626.0,
+      "step": 4350
+    },
+    {
+      "entropy": 3.4040625,
+      "epoch": 0.88,
+      "grad_norm": 132.93264770507812,
+      "learning_rate": 0.00024816786224968245,
+      "loss": 3.7897,
+      "mean_token_accuracy": 0.18009095534682273,
+      "num_tokens": 86450572.0,
+      "step": 4400
+    },
+    {
+      "entropy": 3.39765625,
+      "epoch": 0.89,
+      "grad_norm": 95.76189422607422,
+      "learning_rate": 0.0002450182614801238,
+      "loss": 3.7702,
+      "mean_token_accuracy": 0.16747837640345098,
+      "num_tokens": 87458843.0,
+      "step": 4450
+    },
+    {
+      "entropy": 3.7196875,
+      "epoch": 0.9,
+      "grad_norm": 1875.1121826171875,
+      "learning_rate": 0.00024185685545171547,
+      "loss": 3.7154,
+      "mean_token_accuracy": 0.15143888495862484,
+      "num_tokens": 88453316.0,
+      "step": 4500
+    },
+    {
+      "entropy": 4.118125,
+      "epoch": 0.91,
+      "grad_norm": 4960.1201171875,
+      "learning_rate": 0.0002386844731883971,
+      "loss": 4.1109,
+      "mean_token_accuracy": 0.061175887137651444,
+      "num_tokens": 89427450.0,
+      "step": 4550
+    },
+    {
+      "entropy": 3.519140625,
+      "epoch": 0.92,
+      "grad_norm": 1.4168286323547363,
+      "learning_rate": 0.00023550194659243583,
+      "loss": 10.0555,
+      "mean_token_accuracy": 0.05055650067864917,
+      "num_tokens": 90400676.0,
+      "step": 4600
+    },
+    {
+      "entropy": 3.83765625,
+      "epoch": 0.93,
+      "grad_norm": 0.4862736165523529,
+      "learning_rate": 0.00023231011022627445,
+      "loss": 3.8585,
+      "mean_token_accuracy": 0.10935776714235544,
+      "num_tokens": 91394757.0,
+      "step": 4650
+    },
+    {
+      "entropy": 2.547265625,
+      "epoch": 0.94,
+      "grad_norm": 50807.24609375,
+      "learning_rate": 0.00022910980109368124,
+      "loss": 6.0729,
+      "mean_token_accuracy": 0.10906331218779088,
+      "num_tokens": 92386137.0,
+      "step": 4700
+    },
+    {
+      "entropy": 1.496796875,
+      "epoch": 0.95,
+      "grad_norm": 1488.6181640625,
+      "learning_rate": 0.0002259018584202608,
+      "loss": 19.0649,
+      "mean_token_accuracy": 0.056977251544594765,
+      "num_tokens": 93374756.0,
+      "step": 4750
+    },
+    {
+      "entropy": 2.1346875,
+      "epoch": 0.96,
+      "grad_norm": 994.2299194335938,
+      "learning_rate": 0.00022268712343338168,
+      "loss": 10.3836,
+      "mean_token_accuracy": 0.084834493547678,
+      "num_tokens": 94389846.0,
+      "step": 4800
+    },
+    {
+      "entropy": 1.878046875,
+      "epoch": 0.97,
+      "grad_norm": 1788.4822998046875,
+      "learning_rate": 0.0002194664391415787,
+      "loss": 11.6867,
+      "mean_token_accuracy": 0.06761666610836983,
+      "num_tokens": 95378637.0,
+      "step": 4850
+    },
+    {
+      "entropy": 2.0459375,
+      "epoch": 0.98,
+      "grad_norm": 1082.5526123046875,
+      "learning_rate": 0.00021624065011348876,
+      "loss": 10.0688,
+      "mean_token_accuracy": 0.047716682897880676,
+      "num_tokens": 96357363.0,
+      "step": 4900
+    },
+    {
+      "entropy": 2.08015625,
+      "epoch": 0.99,
+      "grad_norm": 8028.64892578125,
+      "learning_rate": 0.00021301060225637768,
+      "loss": 7.4894,
+      "mean_token_accuracy": 0.04007244031876325,
+      "num_tokens": 97295349.0,
+      "step": 4950
+    },
+    {
+      "entropy": 2.27015625,
+      "epoch": 1.0,
+      "grad_norm": 4037.40625,
+      "learning_rate": 0.00020977714259431496,
+      "loss": 7.6958,
+      "mean_token_accuracy": 0.0680345633905381,
+      "num_tokens": 98262502.0,
+      "step": 5000
+    },
+    {
+      "entropy": 2.731875,
+      "epoch": 1.01,
+      "grad_norm": 2733.033935546875,
+      "learning_rate": 0.0002065411190460571,
+      "loss": 7.9887,
+      "mean_token_accuracy": 0.05633177805691957,
+      "num_tokens": 99235771.0,
+      "step": 5050
+    },
+    {
+      "entropy": 2.44703125,
+      "epoch": 1.02,
+      "grad_norm": 4614.76513671875,
+      "learning_rate": 0.00020330338020269534,
+      "loss": 7.0014,
+      "mean_token_accuracy": 0.05839766040444374,
+      "num_tokens": 100195990.0,
+      "step": 5100
+    },
+    {
+      "entropy": 1.578828125,
+      "epoch": 1.03,
+      "grad_norm": 171283.71875,
+      "learning_rate": 0.00020006477510512718,
+      "loss": 6.812,
+      "mean_token_accuracy": 0.08080782331526279,
+      "num_tokens": 101140160.0,
+      "step": 5150
+    },
+    {
+      "entropy": 2.081015625,
+      "epoch": 1.04,
+      "grad_norm": 17390.23828125,
+      "learning_rate": 0.0001968261530214103,
+      "loss": 6.5982,
+      "mean_token_accuracy": 0.05107002647593618,
+      "num_tokens": 102128346.0,
+      "step": 5200
+    },
+    {
+      "entropy": 2.18921875,
+      "epoch": 1.05,
+      "grad_norm": 4853.02099609375,
+      "learning_rate": 0.00019358836322405675,
+      "loss": 6.8713,
+      "mean_token_accuracy": 0.04993478316813707,
+      "num_tokens": 103125113.0,
+      "step": 5250
+    },
+    {
+      "entropy": 2.514375,
+      "epoch": 1.06,
+      "grad_norm": 27812.669921875,
+      "learning_rate": 0.00019035225476732597,
+      "loss": 6.2907,
+      "mean_token_accuracy": 0.035861225398257375,
+      "num_tokens": 104136988.0,
+      "step": 5300
+    },
+    {
+      "entropy": 2.440625,
+      "epoch": 1.07,
+      "grad_norm": 7923.54345703125,
+      "learning_rate": 0.00018711867626457486,
+      "loss": 7.3383,
+      "mean_token_accuracy": 0.028668028479442,
+      "num_tokens": 105122070.0,
+      "step": 5350
+    },
+    {
+      "entropy": 2.39671875,
+      "epoch": 1.08,
+      "grad_norm": 199.94097900390625,
+      "learning_rate": 0.0001838884756657237,
+      "loss": 7.6583,
+      "mean_token_accuracy": 0.03105319811962545,
+      "num_tokens": 106117109.0,
+      "step": 5400
+    },
+    {
+      "entropy": 2.7834375,
+      "epoch": 1.09,
+      "grad_norm": 1889.8253173828125,
+      "learning_rate": 0.00018066250003489614,
+      "loss": 6.2994,
+      "mean_token_accuracy": 0.03785766938701272,
+      "num_tokens": 107110319.0,
+      "step": 5450
+    },
+    {
+      "entropy": 3.2371875,
+      "epoch": 1.1,
+      "grad_norm": 29.794780731201172,
+      "learning_rate": 0.00017744159532829106,
+      "loss": 5.1424,
+      "mean_token_accuracy": 0.07308705653063953,
+      "num_tokens": 108104820.0,
+      "step": 5500
+    },
+    {
+      "entropy": 3.6965625,
+      "epoch": 1.11,
+      "grad_norm": 77.62666320800781,
+      "learning_rate": 0.00017422660617234545,
+      "loss": 4.0842,
+      "mean_token_accuracy": 0.08951768040657043,
+      "num_tokens": 109091493.0,
+      "step": 5550
+    },
+    {
+      "entropy": 3.7009375,
+      "epoch": 1.12,
+      "grad_norm": 297.0305480957031,
+      "learning_rate": 0.00017101837564224568,
+      "loss": 3.8852,
+      "mean_token_accuracy": 0.09915917724370957,
+      "num_tokens": 110077375.0,
+      "step": 5600
+    },
+    {
+      "entropy": 3.68171875,
+      "epoch": 1.13,
+      "grad_norm": 2722.0419921875,
+      "learning_rate": 0.00016781774504084557,
+      "loss": 3.8538,
+      "mean_token_accuracy": 0.11030505448579789,
+      "num_tokens": 111042295.0,
+      "step": 5650
+    },
+    {
+      "entropy": 3.66859375,
+      "epoch": 1.1400000000000001,
+      "grad_norm": 6916.11572265625,
+      "learning_rate": 0.0001646255536780493,
+      "loss": 3.9748,
+      "mean_token_accuracy": 0.08198527388274669,
+      "num_tokens": 112033836.0,
+      "step": 5700
+    },
+    {
+      "entropy": 3.70671875,
+      "epoch": 1.15,
+      "grad_norm": 27891.65625,
+      "learning_rate": 0.00016144263865071706,
+      "loss": 4.1508,
+      "mean_token_accuracy": 0.08155446078628302,
+      "num_tokens": 113006848.0,
+      "step": 5750
+    },
+    {
+      "entropy": 3.58953125,
+      "epoch": 1.16,
+      "grad_norm": 3785.768310546875,
+      "learning_rate": 0.00015826983462315092,
+      "loss": 4.2916,
+      "mean_token_accuracy": 0.08131240174174309,
+      "num_tokens": 113998253.0,
+      "step": 5800
+    },
+    {
+      "entropy": 3.54375,
+      "epoch": 1.17,
+      "grad_norm": 13191568.0,
+      "learning_rate": 0.0001551079736082182,
+      "loss": 4.4146,
+      "mean_token_accuracy": 0.08286469839513302,
+      "num_tokens": 114979026.0,
+      "step": 5850
+    },
+    {
+      "entropy": 3.4071875,
+      "epoch": 1.18,
+      "grad_norm": 7038425.5,
+      "learning_rate": 0.0001519578847491713,
+      "loss": 4.4638,
+      "mean_token_accuracy": 0.08333390466868877,
+      "num_tokens": 115947279.0,
+      "step": 5900
+    },
+    {
+      "entropy": 3.43546875,
+      "epoch": 1.19,
+      "grad_norm": 14076842.0,
+      "learning_rate": 0.00014882039410221823,
+      "loss": 4.4573,
+      "mean_token_accuracy": 0.08320307277143002,
+      "num_tokens": 116911661.0,
+      "step": 5950
+    },
+    {
+      "entropy": 3.97921875,
+      "epoch": 1.2,
+      "grad_norm": 5726109.5,
+      "learning_rate": 0.00014569632441990463,
+      "loss": 4.1715,
+      "mean_token_accuracy": 0.11475134715437889,
+      "num_tokens": 117890306.0,
+      "step": 6000
+    },
+    {
+      "entropy": 3.99609375,
+      "epoch": 1.21,
+      "grad_norm": 25805514.0,
+      "learning_rate": 0.00014258649493536054,
+      "loss": 4.1931,
+      "mean_token_accuracy": 0.095585173740983,
+      "num_tokens": 118844495.0,
+      "step": 6050
+    },
+    {
+      "entropy": 3.5678125,
+      "epoch": 1.22,
+      "grad_norm": 8431832.0,
+      "learning_rate": 0.00013949172114747105,
+      "loss": 4.4016,
+      "mean_token_accuracy": 0.08313063785433769,
+      "num_tokens": 119807321.0,
+      "step": 6100
+    },
+    {
+      "entropy": 3.57625,
+      "epoch": 1.23,
+      "grad_norm": 6438381.5,
+      "learning_rate": 0.00013641281460702561,
+      "loss": 4.3257,
+      "mean_token_accuracy": 0.08083299946039915,
+      "num_tokens": 120794059.0,
+      "step": 6150
+    },
+    {
+      "entropy": 3.37234375,
+      "epoch": 1.24,
+      "grad_norm": 25868786.0,
+      "learning_rate": 0.00013335058270390316,
+      "loss": 4.4571,
+      "mean_token_accuracy": 0.0822073706611991,
+      "num_tokens": 121765426.0,
+      "step": 6200
+    },
+    {
+      "entropy": 3.7309375,
+      "epoch": 1.25,
+      "grad_norm": 3721589.25,
+      "learning_rate": 0.00013030582845534796,
+      "loss": 4.1714,
+      "mean_token_accuracy": 0.0834744318202138,
+      "num_tokens": 122747186.0,
+      "step": 6250
+    },
+    {
+      "entropy": 3.780625,
+      "epoch": 1.26,
+      "grad_norm": 2903227.75,
+      "learning_rate": 0.00012727935029539222,
+      "loss": 4.1202,
+      "mean_token_accuracy": 0.12598257400095464,
+      "num_tokens": 123735045.0,
+      "step": 6300
+    },
+    {
+      "entropy": 3.75046875,
+      "epoch": 1.27,
+      "grad_norm": 46265.6953125,
+      "learning_rate": 0.00012427194186548075,
+      "loss": 4.0679,
+      "mean_token_accuracy": 0.15891732826828955,
+      "num_tokens": 124748439.0,
+      "step": 6350
+    },
+    {
+      "entropy": 3.36703125,
+      "epoch": 1.28,
+      "grad_norm": 12431.861328125,
+      "learning_rate": 0.00012128439180635206,
+      "loss": 4.1601,
+      "mean_token_accuracy": 0.11257595017552376,
+      "num_tokens": 125723365.0,
+      "step": 6400
+    },
+    {
+      "entropy": 3.55890625,
+      "epoch": 1.29,
+      "grad_norm": 28675.21484375,
+      "learning_rate": 0.00011831748355123116,
+      "loss": 4.0905,
+      "mean_token_accuracy": 0.1111981638520956,
+      "num_tokens": 126678062.0,
+      "step": 6450
+    },
+    {
+      "entropy": 3.3390625,
+      "epoch": 1.3,
+      "grad_norm": 49817.05859375,
+      "learning_rate": 0.00011537199512038759,
+      "loss": 4.1929,
+      "mean_token_accuracy": 0.11387761130928993,
+      "num_tokens": 127675372.0,
+      "step": 6500
+    },
+    {
+      "entropy": 3.4009375,
+      "epoch": 1.31,
+      "grad_norm": 13634.4345703125,
+      "learning_rate": 0.0001124486989171134,
+      "loss": 4.131,
+      "mean_token_accuracy": 0.12340734615921974,
+      "num_tokens": 128665669.0,
+      "step": 6550
+    },
+    {
+      "entropy": 3.62703125,
+      "epoch": 1.32,
+      "grad_norm": 5092.42041015625,
+      "learning_rate": 0.00010954836152517349,
+      "loss": 4.0627,
+      "mean_token_accuracy": 0.15039996787905693,
+      "num_tokens": 129622542.0,
+      "step": 6600
+    },
+    {
+      "entropy": 3.67046875,
+      "epoch": 1.33,
+      "grad_norm": 63338.5859375,
+      "learning_rate": 0.00010667174350778233,
+      "loss": 4.0277,
+      "mean_token_accuracy": 0.11114733677357436,
+      "num_tokens": 130598823.0,
+      "step": 6650
+    },
+    {
+      "entropy": 3.6784375,
+      "epoch": 1.34,
+      "grad_norm": 23187.904296875,
+      "learning_rate": 0.0001038195992081596,
+      "loss": 4.1025,
+      "mean_token_accuracy": 0.08143463462591172,
+      "num_tokens": 131552204.0,
+      "step": 6700
+    },
+    {
+      "entropy": 3.67859375,
+      "epoch": 1.35,
+      "grad_norm": 91772.46875,
+      "learning_rate": 0.00010099267655171663,
+      "loss": 4.1204,
+      "mean_token_accuracy": 0.08422583721578121,
+      "num_tokens": 132526021.0,
+      "step": 6750
+    },
+    {
+      "entropy": 3.6434375,
+      "epoch": 1.3599999999999999,
+      "grad_norm": 8553529.0,
+      "learning_rate": 9.819171684992574e-05,
+      "loss": 4.1191,
+      "mean_token_accuracy": 0.08334969375282526,
+      "num_tokens": 133483207.0,
+      "step": 6800
+    },
+    {
+      "entropy": 3.56390625,
+      "epoch": 1.37,
+      "grad_norm": 24962412.0,
+      "learning_rate": 9.541745460592477e-05,
+      "loss": 4.2056,
+      "mean_token_accuracy": 0.08088707644492388,
+      "num_tokens": 134482802.0,
+      "step": 6850
+    },
+    {
+      "entropy": 3.41,
+      "epoch": 1.38,
+      "grad_norm": 22948340.0,
+      "learning_rate": 9.267061732190565e-05,
+      "loss": 4.4568,
+      "mean_token_accuracy": 0.08181556835770606,
+      "num_tokens": 135484861.0,
+      "step": 6900
+    },
+    {
+      "entropy": 3.5165625,
+      "epoch": 1.3900000000000001,
+      "grad_norm": 48535188.0,
+      "learning_rate": 8.995192530833998e-05,
+      "loss": 4.2961,
+      "mean_token_accuracy": 0.08130559924989939,
+      "num_tokens": 136499177.0,
+      "step": 6950
+    },
+    {
+      "entropy": 3.54609375,
+      "epoch": 1.4,
+      "grad_norm": 22211810.0,
+      "learning_rate": 8.72620914950899e-05,
+      "loss": 4.269,
+      "mean_token_accuracy": 0.08293722718954086,
+      "num_tokens": 137503523.0,
+      "step": 7000
+    },
+    {
+      "entropy": 3.521875,
+      "epoch": 1.41,
+      "grad_norm": 179278528.0,
+      "learning_rate": 8.460182124445483e-05,
+      "loss": 4.3082,
+      "mean_token_accuracy": 0.08092245563864708,
+      "num_tokens": 138497821.0,
+      "step": 7050
+    },
+    {
+      "entropy": 3.51765625,
+      "epoch": 1.42,
+      "grad_norm": 97545912.0,
+      "learning_rate": 8.197181216620193e-05,
+      "loss": 4.3079,
+      "mean_token_accuracy": 0.0814428037405014,
+      "num_tokens": 139480692.0,
+      "step": 7100
+    },
+    {
+      "entropy": 3.50109375,
+      "epoch": 1.43,
+      "grad_norm": 68456640.0,
+      "learning_rate": 7.937275393463079e-05,
+      "loss": 4.4317,
+      "mean_token_accuracy": 0.08111281462013721,
+      "num_tokens": 140475746.0,
+      "step": 7150
+    },
+    {
+      "entropy": 3.545,
+      "epoch": 1.44,
+      "grad_norm": 199703344.0,
+      "learning_rate": 7.680532810771758e-05,
+      "loss": 4.5208,
+      "mean_token_accuracy": 0.08241649590432644,
+      "num_tokens": 141443323.0,
+      "step": 7200
+    },
+    {
+      "entropy": 3.5165625,
+      "epoch": 1.45,
+      "grad_norm": 121526440.0,
+      "learning_rate": 7.427020794838875e-05,
+      "loss": 4.5181,
+      "mean_token_accuracy": 0.08403186947107315,
+      "num_tokens": 142442496.0,
+      "step": 7250
+    },
+    {
+      "entropy": 3.49984375,
+      "epoch": 1.46,
+      "grad_norm": 523378240.0,
+      "learning_rate": 7.176805824796924e-05,
+      "loss": 4.6314,
+      "mean_token_accuracy": 0.08267527181655168,
+      "num_tokens": 143436493.0,
+      "step": 7300
+    },
+    {
+      "entropy": 3.36484375,
+      "epoch": 1.47,
+      "grad_norm": 59219252.0,
+      "learning_rate": 6.929953515185244e-05,
+      "loss": 4.7744,
+      "mean_token_accuracy": 0.08311606187373399,
+      "num_tokens": 144400422.0,
+      "step": 7350
+    },
+    {
+      "entropy": 3.2334375,
+      "epoch": 1.48,
+      "grad_norm": 27287470.0,
+      "learning_rate": 6.686528598743772e-05,
+      "loss": 4.8599,
+      "mean_token_accuracy": 0.08180953335016966,
+      "num_tokens": 145397669.0,
+      "step": 7400
+    },
+    {
+      "entropy": 3.29203125,
+      "epoch": 1.49,
+      "grad_norm": 220870560.0,
+      "learning_rate": 6.446594909437992e-05,
+      "loss": 4.7484,
+      "mean_token_accuracy": 0.08235592387616635,
+      "num_tokens": 146393328.0,
+      "step": 7450
+    },
+    {
+      "entropy": 3.283125,
+      "epoch": 1.5,
+      "grad_norm": 52497260.0,
+      "learning_rate": 6.21021536571957e-05,
+      "loss": 4.7664,
+      "mean_token_accuracy": 0.08202342621982098,
+      "num_tokens": 147355946.0,
+      "step": 7500
+    },
+    {
+      "entropy": 3.21796875,
+      "epoch": 1.51,
+      "grad_norm": 124470384.0,
+      "learning_rate": 5.9774519540270825e-05,
+      "loss": 4.8058,
+      "mean_token_accuracy": 0.08359716083854436,
+      "num_tokens": 148349601.0,
+      "step": 7550
+    },
+    {
+      "entropy": 3.2775,
+      "epoch": 1.52,
+      "grad_norm": 218581520.0,
+      "learning_rate": 5.748365712531172e-05,
+      "loss": 4.753,
+      "mean_token_accuracy": 0.0818595914170146,
+      "num_tokens": 149313070.0,
+      "step": 7600
+    },
+    {
+      "entropy": 3.32921875,
+      "epoch": 1.53,
+      "grad_norm": 96119168.0,
+      "learning_rate": 5.523016715128302e-05,
+      "loss": 4.7178,
+      "mean_token_accuracy": 0.08181910429149866,
+      "num_tokens": 150295432.0,
+      "step": 7650
+    },
+    {
+      "entropy": 3.23734375,
+      "epoch": 1.54,
+      "grad_norm": 123875408.0,
+      "learning_rate": 5.3014640556874596e-05,
+      "loss": 4.8026,
+      "mean_token_accuracy": 0.08410819683223963,
+      "num_tokens": 151287796.0,
+      "step": 7700
+    },
+    {
+      "entropy": 3.2475,
+      "epoch": 1.55,
+      "grad_norm": 89756040.0,
+      "learning_rate": 5.083765832553811e-05,
+      "loss": 4.8296,
+      "mean_token_accuracy": 0.08127258587628602,
+      "num_tokens": 152301411.0,
+      "step": 7750
+    },
+    {
+      "entropy": 3.2421875,
+      "epoch": 1.56,
+      "grad_norm": 52244988.0,
+      "learning_rate": 4.8699791333133734e-05,
+      "loss": 4.8423,
+      "mean_token_accuracy": 0.08390770640224218,
+      "num_tokens": 153267430.0,
+      "step": 7800
+    },
+    {
+      "entropy": 3.26671875,
+      "epoch": 1.5699999999999998,
+      "grad_norm": 63418748.0,
+      "learning_rate": 4.660160019822837e-05,
+      "loss": 4.8218,
+      "mean_token_accuracy": 0.08359142549335957,
+      "num_tokens": 154223984.0,
+      "step": 7850
+    },
+    {
+      "entropy": 3.285625,
+      "epoch": 1.58,
+      "grad_norm": 31121694.0,
+      "learning_rate": 4.4543635135082864e-05,
+      "loss": 4.8061,
+      "mean_token_accuracy": 0.08465287633240223,
+      "num_tokens": 155196164.0,
+      "step": 7900
+    },
+    {
+      "entropy": 3.33234375,
+      "epoch": 1.5899999999999999,
+      "grad_norm": 219523168.0,
+      "learning_rate": 4.252643580936777e-05,
+      "loss": 4.778,
+      "mean_token_accuracy": 0.08125597681850195,
+      "num_tokens": 156148276.0,
+      "step": 7950
+    },
+    {
+      "entropy": 3.3496875,
+      "epoch": 1.6,
+      "grad_norm": 314958400.0,
+      "learning_rate": 4.055053119664565e-05,
+      "loss": 4.7509,
+      "mean_token_accuracy": 0.08156558005139232,
+      "num_tokens": 157142308.0,
+      "step": 8000
+    },
+    {
+      "entropy": 3.3684375,
+      "epoch": 1.6099999999999999,
+      "grad_norm": 173463392.0,
+      "learning_rate": 3.861643944365638e-05,
+      "loss": 4.7166,
+      "mean_token_accuracy": 0.08095594022423029,
+      "num_tokens": 158121229.0,
+      "step": 8050
+    },
+    {
+      "entropy": 3.345,
+      "epoch": 1.62,
+      "grad_norm": 352727168.0,
+      "learning_rate": 3.672466773244187e-05,
+      "loss": 4.7276,
+      "mean_token_accuracy": 0.08375480145215988,
+      "num_tokens": 159092550.0,
+      "step": 8100
+    },
+    {
+      "entropy": 3.3546875,
+      "epoch": 1.63,
+      "grad_norm": 392757440.0,
+      "learning_rate": 3.487571214734664e-05,
+      "loss": 4.701,
+      "mean_token_accuracy": 0.08215148936957122,
+      "num_tokens": 160079462.0,
+      "step": 8150
+    },
+    {
+      "entropy": 3.3859375,
+      "epoch": 1.6400000000000001,
+      "grad_norm": 494162432.0,
+      "learning_rate": 3.3070057544928024e-05,
+      "loss": 4.6926,
+      "mean_token_accuracy": 0.08144044734537602,
+      "num_tokens": 161078873.0,
+      "step": 8200
+    },
+    {
+      "entropy": 3.3778125,
+      "epoch": 1.65,
+      "grad_norm": 427262176.0,
+      "learning_rate": 3.130817742681129e-05,
+      "loss": 4.7145,
+      "mean_token_accuracy": 0.08182350981980563,
+      "num_tokens": 162066123.0,
+      "step": 8250
+    },
+    {
+      "entropy": 3.40921875,
+      "epoch": 1.6600000000000001,
+      "grad_norm": 116112840.0,
+      "learning_rate": 2.959053381552179e-05,
+      "loss": 4.6951,
+      "mean_token_accuracy": 0.08311355136334896,
+      "num_tokens": 163064136.0,
+      "step": 8300
+    },
+    {
+      "entropy": 3.38828125,
+      "epoch": 1.67,
+      "grad_norm": 303733184.0,
+      "learning_rate": 2.7917577133327698e-05,
+      "loss": 4.7172,
+      "mean_token_accuracy": 0.08152598340064288,
+      "num_tokens": 164044744.0,
+      "step": 8350
+    },
+    {
+      "entropy": 3.35984375,
+      "epoch": 1.6800000000000002,
+      "grad_norm": 483221248.0,
+      "learning_rate": 2.6289746084124445e-05,
+      "loss": 4.7452,
+      "mean_token_accuracy": 0.08084806479513645,
+      "num_tokens": 165046152.0,
+      "step": 8400
+    },
+    {
+      "entropy": 3.34359375,
+      "epoch": 1.69,
+      "grad_norm": 301201760.0,
+      "learning_rate": 2.470746753839235e-05,
+      "loss": 4.7709,
+      "mean_token_accuracy": 0.08231629129499198,
+      "num_tokens": 165986955.0,
+      "step": 8450
+    },
+    {
+      "entropy": 3.35984375,
+      "epoch": 1.7,
+      "grad_norm": 1031817536.0,
+      "learning_rate": 2.3171156421257046e-05,
+      "loss": 4.7584,
+      "mean_token_accuracy": 0.0808232320100069,
+      "num_tokens": 166977751.0,
+      "step": 8500
+    },
+    {
+      "entropy": 3.3728125,
+      "epoch": 1.71,
+      "grad_norm": 213066832.0,
+      "learning_rate": 2.1681215603682858e-05,
+      "loss": 4.7473,
+      "mean_token_accuracy": 0.08175162333995104,
+      "num_tokens": 167953599.0,
+      "step": 8550
+    },
+    {
+      "entropy": 3.35109375,
+      "epoch": 1.72,
+      "grad_norm": 508263424.0,
+      "learning_rate": 2.023803579682657e-05,
+      "loss": 4.7536,
+      "mean_token_accuracy": 0.08210944768041373,
+      "num_tokens": 168959690.0,
+      "step": 8600
+    },
+    {
+      "entropy": 3.3259375,
+      "epoch": 1.73,
+      "grad_norm": 447662976.0,
+      "learning_rate": 1.8841995449580473e-05,
+      "loss": 4.792,
+      "mean_token_accuracy": 0.08307334572076798,
+      "num_tokens": 169960239.0,
+      "step": 8650
+    },
+    {
+      "entropy": 3.334375,
+      "epoch": 1.74,
+      "grad_norm": 139140064.0,
+      "learning_rate": 1.749346064933053e-05,
+      "loss": 4.7966,
+      "mean_token_accuracy": 0.0822683822363615,
+      "num_tokens": 170919546.0,
+      "step": 8700
+    },
+    {
+      "entropy": 3.27921875,
+      "epoch": 1.75,
+      "grad_norm": 271223424.0,
+      "learning_rate": 1.619278502595618e-05,
+      "loss": 4.85,
+      "mean_token_accuracy": 0.08204214256256818,
+      "num_tokens": 171904243.0,
+      "step": 8750
+    },
+    {
+      "entropy": 3.28734375,
+      "epoch": 1.76,
+      "grad_norm": 378753440.0,
+      "learning_rate": 1.494030965909734e-05,
+      "loss": 4.8426,
+      "mean_token_accuracy": 0.0809002497047186,
+      "num_tokens": 172906287.0,
+      "step": 8800
+    },
+    {
+      "entropy": 3.30796875,
+      "epoch": 1.77,
+      "grad_norm": 1111371776.0,
+      "learning_rate": 1.3736362988711993e-05,
+      "loss": 4.8231,
+      "mean_token_accuracy": 0.08123100224882364,
+      "num_tokens": 173893909.0,
+      "step": 8850
+    },
+    {
+      "entropy": 3.30984375,
+      "epoch": 1.78,
+      "grad_norm": 479797888.0,
+      "learning_rate": 1.2581260728948496e-05,
+      "loss": 4.8127,
+      "mean_token_accuracy": 0.083384810090065,
+      "num_tokens": 174890261.0,
+      "step": 8900
+    },
+    {
+      "entropy": 3.2878125,
+      "epoch": 1.79,
+      "grad_norm": 135514880.0,
+      "learning_rate": 1.1475305785355183e-05,
+      "loss": 4.8396,
+      "mean_token_accuracy": 0.08197552632540464,
+      "num_tokens": 175854165.0,
+      "step": 8950
+    },
+    {
+      "entropy": 3.2778125,
+      "epoch": 1.8,
+      "grad_norm": 104816560.0,
+      "learning_rate": 1.0418788175448346e-05,
+      "loss": 4.8572,
+      "mean_token_accuracy": 0.08044139757752418,
+      "num_tokens": 176810914.0,
+      "step": 9000
+    },
+    {
+      "entropy": 3.29140625,
+      "epoch": 1.81,
+      "grad_norm": 632302080.0,
+      "learning_rate": 9.411984952660402e-06,
+      "loss": 4.837,
+      "mean_token_accuracy": 0.0821102448180318,
+      "num_tokens": 177771417.0,
+      "step": 9050
+    },
+    {
+      "entropy": 3.30328125,
+      "epoch": 1.8199999999999998,
+      "grad_norm": 239297808.0,
+      "learning_rate": 8.455160133687368e-06,
+      "loss": 4.8289,
+      "mean_token_accuracy": 0.08218734085559845,
+      "num_tokens": 178723204.0,
+      "step": 9100
+    },
+    {
+      "entropy": 3.3075,
+      "epoch": 1.83,
+      "grad_norm": 620170880.0,
+      "learning_rate": 7.5485646292549946e-06,
+      "loss": 4.8307,
+      "mean_token_accuracy": 0.08314984381198882,
+      "num_tokens": 179696113.0,
+      "step": 9150
+    },
+    {
+      "entropy": 3.2915625,
+      "epoch": 1.8399999999999999,
+      "grad_norm": 468007296.0,
+      "learning_rate": 6.692436178321915e-06,
+      "loss": 4.8417,
+      "mean_token_accuracy": 0.08378895036876202,
+      "num_tokens": 180706266.0,
+      "step": 9200
+    },
+    {
+      "entropy": 3.29984375,
+      "epoch": 1.85,
+      "grad_norm": 249302096.0,
+      "learning_rate": 5.886999285736594e-06,
+      "loss": 4.8375,
+      "mean_token_accuracy": 0.0823655678331852,
+      "num_tokens": 181699421.0,
+      "step": 9250
+    },
+    {
+      "entropy": 3.28328125,
+      "epoch": 1.8599999999999999,
+      "grad_norm": 338011424.0,
+      "learning_rate": 5.132465163364897e-06,
+      "loss": 4.845,
+      "mean_token_accuracy": 0.08341132100671529,
+      "num_tokens": 182712460.0,
+      "step": 9300
+    },
+    {
+      "entropy": 3.27734375,
+      "epoch": 1.87,
+      "grad_norm": 231977024.0,
+      "learning_rate": 4.429031674703321e-06,
+      "loss": 4.8598,
+      "mean_token_accuracy": 0.08164244964718818,
+      "num_tokens": 183707370.0,
+      "step": 9350
+    },
+    {
+      "entropy": 3.275625,
+      "epoch": 1.88,
+      "grad_norm": 163458608.0,
+      "learning_rate": 3.7768832829929046e-06,
+      "loss": 4.85,
+      "mean_token_accuracy": 0.08151648089289665,
+      "num_tokens": 184727025.0,
+      "step": 9400
+    },
+    {
+      "entropy": 3.27890625,
+      "epoch": 1.8900000000000001,
+      "grad_norm": 181780752.0,
+      "learning_rate": 3.1761910028468288e-06,
+      "loss": 4.8514,
+      "mean_token_accuracy": 0.08327789537608624,
+      "num_tokens": 185717464.0,
+      "step": 9450
+    },
+    {
+      "entropy": 3.28515625,
+      "epoch": 1.9,
+      "grad_norm": 403118528.0,
+      "learning_rate": 2.6271123554048037e-06,
+      "loss": 4.846,
+      "mean_token_accuracy": 0.08180779391899705,
+      "num_tokens": 186699086.0,
+      "step": 9500
+    },
+    {
+      "entropy": 3.2878125,
+      "epoch": 1.9100000000000001,
+      "grad_norm": 248453312.0,
+      "learning_rate": 2.1297913270259094e-06,
+      "loss": 4.8419,
+      "mean_token_accuracy": 0.08341263737529517,
+      "num_tokens": 187677214.0,
+      "step": 9550
+    },
+    {
+      "entropy": 3.2875,
+      "epoch": 1.92,
+      "grad_norm": 759850176.0,
+      "learning_rate": 1.6843583315305777e-06,
+      "loss": 4.838,
+      "mean_token_accuracy": 0.08179443206638098,
+      "num_tokens": 188656836.0,
+      "step": 9600
+    },
+    {
+      "entropy": 3.28359375,
+      "epoch": 1.9300000000000002,
+      "grad_norm": 225214656.0,
+      "learning_rate": 1.2909301760020365e-06,
+      "loss": 4.848,
+      "mean_token_accuracy": 0.08185519728809595,
+      "num_tokens": 189644075.0,
+      "step": 9650
+    },
+    {
+      "entropy": 3.283125,
+      "epoch": 1.94,
+      "grad_norm": 160634608.0,
+      "learning_rate": 9.4961003015539e-07,
+      "loss": 4.8454,
+      "mean_token_accuracy": 0.08264562945812941,
+      "num_tokens": 190652489.0,
+      "step": 9700
+    },
+    {
+      "entropy": 3.28,
+      "epoch": 1.95,
+      "grad_norm": 466191232.0,
+      "learning_rate": 6.604873992834382e-07,
+      "loss": 4.848,
+      "mean_token_accuracy": 0.08429396256804467,
+      "num_tokens": 191635957.0,
+      "step": 9750
+    },
+    {
+      "entropy": 3.28234375,
+      "epoch": 1.96,
+      "grad_norm": 360816064.0,
+      "learning_rate": 4.236381007852286e-07,
+      "loss": 4.8496,
+      "mean_token_accuracy": 0.08175841763615609,
+      "num_tokens": 192605918.0,
+      "step": 9800
+    },
+    {
+      "entropy": 3.28171875,
+      "epoch": 1.97,
+      "grad_norm": 250428096.0,
+      "learning_rate": 2.391242442844277e-07,
+      "loss": 4.8551,
+      "mean_token_accuracy": 0.08145457282662391,
+      "num_tokens": 193565571.0,
+      "step": 9850
+    },
+    {
+      "entropy": 3.2834375,
+      "epoch": 1.98,
+      "grad_norm": 142640112.0,
+      "learning_rate": 1.069942153420156e-07,
+      "loss": 4.8516,
+      "mean_token_accuracy": 0.07988192901015281,
+      "num_tokens": 194558732.0,
+      "step": 9900
+    },
+    {
+      "entropy": 3.279375,
+      "epoch": 1.99,
+      "grad_norm": 380546368.0,
+      "learning_rate": 2.7282662768124412e-08,
+      "loss": 4.8546,
+      "mean_token_accuracy": 0.08176410280168056,
+      "num_tokens": 195544901.0,
+      "step": 9950
+    },
+    {
+      "entropy": 3.28140625,
+      "epoch": 2.0,
+      "grad_norm": 450577984.0,
+      "learning_rate": 1.0489535884339318e-11,
+      "loss": 4.8535,
+      "mean_token_accuracy": 0.08271308846771717,
+      "num_tokens": 196525004.0,
+      "step": 10000
+    }
+  ],
+  "logging_steps": 50,
+  "max_steps": 10000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 3.4130592242031657e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-10000/zero_to_fp32.py

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

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-3500/README.md

@@ -0,0 +1,209 @@
+---
+base_model: Qwen/Qwen2.5-Coder-3B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen2.5-Coder-3B
+- lora
+- sft
+- transformers
+- trl
+---
+
+# 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 0.18.0

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-3500/adapter_config.json

@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-Coder-3B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "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": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-3500/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
+}

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-3500/chat_template.jinja

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

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-3500/latest

@@ -0,0 +1 @@
+global_step3500

qwen2.5-coder-3B_r64_lr4e-4/checkpoint-3500/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
+  }
+}