Add files using upload-large-folder tool

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-4000/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_lr1e-4/checkpoint-4000/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": [
+    "up_proj",
+    "q_proj",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "v_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_lr1e-4/checkpoint-4000/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_lr1e-4/checkpoint-4000/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_lr1e-4/checkpoint-4000/latest

@@ -0,0 +1 @@
+global_step4000

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-4000/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_lr1e-4/checkpoint-4000/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_lr1e-4/checkpoint-4000/trainer_state.json

@@ -0,0 +1,834 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.8,
+  "eval_steps": 500,
+  "global_step": 4000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "entropy": 0.166884765625,
+      "epoch": 0.01,
+      "grad_norm": 0.09714163839817047,
+      "learning_rate": 1.6333333333333335e-05,
+      "loss": 0.1512,
+      "mean_token_accuracy": 0.966504562497139,
+      "num_tokens": 973322.0,
+      "step": 50
+    },
+    {
+      "entropy": 0.048140869140625,
+      "epoch": 0.02,
+      "grad_norm": 0.057216521352529526,
+      "learning_rate": 3.3e-05,
+      "loss": 0.0424,
+      "mean_token_accuracy": 0.9896974349021912,
+      "num_tokens": 1965908.0,
+      "step": 100
+    },
+    {
+      "entropy": 0.029632568359375,
+      "epoch": 0.03,
+      "grad_norm": 0.054412033408880234,
+      "learning_rate": 4.966666666666667e-05,
+      "loss": 0.0248,
+      "mean_token_accuracy": 0.9936319047212601,
+      "num_tokens": 2931943.0,
+      "step": 150
+    },
+    {
+      "entropy": 0.02478515625,
+      "epoch": 0.04,
+      "grad_norm": 0.08800562471151352,
+      "learning_rate": 6.633333333333334e-05,
+      "loss": 0.0214,
+      "mean_token_accuracy": 0.9943305045366287,
+      "num_tokens": 3906979.0,
+      "step": 200
+    },
+    {
+      "entropy": 0.0224652099609375,
+      "epoch": 0.05,
+      "grad_norm": 0.041793134063482285,
+      "learning_rate": 8.3e-05,
+      "loss": 0.0191,
+      "mean_token_accuracy": 0.9948792272806167,
+      "num_tokens": 4891791.0,
+      "step": 250
+    },
+    {
+      "entropy": 0.01969482421875,
+      "epoch": 0.06,
+      "grad_norm": 0.04181526601314545,
+      "learning_rate": 9.966666666666667e-05,
+      "loss": 0.0173,
+      "mean_token_accuracy": 0.995435175895691,
+      "num_tokens": 5923811.0,
+      "step": 300
+    },
+    {
+      "entropy": 0.019368896484375,
+      "epoch": 0.07,
+      "grad_norm": 0.03995412588119507,
+      "learning_rate": 9.999370378817499e-05,
+      "loss": 0.0164,
+      "mean_token_accuracy": 0.9956372672319412,
+      "num_tokens": 6923169.0,
+      "step": 350
+    },
+    {
+      "entropy": 0.0204400634765625,
+      "epoch": 0.08,
+      "grad_norm": 0.06872096657752991,
+      "learning_rate": 9.997430021636957e-05,
+      "loss": 0.0163,
+      "mean_token_accuracy": 0.9956023478507996,
+      "num_tokens": 7901360.0,
+      "step": 400
+    },
+    {
+      "entropy": 0.0186175537109375,
+      "epoch": 0.09,
+      "grad_norm": 0.037068989127874374,
+      "learning_rate": 9.994179175045023e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958296239376068,
+      "num_tokens": 8899074.0,
+      "step": 450
+    },
+    {
+      "entropy": 0.0173492431640625,
+      "epoch": 0.1,
+      "grad_norm": 0.014651700854301453,
+      "learning_rate": 9.989618691519873e-05,
+      "loss": 0.015,
+      "mean_token_accuracy": 0.9959587389230729,
+      "num_tokens": 9909340.0,
+      "step": 500
+    },
+    {
+      "entropy": 0.0183203125,
+      "epoch": 0.11,
+      "grad_norm": 0.016151444986462593,
+      "learning_rate": 9.983749766969271e-05,
+      "loss": 0.0151,
+      "mean_token_accuracy": 0.9959770923852921,
+      "num_tokens": 10894801.0,
+      "step": 550
+    },
+    {
+      "entropy": 0.02047607421875,
+      "epoch": 0.12,
+      "grad_norm": 0.021302781999111176,
+      "learning_rate": 9.976573940416966e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958055526018142,
+      "num_tokens": 11862148.0,
+      "step": 600
+    },
+    {
+      "entropy": 0.017862548828125,
+      "epoch": 0.13,
+      "grad_norm": 0.013205700553953648,
+      "learning_rate": 9.968093093599106e-05,
+      "loss": 0.0152,
+      "mean_token_accuracy": 0.9958789277076722,
+      "num_tokens": 12851463.0,
+      "step": 650
+    },
+    {
+      "entropy": 0.0182177734375,
+      "epoch": 0.14,
+      "grad_norm": 0.013120726682245731,
+      "learning_rate": 9.958309450470784e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9962116169929505,
+      "num_tokens": 13849228.0,
+      "step": 700
+    },
+    {
+      "entropy": 0.0179010009765625,
+      "epoch": 0.15,
+      "grad_norm": 0.014701537787914276,
+      "learning_rate": 9.947225576622847e-05,
+      "loss": 0.0146,
+      "mean_token_accuracy": 0.9961142331361771,
+      "num_tokens": 14812166.0,
+      "step": 750
+    },
+    {
+      "entropy": 0.0173638916015625,
+      "epoch": 0.16,
+      "grad_norm": 0.033504419028759,
+      "learning_rate": 9.934844378609117e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.9961155968904495,
+      "num_tokens": 15789120.0,
+      "step": 800
+    },
+    {
+      "entropy": 0.0183349609375,
+      "epoch": 0.17,
+      "grad_norm": 0.010470729321241379,
+      "learning_rate": 9.921169103184187e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9961740493774414,
+      "num_tokens": 16764645.0,
+      "step": 850
+    },
+    {
+      "entropy": 0.016279296875,
+      "epoch": 0.18,
+      "grad_norm": 0.02781980112195015,
+      "learning_rate": 9.906203336452029e-05,
+      "loss": 0.0142,
+      "mean_token_accuracy": 0.9962105673551559,
+      "num_tokens": 17730003.0,
+      "step": 900
+    },
+    {
+      "entropy": 0.0166461181640625,
+      "epoch": 0.19,
+      "grad_norm": 0.020350394770503044,
+      "learning_rate": 9.889951002925593e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962483876943589,
+      "num_tokens": 18723861.0,
+      "step": 950
+    },
+    {
+      "entropy": 0.0165386962890625,
+      "epoch": 0.2,
+      "grad_norm": 0.013212243095040321,
+      "learning_rate": 9.872416364497675e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9961817175149917,
+      "num_tokens": 19695091.0,
+      "step": 1000
+    },
+    {
+      "entropy": 0.017391357421875,
+      "epoch": 0.21,
+      "grad_norm": 0.014041919261217117,
+      "learning_rate": 9.853604019323301e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962570595741272,
+      "num_tokens": 20662255.0,
+      "step": 1050
+    },
+    {
+      "entropy": 0.016884765625,
+      "epoch": 0.22,
+      "grad_norm": 0.017174789682030678,
+      "learning_rate": 9.833518900613956e-05,
+      "loss": 0.0138,
+      "mean_token_accuracy": 0.9962636953592301,
+      "num_tokens": 21642092.0,
+      "step": 1100
+    },
+    {
+      "entropy": 0.019366455078125,
+      "epoch": 0.23,
+      "grad_norm": 0.017337976023554802,
+      "learning_rate": 9.812166275343917e-05,
+      "loss": 0.0143,
+      "mean_token_accuracy": 0.996159919500351,
+      "num_tokens": 22607505.0,
+      "step": 1150
+    },
+    {
+      "entropy": 0.0164886474609375,
+      "epoch": 0.24,
+      "grad_norm": 0.018486863002181053,
+      "learning_rate": 9.789551742869096e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963399863243103,
+      "num_tokens": 23584579.0,
+      "step": 1200
+    },
+    {
+      "entropy": 0.016282958984375,
+      "epoch": 0.25,
+      "grad_norm": 0.009195365011692047,
+      "learning_rate": 9.765681233458693e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9963367432355881,
+      "num_tokens": 24580285.0,
+      "step": 1250
+    },
+    {
+      "entropy": 0.0197235107421875,
+      "epoch": 0.26,
+      "grad_norm": 0.01205432415008545,
+      "learning_rate": 9.740561006740098e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.996049216389656,
+      "num_tokens": 25525513.0,
+      "step": 1300
+    },
+    {
+      "entropy": 0.0153009033203125,
+      "epoch": 0.27,
+      "grad_norm": 0.01174497976899147,
+      "learning_rate": 9.7141976500574e-05,
+      "loss": 0.0135,
+      "mean_token_accuracy": 0.9964245760440826,
+      "num_tokens": 26523426.0,
+      "step": 1350
+    },
+    {
+      "entropy": 0.0169378662109375,
+      "epoch": 0.28,
+      "grad_norm": 0.014242298901081085,
+      "learning_rate": 9.686598076743981e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962529343366623,
+      "num_tokens": 27493091.0,
+      "step": 1400
+    },
+    {
+      "entropy": 0.017811279296875,
+      "epoch": 0.29,
+      "grad_norm": 0.0160844549536705,
+      "learning_rate": 9.657769524309605e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.996181645989418,
+      "num_tokens": 28458553.0,
+      "step": 1450
+    },
+    {
+      "entropy": 0.0169232177734375,
+      "epoch": 0.3,
+      "grad_norm": 0.03241236135363579,
+      "learning_rate": 9.627719552542516e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9962622857093811,
+      "num_tokens": 29430497.0,
+      "step": 1500
+    },
+    {
+      "entropy": 0.0163720703125,
+      "epoch": 0.31,
+      "grad_norm": 0.01750756986439228,
+      "learning_rate": 9.596456041527001e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9964328509569168,
+      "num_tokens": 30414152.0,
+      "step": 1550
+    },
+    {
+      "entropy": 0.017760009765625,
+      "epoch": 0.32,
+      "grad_norm": 0.02182549238204956,
+      "learning_rate": 9.563987189576991e-05,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.9961263346672058,
+      "num_tokens": 31396417.0,
+      "step": 1600
+    },
+    {
+      "entropy": 0.0171954345703125,
+      "epoch": 0.33,
+      "grad_norm": 0.011301269754767418,
+      "learning_rate": 9.530321511086183e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962880289554596,
+      "num_tokens": 32364443.0,
+      "step": 1650
+    },
+    {
+      "entropy": 0.015126953125,
+      "epoch": 0.34,
+      "grad_norm": 0.009061276912689209,
+      "learning_rate": 9.495467834295291e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9964066076278687,
+      "num_tokens": 33374821.0,
+      "step": 1700
+    },
+    {
+      "entropy": 0.018043212890625,
+      "epoch": 0.35,
+      "grad_norm": 0.014127642847597599,
+      "learning_rate": 9.459435298976998e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963885354995727,
+      "num_tokens": 34338570.0,
+      "step": 1750
+    },
+    {
+      "entropy": 0.0155584716796875,
+      "epoch": 0.36,
+      "grad_norm": 0.00913177989423275,
+      "learning_rate": 9.422233354039198e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965262812376022,
+      "num_tokens": 35333253.0,
+      "step": 1800
+    },
+    {
+      "entropy": 0.0177911376953125,
+      "epoch": 0.37,
+      "grad_norm": 0.009629561565816402,
+      "learning_rate": 9.383871755047198e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9964758467674255,
+      "num_tokens": 36316255.0,
+      "step": 1850
+    },
+    {
+      "entropy": 0.0159539794921875,
+      "epoch": 0.38,
+      "grad_norm": 0.01395715307444334,
+      "learning_rate": 9.344360561665476e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964872688055039,
+      "num_tokens": 37303247.0,
+      "step": 1900
+    },
+    {
+      "entropy": 0.0152783203125,
+      "epoch": 0.39,
+      "grad_norm": 0.007368483114987612,
+      "learning_rate": 9.30371013501972e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965988719463348,
+      "num_tokens": 38287463.0,
+      "step": 1950
+    },
+    {
+      "entropy": 0.0158270263671875,
+      "epoch": 0.4,
+      "grad_norm": 0.01508031040430069,
+      "learning_rate": 9.261931134979791e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9965020543336869,
+      "num_tokens": 39273383.0,
+      "step": 2000
+    },
+    {
+      "entropy": 0.0167919921875,
+      "epoch": 0.41,
+      "grad_norm": 0.009929426945745945,
+      "learning_rate": 9.219034517364369e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965337771177292,
+      "num_tokens": 40251665.0,
+      "step": 2050
+    },
+    {
+      "entropy": 0.0153082275390625,
+      "epoch": 0.42,
+      "grad_norm": 0.007909618318080902,
+      "learning_rate": 9.17503153106797e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965605139732361,
+      "num_tokens": 41243731.0,
+      "step": 2100
+    },
+    {
+      "entropy": 0.0154461669921875,
+      "epoch": 0.43,
+      "grad_norm": 0.015921294689178467,
+      "learning_rate": 9.129933715111125e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9964988273382187,
+      "num_tokens": 42247346.0,
+      "step": 2150
+    },
+    {
+      "entropy": 0.01652587890625,
+      "epoch": 0.44,
+      "grad_norm": 0.006036434322595596,
+      "learning_rate": 9.083752895614464e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964230692386628,
+      "num_tokens": 43220001.0,
+      "step": 2200
+    },
+    {
+      "entropy": 0.015128173828125,
+      "epoch": 0.45,
+      "grad_norm": 0.009778963401913643,
+      "learning_rate": 9.03650118269753e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965469115972518,
+      "num_tokens": 44205197.0,
+      "step": 2250
+    },
+    {
+      "entropy": 0.015570068359375,
+      "epoch": 0.46,
+      "grad_norm": 0.014646470546722412,
+      "learning_rate": 8.988190967303101e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965356832742691,
+      "num_tokens": 45185181.0,
+      "step": 2300
+    },
+    {
+      "entropy": 0.01632568359375,
+      "epoch": 0.47,
+      "grad_norm": 0.006433142349123955,
+      "learning_rate": 8.938834917947889e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965211725234986,
+      "num_tokens": 46170545.0,
+      "step": 2350
+    },
+    {
+      "entropy": 0.01555419921875,
+      "epoch": 0.48,
+      "grad_norm": 0.007329673506319523,
+      "learning_rate": 8.888445977400435e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965089631080627,
+      "num_tokens": 47152165.0,
+      "step": 2400
+    },
+    {
+      "entropy": 0.0156890869140625,
+      "epoch": 0.49,
+      "grad_norm": 0.010034145787358284,
+      "learning_rate": 8.837037359287092e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965383523702621,
+      "num_tokens": 48122023.0,
+      "step": 2450
+    },
+    {
+      "entropy": 0.0163201904296875,
+      "epoch": 0.5,
+      "grad_norm": 0.008338450454175472,
+      "learning_rate": 8.784622544626985e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965311539173126,
+      "num_tokens": 49094218.0,
+      "step": 2500
+    },
+    {
+      "entropy": 0.0152789306640625,
+      "epoch": 0.51,
+      "grad_norm": 0.007771766744554043,
+      "learning_rate": 8.731215278296843e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965461915731431,
+      "num_tokens": 50090248.0,
+      "step": 2550
+    },
+    {
+      "entropy": 0.0171893310546875,
+      "epoch": 0.52,
+      "grad_norm": 0.010534558445215225,
+      "learning_rate": 8.676829565426646e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964153295755387,
+      "num_tokens": 51055771.0,
+      "step": 2600
+    },
+    {
+      "entropy": 0.0148272705078125,
+      "epoch": 0.53,
+      "grad_norm": 0.007072142791002989,
+      "learning_rate": 8.62147966772702e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966214698553085,
+      "num_tokens": 52051029.0,
+      "step": 2650
+    },
+    {
+      "entropy": 0.0146075439453125,
+      "epoch": 0.54,
+      "grad_norm": 0.007414950989186764,
+      "learning_rate": 8.565180099749355e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966325616836548,
+      "num_tokens": 53064519.0,
+      "step": 2700
+    },
+    {
+      "entropy": 0.0161309814453125,
+      "epoch": 0.55,
+      "grad_norm": 0.0070669627748429775,
+      "learning_rate": 8.50794562507961e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9964843678474427,
+      "num_tokens": 54043844.0,
+      "step": 2750
+    },
+    {
+      "entropy": 0.015706787109375,
+      "epoch": 0.56,
+      "grad_norm": 0.005744470749050379,
+      "learning_rate": 8.449791252466819e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965711253881454,
+      "num_tokens": 55026422.0,
+      "step": 2800
+    },
+    {
+      "entropy": 0.01641357421875,
+      "epoch": 0.57,
+      "grad_norm": 0.0069727362133562565,
+      "learning_rate": 8.390732231887314e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.996541822552681,
+      "num_tokens": 56009908.0,
+      "step": 2850
+    },
+    {
+      "entropy": 0.016571044921875,
+      "epoch": 0.58,
+      "grad_norm": 0.007201408036053181,
+      "learning_rate": 8.330784050545672e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964863443374634,
+      "num_tokens": 56948904.0,
+      "step": 2900
+    },
+    {
+      "entropy": 0.0156451416015625,
+      "epoch": 0.59,
+      "grad_norm": 0.005370237864553928,
+      "learning_rate": 8.269962428813474e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966201782226562,
+      "num_tokens": 57925745.0,
+      "step": 2950
+    },
+    {
+      "entropy": 0.01558837890625,
+      "epoch": 0.6,
+      "grad_norm": 0.005006297491490841,
+      "learning_rate": 8.208283316106902e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966268092393875,
+      "num_tokens": 58902798.0,
+      "step": 3000
+    },
+    {
+      "entropy": 0.015733642578125,
+      "epoch": 0.61,
+      "grad_norm": 0.06520090997219086,
+      "learning_rate": 8.145762886704286e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965688252449035,
+      "num_tokens": 59860890.0,
+      "step": 3050
+    },
+    {
+      "entropy": 0.015631103515625,
+      "epoch": 0.62,
+      "grad_norm": 0.017850181087851524,
+      "learning_rate": 8.082417535504683e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9963969469070435,
+      "num_tokens": 60845473.0,
+      "step": 3100
+    },
+    {
+      "entropy": 0.0175927734375,
+      "epoch": 0.63,
+      "grad_norm": 0.017514687031507492,
+      "learning_rate": 8.018263873728585e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9963253819942475,
+      "num_tokens": 61784192.0,
+      "step": 3150
+    },
+    {
+      "entropy": 0.0167742919921875,
+      "epoch": 0.64,
+      "grad_norm": 0.008665528148412704,
+      "learning_rate": 7.953318724561932e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964195656776428,
+      "num_tokens": 62766207.0,
+      "step": 3200
+    },
+    {
+      "entropy": 0.0157501220703125,
+      "epoch": 0.65,
+      "grad_norm": 0.019448991864919662,
+      "learning_rate": 7.887599118744509e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966711962223053,
+      "num_tokens": 63762073.0,
+      "step": 3250
+    },
+    {
+      "entropy": 0.0180609130859375,
+      "epoch": 0.66,
+      "grad_norm": 0.004935233388096094,
+      "learning_rate": 7.821122290103938e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964338165521621,
+      "num_tokens": 64710323.0,
+      "step": 3300
+    },
+    {
+      "entropy": 0.015982666015625,
+      "epoch": 0.67,
+      "grad_norm": 0.004395989701151848,
+      "learning_rate": 7.753905671036403e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966453295946122,
+      "num_tokens": 65703355.0,
+      "step": 3350
+    },
+    {
+      "entropy": 0.0160992431640625,
+      "epoch": 0.68,
+      "grad_norm": 0.01108239684253931,
+      "learning_rate": 7.685966887935309e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996572095155716,
+      "num_tokens": 66693270.0,
+      "step": 3400
+    },
+    {
+      "entropy": 0.0156439208984375,
+      "epoch": 0.69,
+      "grad_norm": 0.007867238484323025,
+      "learning_rate": 7.617323756569053e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965278363227844,
+      "num_tokens": 67662491.0,
+      "step": 3450
+    },
+    {
+      "entropy": 0.0155462646484375,
+      "epoch": 0.7,
+      "grad_norm": 0.006034135352820158,
+      "learning_rate": 7.547994277409168e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966409939527512,
+      "num_tokens": 68643197.0,
+      "step": 3500
+    },
+    {
+      "entropy": 0.015638427734375,
+      "epoch": 0.71,
+      "grad_norm": 0.007907884195446968,
+      "learning_rate": 7.477996630909994e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966798382997513,
+      "num_tokens": 69638359.0,
+      "step": 3550
+    },
+    {
+      "entropy": 0.0161322021484375,
+      "epoch": 0.72,
+      "grad_norm": 0.007897689007222652,
+      "learning_rate": 7.40734917274118e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966677987575531,
+      "num_tokens": 70621331.0,
+      "step": 3600
+    },
+    {
+      "entropy": 0.015335693359375,
+      "epoch": 0.73,
+      "grad_norm": 0.006248582154512405,
+      "learning_rate": 7.336070428974218e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966181468963623,
+      "num_tokens": 71609172.0,
+      "step": 3650
+    },
+    {
+      "entropy": 0.0166754150390625,
+      "epoch": 0.74,
+      "grad_norm": 0.009811230935156345,
+      "learning_rate": 7.26417909122431e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9965909200906754,
+      "num_tokens": 72589978.0,
+      "step": 3700
+    },
+    {
+      "entropy": 0.0143695068359375,
+      "epoch": 0.75,
+      "grad_norm": 0.012172735296189785,
+      "learning_rate": 7.191694011748818e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965182906389236,
+      "num_tokens": 73606622.0,
+      "step": 3750
+    },
+    {
+      "entropy": 0.0156829833984375,
+      "epoch": 0.76,
+      "grad_norm": 0.004879123531281948,
+      "learning_rate": 7.118634198503571e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965479630231857,
+      "num_tokens": 74626982.0,
+      "step": 3800
+    },
+    {
+      "entropy": 0.016346435546875,
+      "epoch": 0.77,
+      "grad_norm": 0.005593061912804842,
+      "learning_rate": 7.045018810158375e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966219502687454,
+      "num_tokens": 75622857.0,
+      "step": 3850
+    },
+    {
+      "entropy": 0.015889892578125,
+      "epoch": 0.78,
+      "grad_norm": 0.006423663813620806,
+      "learning_rate": 6.97086715107298e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966537064313888,
+      "num_tokens": 76601614.0,
+      "step": 3900
+    },
+    {
+      "entropy": 0.015166015625,
+      "epoch": 0.79,
+      "grad_norm": 0.009724115021526814,
+      "learning_rate": 6.896198666234833e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966633880138397,
+      "num_tokens": 77598059.0,
+      "step": 3950
+    },
+    {
+      "entropy": 0.0157177734375,
+      "epoch": 0.8,
+      "grad_norm": 0.014047837816178799,
+      "learning_rate": 6.821032936159986e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996613291501999,
+      "num_tokens": 78570180.0,
+      "step": 4000
+    }
+  ],
+  "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": 1.3645321080343101e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-4000/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_lr1e-4/checkpoint-5000/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_lr1e-4/checkpoint-5000/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": [
+    "up_proj",
+    "q_proj",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "v_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_lr1e-4/checkpoint-5000/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_lr1e-4/checkpoint-5000/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_lr1e-4/checkpoint-5000/latest

@@ -0,0 +1 @@
+global_step5000

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-5000/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_lr1e-4/checkpoint-5000/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_lr1e-4/checkpoint-5000/trainer_state.json

@@ -0,0 +1,1034 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.0,
+  "eval_steps": 500,
+  "global_step": 5000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "entropy": 0.166884765625,
+      "epoch": 0.01,
+      "grad_norm": 0.09714163839817047,
+      "learning_rate": 1.6333333333333335e-05,
+      "loss": 0.1512,
+      "mean_token_accuracy": 0.966504562497139,
+      "num_tokens": 973322.0,
+      "step": 50
+    },
+    {
+      "entropy": 0.048140869140625,
+      "epoch": 0.02,
+      "grad_norm": 0.057216521352529526,
+      "learning_rate": 3.3e-05,
+      "loss": 0.0424,
+      "mean_token_accuracy": 0.9896974349021912,
+      "num_tokens": 1965908.0,
+      "step": 100
+    },
+    {
+      "entropy": 0.029632568359375,
+      "epoch": 0.03,
+      "grad_norm": 0.054412033408880234,
+      "learning_rate": 4.966666666666667e-05,
+      "loss": 0.0248,
+      "mean_token_accuracy": 0.9936319047212601,
+      "num_tokens": 2931943.0,
+      "step": 150
+    },
+    {
+      "entropy": 0.02478515625,
+      "epoch": 0.04,
+      "grad_norm": 0.08800562471151352,
+      "learning_rate": 6.633333333333334e-05,
+      "loss": 0.0214,
+      "mean_token_accuracy": 0.9943305045366287,
+      "num_tokens": 3906979.0,
+      "step": 200
+    },
+    {
+      "entropy": 0.0224652099609375,
+      "epoch": 0.05,
+      "grad_norm": 0.041793134063482285,
+      "learning_rate": 8.3e-05,
+      "loss": 0.0191,
+      "mean_token_accuracy": 0.9948792272806167,
+      "num_tokens": 4891791.0,
+      "step": 250
+    },
+    {
+      "entropy": 0.01969482421875,
+      "epoch": 0.06,
+      "grad_norm": 0.04181526601314545,
+      "learning_rate": 9.966666666666667e-05,
+      "loss": 0.0173,
+      "mean_token_accuracy": 0.995435175895691,
+      "num_tokens": 5923811.0,
+      "step": 300
+    },
+    {
+      "entropy": 0.019368896484375,
+      "epoch": 0.07,
+      "grad_norm": 0.03995412588119507,
+      "learning_rate": 9.999370378817499e-05,
+      "loss": 0.0164,
+      "mean_token_accuracy": 0.9956372672319412,
+      "num_tokens": 6923169.0,
+      "step": 350
+    },
+    {
+      "entropy": 0.0204400634765625,
+      "epoch": 0.08,
+      "grad_norm": 0.06872096657752991,
+      "learning_rate": 9.997430021636957e-05,
+      "loss": 0.0163,
+      "mean_token_accuracy": 0.9956023478507996,
+      "num_tokens": 7901360.0,
+      "step": 400
+    },
+    {
+      "entropy": 0.0186175537109375,
+      "epoch": 0.09,
+      "grad_norm": 0.037068989127874374,
+      "learning_rate": 9.994179175045023e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958296239376068,
+      "num_tokens": 8899074.0,
+      "step": 450
+    },
+    {
+      "entropy": 0.0173492431640625,
+      "epoch": 0.1,
+      "grad_norm": 0.014651700854301453,
+      "learning_rate": 9.989618691519873e-05,
+      "loss": 0.015,
+      "mean_token_accuracy": 0.9959587389230729,
+      "num_tokens": 9909340.0,
+      "step": 500
+    },
+    {
+      "entropy": 0.0183203125,
+      "epoch": 0.11,
+      "grad_norm": 0.016151444986462593,
+      "learning_rate": 9.983749766969271e-05,
+      "loss": 0.0151,
+      "mean_token_accuracy": 0.9959770923852921,
+      "num_tokens": 10894801.0,
+      "step": 550
+    },
+    {
+      "entropy": 0.02047607421875,
+      "epoch": 0.12,
+      "grad_norm": 0.021302781999111176,
+      "learning_rate": 9.976573940416966e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958055526018142,
+      "num_tokens": 11862148.0,
+      "step": 600
+    },
+    {
+      "entropy": 0.017862548828125,
+      "epoch": 0.13,
+      "grad_norm": 0.013205700553953648,
+      "learning_rate": 9.968093093599106e-05,
+      "loss": 0.0152,
+      "mean_token_accuracy": 0.9958789277076722,
+      "num_tokens": 12851463.0,
+      "step": 650
+    },
+    {
+      "entropy": 0.0182177734375,
+      "epoch": 0.14,
+      "grad_norm": 0.013120726682245731,
+      "learning_rate": 9.958309450470784e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9962116169929505,
+      "num_tokens": 13849228.0,
+      "step": 700
+    },
+    {
+      "entropy": 0.0179010009765625,
+      "epoch": 0.15,
+      "grad_norm": 0.014701537787914276,
+      "learning_rate": 9.947225576622847e-05,
+      "loss": 0.0146,
+      "mean_token_accuracy": 0.9961142331361771,
+      "num_tokens": 14812166.0,
+      "step": 750
+    },
+    {
+      "entropy": 0.0173638916015625,
+      "epoch": 0.16,
+      "grad_norm": 0.033504419028759,
+      "learning_rate": 9.934844378609117e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.9961155968904495,
+      "num_tokens": 15789120.0,
+      "step": 800
+    },
+    {
+      "entropy": 0.0183349609375,
+      "epoch": 0.17,
+      "grad_norm": 0.010470729321241379,
+      "learning_rate": 9.921169103184187e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9961740493774414,
+      "num_tokens": 16764645.0,
+      "step": 850
+    },
+    {
+      "entropy": 0.016279296875,
+      "epoch": 0.18,
+      "grad_norm": 0.02781980112195015,
+      "learning_rate": 9.906203336452029e-05,
+      "loss": 0.0142,
+      "mean_token_accuracy": 0.9962105673551559,
+      "num_tokens": 17730003.0,
+      "step": 900
+    },
+    {
+      "entropy": 0.0166461181640625,
+      "epoch": 0.19,
+      "grad_norm": 0.020350394770503044,
+      "learning_rate": 9.889951002925593e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962483876943589,
+      "num_tokens": 18723861.0,
+      "step": 950
+    },
+    {
+      "entropy": 0.0165386962890625,
+      "epoch": 0.2,
+      "grad_norm": 0.013212243095040321,
+      "learning_rate": 9.872416364497675e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9961817175149917,
+      "num_tokens": 19695091.0,
+      "step": 1000
+    },
+    {
+      "entropy": 0.017391357421875,
+      "epoch": 0.21,
+      "grad_norm": 0.014041919261217117,
+      "learning_rate": 9.853604019323301e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962570595741272,
+      "num_tokens": 20662255.0,
+      "step": 1050
+    },
+    {
+      "entropy": 0.016884765625,
+      "epoch": 0.22,
+      "grad_norm": 0.017174789682030678,
+      "learning_rate": 9.833518900613956e-05,
+      "loss": 0.0138,
+      "mean_token_accuracy": 0.9962636953592301,
+      "num_tokens": 21642092.0,
+      "step": 1100
+    },
+    {
+      "entropy": 0.019366455078125,
+      "epoch": 0.23,
+      "grad_norm": 0.017337976023554802,
+      "learning_rate": 9.812166275343917e-05,
+      "loss": 0.0143,
+      "mean_token_accuracy": 0.996159919500351,
+      "num_tokens": 22607505.0,
+      "step": 1150
+    },
+    {
+      "entropy": 0.0164886474609375,
+      "epoch": 0.24,
+      "grad_norm": 0.018486863002181053,
+      "learning_rate": 9.789551742869096e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963399863243103,
+      "num_tokens": 23584579.0,
+      "step": 1200
+    },
+    {
+      "entropy": 0.016282958984375,
+      "epoch": 0.25,
+      "grad_norm": 0.009195365011692047,
+      "learning_rate": 9.765681233458693e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9963367432355881,
+      "num_tokens": 24580285.0,
+      "step": 1250
+    },
+    {
+      "entropy": 0.0197235107421875,
+      "epoch": 0.26,
+      "grad_norm": 0.01205432415008545,
+      "learning_rate": 9.740561006740098e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.996049216389656,
+      "num_tokens": 25525513.0,
+      "step": 1300
+    },
+    {
+      "entropy": 0.0153009033203125,
+      "epoch": 0.27,
+      "grad_norm": 0.01174497976899147,
+      "learning_rate": 9.7141976500574e-05,
+      "loss": 0.0135,
+      "mean_token_accuracy": 0.9964245760440826,
+      "num_tokens": 26523426.0,
+      "step": 1350
+    },
+    {
+      "entropy": 0.0169378662109375,
+      "epoch": 0.28,
+      "grad_norm": 0.014242298901081085,
+      "learning_rate": 9.686598076743981e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962529343366623,
+      "num_tokens": 27493091.0,
+      "step": 1400
+    },
+    {
+      "entropy": 0.017811279296875,
+      "epoch": 0.29,
+      "grad_norm": 0.0160844549536705,
+      "learning_rate": 9.657769524309605e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.996181645989418,
+      "num_tokens": 28458553.0,
+      "step": 1450
+    },
+    {
+      "entropy": 0.0169232177734375,
+      "epoch": 0.3,
+      "grad_norm": 0.03241236135363579,
+      "learning_rate": 9.627719552542516e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9962622857093811,
+      "num_tokens": 29430497.0,
+      "step": 1500
+    },
+    {
+      "entropy": 0.0163720703125,
+      "epoch": 0.31,
+      "grad_norm": 0.01750756986439228,
+      "learning_rate": 9.596456041527001e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9964328509569168,
+      "num_tokens": 30414152.0,
+      "step": 1550
+    },
+    {
+      "entropy": 0.017760009765625,
+      "epoch": 0.32,
+      "grad_norm": 0.02182549238204956,
+      "learning_rate": 9.563987189576991e-05,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.9961263346672058,
+      "num_tokens": 31396417.0,
+      "step": 1600
+    },
+    {
+      "entropy": 0.0171954345703125,
+      "epoch": 0.33,
+      "grad_norm": 0.011301269754767418,
+      "learning_rate": 9.530321511086183e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962880289554596,
+      "num_tokens": 32364443.0,
+      "step": 1650
+    },
+    {
+      "entropy": 0.015126953125,
+      "epoch": 0.34,
+      "grad_norm": 0.009061276912689209,
+      "learning_rate": 9.495467834295291e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9964066076278687,
+      "num_tokens": 33374821.0,
+      "step": 1700
+    },
+    {
+      "entropy": 0.018043212890625,
+      "epoch": 0.35,
+      "grad_norm": 0.014127642847597599,
+      "learning_rate": 9.459435298976998e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963885354995727,
+      "num_tokens": 34338570.0,
+      "step": 1750
+    },
+    {
+      "entropy": 0.0155584716796875,
+      "epoch": 0.36,
+      "grad_norm": 0.00913177989423275,
+      "learning_rate": 9.422233354039198e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965262812376022,
+      "num_tokens": 35333253.0,
+      "step": 1800
+    },
+    {
+      "entropy": 0.0177911376953125,
+      "epoch": 0.37,
+      "grad_norm": 0.009629561565816402,
+      "learning_rate": 9.383871755047198e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9964758467674255,
+      "num_tokens": 36316255.0,
+      "step": 1850
+    },
+    {
+      "entropy": 0.0159539794921875,
+      "epoch": 0.38,
+      "grad_norm": 0.01395715307444334,
+      "learning_rate": 9.344360561665476e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964872688055039,
+      "num_tokens": 37303247.0,
+      "step": 1900
+    },
+    {
+      "entropy": 0.0152783203125,
+      "epoch": 0.39,
+      "grad_norm": 0.007368483114987612,
+      "learning_rate": 9.30371013501972e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965988719463348,
+      "num_tokens": 38287463.0,
+      "step": 1950
+    },
+    {
+      "entropy": 0.0158270263671875,
+      "epoch": 0.4,
+      "grad_norm": 0.01508031040430069,
+      "learning_rate": 9.261931134979791e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9965020543336869,
+      "num_tokens": 39273383.0,
+      "step": 2000
+    },
+    {
+      "entropy": 0.0167919921875,
+      "epoch": 0.41,
+      "grad_norm": 0.009929426945745945,
+      "learning_rate": 9.219034517364369e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965337771177292,
+      "num_tokens": 40251665.0,
+      "step": 2050
+    },
+    {
+      "entropy": 0.0153082275390625,
+      "epoch": 0.42,
+      "grad_norm": 0.007909618318080902,
+      "learning_rate": 9.17503153106797e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965605139732361,
+      "num_tokens": 41243731.0,
+      "step": 2100
+    },
+    {
+      "entropy": 0.0154461669921875,
+      "epoch": 0.43,
+      "grad_norm": 0.015921294689178467,
+      "learning_rate": 9.129933715111125e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9964988273382187,
+      "num_tokens": 42247346.0,
+      "step": 2150
+    },
+    {
+      "entropy": 0.01652587890625,
+      "epoch": 0.44,
+      "grad_norm": 0.006036434322595596,
+      "learning_rate": 9.083752895614464e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964230692386628,
+      "num_tokens": 43220001.0,
+      "step": 2200
+    },
+    {
+      "entropy": 0.015128173828125,
+      "epoch": 0.45,
+      "grad_norm": 0.009778963401913643,
+      "learning_rate": 9.03650118269753e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965469115972518,
+      "num_tokens": 44205197.0,
+      "step": 2250
+    },
+    {
+      "entropy": 0.015570068359375,
+      "epoch": 0.46,
+      "grad_norm": 0.014646470546722412,
+      "learning_rate": 8.988190967303101e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965356832742691,
+      "num_tokens": 45185181.0,
+      "step": 2300
+    },
+    {
+      "entropy": 0.01632568359375,
+      "epoch": 0.47,
+      "grad_norm": 0.006433142349123955,
+      "learning_rate": 8.938834917947889e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965211725234986,
+      "num_tokens": 46170545.0,
+      "step": 2350
+    },
+    {
+      "entropy": 0.01555419921875,
+      "epoch": 0.48,
+      "grad_norm": 0.007329673506319523,
+      "learning_rate": 8.888445977400435e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965089631080627,
+      "num_tokens": 47152165.0,
+      "step": 2400
+    },
+    {
+      "entropy": 0.0156890869140625,
+      "epoch": 0.49,
+      "grad_norm": 0.010034145787358284,
+      "learning_rate": 8.837037359287092e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965383523702621,
+      "num_tokens": 48122023.0,
+      "step": 2450
+    },
+    {
+      "entropy": 0.0163201904296875,
+      "epoch": 0.5,
+      "grad_norm": 0.008338450454175472,
+      "learning_rate": 8.784622544626985e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965311539173126,
+      "num_tokens": 49094218.0,
+      "step": 2500
+    },
+    {
+      "entropy": 0.0152789306640625,
+      "epoch": 0.51,
+      "grad_norm": 0.007771766744554043,
+      "learning_rate": 8.731215278296843e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965461915731431,
+      "num_tokens": 50090248.0,
+      "step": 2550
+    },
+    {
+      "entropy": 0.0171893310546875,
+      "epoch": 0.52,
+      "grad_norm": 0.010534558445215225,
+      "learning_rate": 8.676829565426646e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964153295755387,
+      "num_tokens": 51055771.0,
+      "step": 2600
+    },
+    {
+      "entropy": 0.0148272705078125,
+      "epoch": 0.53,
+      "grad_norm": 0.007072142791002989,
+      "learning_rate": 8.62147966772702e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966214698553085,
+      "num_tokens": 52051029.0,
+      "step": 2650
+    },
+    {
+      "entropy": 0.0146075439453125,
+      "epoch": 0.54,
+      "grad_norm": 0.007414950989186764,
+      "learning_rate": 8.565180099749355e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966325616836548,
+      "num_tokens": 53064519.0,
+      "step": 2700
+    },
+    {
+      "entropy": 0.0161309814453125,
+      "epoch": 0.55,
+      "grad_norm": 0.0070669627748429775,
+      "learning_rate": 8.50794562507961e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9964843678474427,
+      "num_tokens": 54043844.0,
+      "step": 2750
+    },
+    {
+      "entropy": 0.015706787109375,
+      "epoch": 0.56,
+      "grad_norm": 0.005744470749050379,
+      "learning_rate": 8.449791252466819e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965711253881454,
+      "num_tokens": 55026422.0,
+      "step": 2800
+    },
+    {
+      "entropy": 0.01641357421875,
+      "epoch": 0.57,
+      "grad_norm": 0.0069727362133562565,
+      "learning_rate": 8.390732231887314e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.996541822552681,
+      "num_tokens": 56009908.0,
+      "step": 2850
+    },
+    {
+      "entropy": 0.016571044921875,
+      "epoch": 0.58,
+      "grad_norm": 0.007201408036053181,
+      "learning_rate": 8.330784050545672e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964863443374634,
+      "num_tokens": 56948904.0,
+      "step": 2900
+    },
+    {
+      "entropy": 0.0156451416015625,
+      "epoch": 0.59,
+      "grad_norm": 0.005370237864553928,
+      "learning_rate": 8.269962428813474e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966201782226562,
+      "num_tokens": 57925745.0,
+      "step": 2950
+    },
+    {
+      "entropy": 0.01558837890625,
+      "epoch": 0.6,
+      "grad_norm": 0.005006297491490841,
+      "learning_rate": 8.208283316106902e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966268092393875,
+      "num_tokens": 58902798.0,
+      "step": 3000
+    },
+    {
+      "entropy": 0.015733642578125,
+      "epoch": 0.61,
+      "grad_norm": 0.06520090997219086,
+      "learning_rate": 8.145762886704286e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965688252449035,
+      "num_tokens": 59860890.0,
+      "step": 3050
+    },
+    {
+      "entropy": 0.015631103515625,
+      "epoch": 0.62,
+      "grad_norm": 0.017850181087851524,
+      "learning_rate": 8.082417535504683e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9963969469070435,
+      "num_tokens": 60845473.0,
+      "step": 3100
+    },
+    {
+      "entropy": 0.0175927734375,
+      "epoch": 0.63,
+      "grad_norm": 0.017514687031507492,
+      "learning_rate": 8.018263873728585e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9963253819942475,
+      "num_tokens": 61784192.0,
+      "step": 3150
+    },
+    {
+      "entropy": 0.0167742919921875,
+      "epoch": 0.64,
+      "grad_norm": 0.008665528148412704,
+      "learning_rate": 7.953318724561932e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964195656776428,
+      "num_tokens": 62766207.0,
+      "step": 3200
+    },
+    {
+      "entropy": 0.0157501220703125,
+      "epoch": 0.65,
+      "grad_norm": 0.019448991864919662,
+      "learning_rate": 7.887599118744509e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966711962223053,
+      "num_tokens": 63762073.0,
+      "step": 3250
+    },
+    {
+      "entropy": 0.0180609130859375,
+      "epoch": 0.66,
+      "grad_norm": 0.004935233388096094,
+      "learning_rate": 7.821122290103938e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964338165521621,
+      "num_tokens": 64710323.0,
+      "step": 3300
+    },
+    {
+      "entropy": 0.015982666015625,
+      "epoch": 0.67,
+      "grad_norm": 0.004395989701151848,
+      "learning_rate": 7.753905671036403e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966453295946122,
+      "num_tokens": 65703355.0,
+      "step": 3350
+    },
+    {
+      "entropy": 0.0160992431640625,
+      "epoch": 0.68,
+      "grad_norm": 0.01108239684253931,
+      "learning_rate": 7.685966887935309e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996572095155716,
+      "num_tokens": 66693270.0,
+      "step": 3400
+    },
+    {
+      "entropy": 0.0156439208984375,
+      "epoch": 0.69,
+      "grad_norm": 0.007867238484323025,
+      "learning_rate": 7.617323756569053e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965278363227844,
+      "num_tokens": 67662491.0,
+      "step": 3450
+    },
+    {
+      "entropy": 0.0155462646484375,
+      "epoch": 0.7,
+      "grad_norm": 0.006034135352820158,
+      "learning_rate": 7.547994277409168e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966409939527512,
+      "num_tokens": 68643197.0,
+      "step": 3500
+    },
+    {
+      "entropy": 0.015638427734375,
+      "epoch": 0.71,
+      "grad_norm": 0.007907884195446968,
+      "learning_rate": 7.477996630909994e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966798382997513,
+      "num_tokens": 69638359.0,
+      "step": 3550
+    },
+    {
+      "entropy": 0.0161322021484375,
+      "epoch": 0.72,
+      "grad_norm": 0.007897689007222652,
+      "learning_rate": 7.40734917274118e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966677987575531,
+      "num_tokens": 70621331.0,
+      "step": 3600
+    },
+    {
+      "entropy": 0.015335693359375,
+      "epoch": 0.73,
+      "grad_norm": 0.006248582154512405,
+      "learning_rate": 7.336070428974218e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966181468963623,
+      "num_tokens": 71609172.0,
+      "step": 3650
+    },
+    {
+      "entropy": 0.0166754150390625,
+      "epoch": 0.74,
+      "grad_norm": 0.009811230935156345,
+      "learning_rate": 7.26417909122431e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9965909200906754,
+      "num_tokens": 72589978.0,
+      "step": 3700
+    },
+    {
+      "entropy": 0.0143695068359375,
+      "epoch": 0.75,
+      "grad_norm": 0.012172735296189785,
+      "learning_rate": 7.191694011748818e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965182906389236,
+      "num_tokens": 73606622.0,
+      "step": 3750
+    },
+    {
+      "entropy": 0.0156829833984375,
+      "epoch": 0.76,
+      "grad_norm": 0.004879123531281948,
+      "learning_rate": 7.118634198503571e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965479630231857,
+      "num_tokens": 74626982.0,
+      "step": 3800
+    },
+    {
+      "entropy": 0.016346435546875,
+      "epoch": 0.77,
+      "grad_norm": 0.005593061912804842,
+      "learning_rate": 7.045018810158375e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966219502687454,
+      "num_tokens": 75622857.0,
+      "step": 3850
+    },
+    {
+      "entropy": 0.015889892578125,
+      "epoch": 0.78,
+      "grad_norm": 0.006423663813620806,
+      "learning_rate": 6.97086715107298e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966537064313888,
+      "num_tokens": 76601614.0,
+      "step": 3900
+    },
+    {
+      "entropy": 0.015166015625,
+      "epoch": 0.79,
+      "grad_norm": 0.009724115021526814,
+      "learning_rate": 6.896198666234833e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966633880138397,
+      "num_tokens": 77598059.0,
+      "step": 3950
+    },
+    {
+      "entropy": 0.0157177734375,
+      "epoch": 0.8,
+      "grad_norm": 0.014047837816178799,
+      "learning_rate": 6.821032936159986e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996613291501999,
+      "num_tokens": 78570180.0,
+      "step": 4000
+    },
+    {
+      "entropy": 0.0155902099609375,
+      "epoch": 0.81,
+      "grad_norm": 0.008200579322874546,
+      "learning_rate": 6.745389671758435e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9965583562850953,
+      "num_tokens": 79561943.0,
+      "step": 4050
+    },
+    {
+      "entropy": 0.014603271484375,
+      "epoch": 0.82,
+      "grad_norm": 0.0051344577223062515,
+      "learning_rate": 6.669288709165276e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967030560970307,
+      "num_tokens": 80576757.0,
+      "step": 4100
+    },
+    {
+      "entropy": 0.0161895751953125,
+      "epoch": 0.83,
+      "grad_norm": 0.00540529889985919,
+      "learning_rate": 6.59275000453902e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966316163539887,
+      "num_tokens": 81549255.0,
+      "step": 4150
+    },
+    {
+      "entropy": 0.01657470703125,
+      "epoch": 0.84,
+      "grad_norm": 0.005644636228680611,
+      "learning_rate": 6.515793628828447e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965110903978348,
+      "num_tokens": 82499683.0,
+      "step": 4200
+    },
+    {
+      "entropy": 0.0160162353515625,
+      "epoch": 0.85,
+      "grad_norm": 0.004190398845821619,
+      "learning_rate": 6.438439762509332e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966331452131272,
+      "num_tokens": 83475810.0,
+      "step": 4250
+    },
+    {
+      "entropy": 0.01598388671875,
+      "epoch": 0.86,
+      "grad_norm": 0.005400759633630514,
+      "learning_rate": 6.360708690292479e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966162234544754,
+      "num_tokens": 84432102.0,
+      "step": 4300
+    },
+    {
+      "entropy": 0.014876708984375,
+      "epoch": 0.87,
+      "grad_norm": 0.012628648430109024,
+      "learning_rate": 6.2826207958044e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967326593399047,
+      "num_tokens": 85429626.0,
+      "step": 4350
+    },
+    {
+      "entropy": 0.014339599609375,
+      "epoch": 0.88,
+      "grad_norm": 0.009649231098592281,
+      "learning_rate": 6.204196556242061e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967543077468872,
+      "num_tokens": 86450572.0,
+      "step": 4400
+    },
+    {
+      "entropy": 0.014298095703125,
+      "epoch": 0.89,
+      "grad_norm": 0.0049432734958827496,
+      "learning_rate": 6.125456537003095e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967639350891113,
+      "num_tokens": 87458843.0,
+      "step": 4450
+    },
+    {
+      "entropy": 0.015306396484375,
+      "epoch": 0.9,
+      "grad_norm": 0.004098173696547747,
+      "learning_rate": 6.046421386292887e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9966574615240097,
+      "num_tokens": 88453316.0,
+      "step": 4500
+    },
+    {
+      "entropy": 0.016375732421875,
+      "epoch": 0.91,
+      "grad_norm": 0.005302398465573788,
+      "learning_rate": 5.9671118297099274e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966228741407395,
+      "num_tokens": 89427450.0,
+      "step": 4550
+    },
+    {
+      "entropy": 0.015936279296875,
+      "epoch": 0.92,
+      "grad_norm": 0.004837509244680405,
+      "learning_rate": 5.887548664810896e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966443425416946,
+      "num_tokens": 90400676.0,
+      "step": 4600
+    },
+    {
+      "entropy": 0.0147259521484375,
+      "epoch": 0.93,
+      "grad_norm": 0.008831442333757877,
+      "learning_rate": 5.8077527556568614e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967519730329514,
+      "num_tokens": 91394757.0,
+      "step": 4650
+    },
+    {
+      "entropy": 0.01557373046875,
+      "epoch": 0.94,
+      "grad_norm": 0.0063737258315086365,
+      "learning_rate": 5.727745027342031e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966205650568009,
+      "num_tokens": 92386137.0,
+      "step": 4700
+    },
+    {
+      "entropy": 0.0151513671875,
+      "epoch": 0.95,
+      "grad_norm": 0.005380143877118826,
+      "learning_rate": 5.64754646050652e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9965577948093415,
+      "num_tokens": 93374756.0,
+      "step": 4750
+    },
+    {
+      "entropy": 0.0138897705078125,
+      "epoch": 0.96,
+      "grad_norm": 0.005837676115334034,
+      "learning_rate": 5.567178085834542e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967484200000762,
+      "num_tokens": 94389846.0,
+      "step": 4800
+    },
+    {
+      "entropy": 0.0154791259765625,
+      "epoch": 0.97,
+      "grad_norm": 0.008733139373362064,
+      "learning_rate": 5.486660978539468e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966599184274674,
+      "num_tokens": 95378637.0,
+      "step": 4850
+    },
+    {
+      "entropy": 0.015618896484375,
+      "epoch": 0.98,
+      "grad_norm": 0.016466081142425537,
+      "learning_rate": 5.406016252837219e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966346627473831,
+      "num_tokens": 96357363.0,
+      "step": 4900
+    },
+    {
+      "entropy": 0.0171966552734375,
+      "epoch": 0.99,
+      "grad_norm": 0.007259083911776543,
+      "learning_rate": 5.325265056409442e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965799397230148,
+      "num_tokens": 97295349.0,
+      "step": 4950
+    },
+    {
+      "entropy": 0.01597900390625,
+      "epoch": 1.0,
+      "grad_norm": 0.007410865277051926,
+      "learning_rate": 5.244428564857874e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.996621105670929,
+      "num_tokens": 98262502.0,
+      "step": 5000
+    }
+  ],
+  "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": 1.7065296121015828e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-5000/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_lr1e-4/checkpoint-6000/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": [
+    "up_proj",
+    "q_proj",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "v_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_lr1e-4/checkpoint-6000/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_lr1e-4/checkpoint-6000/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_lr1e-4/checkpoint-6000/latest

@@ -0,0 +1 @@
+global_step6000

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-6000/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_lr1e-4/checkpoint-6000/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_lr1e-4/checkpoint-7000/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": [
+    "up_proj",
+    "q_proj",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "v_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_lr1e-4/checkpoint-7000/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_lr1e-4/checkpoint-7000/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_lr1e-4/checkpoint-7000/latest

@@ -0,0 +1 @@
+global_step7000

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-7000/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_lr1e-4/checkpoint-7000/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_lr1e-4/checkpoint-7000/trainer_state.json

@@ -0,0 +1,1434 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.4,
+  "eval_steps": 500,
+  "global_step": 7000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "entropy": 0.166884765625,
+      "epoch": 0.01,
+      "grad_norm": 0.09714163839817047,
+      "learning_rate": 1.6333333333333335e-05,
+      "loss": 0.1512,
+      "mean_token_accuracy": 0.966504562497139,
+      "num_tokens": 973322.0,
+      "step": 50
+    },
+    {
+      "entropy": 0.048140869140625,
+      "epoch": 0.02,
+      "grad_norm": 0.057216521352529526,
+      "learning_rate": 3.3e-05,
+      "loss": 0.0424,
+      "mean_token_accuracy": 0.9896974349021912,
+      "num_tokens": 1965908.0,
+      "step": 100
+    },
+    {
+      "entropy": 0.029632568359375,
+      "epoch": 0.03,
+      "grad_norm": 0.054412033408880234,
+      "learning_rate": 4.966666666666667e-05,
+      "loss": 0.0248,
+      "mean_token_accuracy": 0.9936319047212601,
+      "num_tokens": 2931943.0,
+      "step": 150
+    },
+    {
+      "entropy": 0.02478515625,
+      "epoch": 0.04,
+      "grad_norm": 0.08800562471151352,
+      "learning_rate": 6.633333333333334e-05,
+      "loss": 0.0214,
+      "mean_token_accuracy": 0.9943305045366287,
+      "num_tokens": 3906979.0,
+      "step": 200
+    },
+    {
+      "entropy": 0.0224652099609375,
+      "epoch": 0.05,
+      "grad_norm": 0.041793134063482285,
+      "learning_rate": 8.3e-05,
+      "loss": 0.0191,
+      "mean_token_accuracy": 0.9948792272806167,
+      "num_tokens": 4891791.0,
+      "step": 250
+    },
+    {
+      "entropy": 0.01969482421875,
+      "epoch": 0.06,
+      "grad_norm": 0.04181526601314545,
+      "learning_rate": 9.966666666666667e-05,
+      "loss": 0.0173,
+      "mean_token_accuracy": 0.995435175895691,
+      "num_tokens": 5923811.0,
+      "step": 300
+    },
+    {
+      "entropy": 0.019368896484375,
+      "epoch": 0.07,
+      "grad_norm": 0.03995412588119507,
+      "learning_rate": 9.999370378817499e-05,
+      "loss": 0.0164,
+      "mean_token_accuracy": 0.9956372672319412,
+      "num_tokens": 6923169.0,
+      "step": 350
+    },
+    {
+      "entropy": 0.0204400634765625,
+      "epoch": 0.08,
+      "grad_norm": 0.06872096657752991,
+      "learning_rate": 9.997430021636957e-05,
+      "loss": 0.0163,
+      "mean_token_accuracy": 0.9956023478507996,
+      "num_tokens": 7901360.0,
+      "step": 400
+    },
+    {
+      "entropy": 0.0186175537109375,
+      "epoch": 0.09,
+      "grad_norm": 0.037068989127874374,
+      "learning_rate": 9.994179175045023e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958296239376068,
+      "num_tokens": 8899074.0,
+      "step": 450
+    },
+    {
+      "entropy": 0.0173492431640625,
+      "epoch": 0.1,
+      "grad_norm": 0.014651700854301453,
+      "learning_rate": 9.989618691519873e-05,
+      "loss": 0.015,
+      "mean_token_accuracy": 0.9959587389230729,
+      "num_tokens": 9909340.0,
+      "step": 500
+    },
+    {
+      "entropy": 0.0183203125,
+      "epoch": 0.11,
+      "grad_norm": 0.016151444986462593,
+      "learning_rate": 9.983749766969271e-05,
+      "loss": 0.0151,
+      "mean_token_accuracy": 0.9959770923852921,
+      "num_tokens": 10894801.0,
+      "step": 550
+    },
+    {
+      "entropy": 0.02047607421875,
+      "epoch": 0.12,
+      "grad_norm": 0.021302781999111176,
+      "learning_rate": 9.976573940416966e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958055526018142,
+      "num_tokens": 11862148.0,
+      "step": 600
+    },
+    {
+      "entropy": 0.017862548828125,
+      "epoch": 0.13,
+      "grad_norm": 0.013205700553953648,
+      "learning_rate": 9.968093093599106e-05,
+      "loss": 0.0152,
+      "mean_token_accuracy": 0.9958789277076722,
+      "num_tokens": 12851463.0,
+      "step": 650
+    },
+    {
+      "entropy": 0.0182177734375,
+      "epoch": 0.14,
+      "grad_norm": 0.013120726682245731,
+      "learning_rate": 9.958309450470784e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9962116169929505,
+      "num_tokens": 13849228.0,
+      "step": 700
+    },
+    {
+      "entropy": 0.0179010009765625,
+      "epoch": 0.15,
+      "grad_norm": 0.014701537787914276,
+      "learning_rate": 9.947225576622847e-05,
+      "loss": 0.0146,
+      "mean_token_accuracy": 0.9961142331361771,
+      "num_tokens": 14812166.0,
+      "step": 750
+    },
+    {
+      "entropy": 0.0173638916015625,
+      "epoch": 0.16,
+      "grad_norm": 0.033504419028759,
+      "learning_rate": 9.934844378609117e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.9961155968904495,
+      "num_tokens": 15789120.0,
+      "step": 800
+    },
+    {
+      "entropy": 0.0183349609375,
+      "epoch": 0.17,
+      "grad_norm": 0.010470729321241379,
+      "learning_rate": 9.921169103184187e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9961740493774414,
+      "num_tokens": 16764645.0,
+      "step": 850
+    },
+    {
+      "entropy": 0.016279296875,
+      "epoch": 0.18,
+      "grad_norm": 0.02781980112195015,
+      "learning_rate": 9.906203336452029e-05,
+      "loss": 0.0142,
+      "mean_token_accuracy": 0.9962105673551559,
+      "num_tokens": 17730003.0,
+      "step": 900
+    },
+    {
+      "entropy": 0.0166461181640625,
+      "epoch": 0.19,
+      "grad_norm": 0.020350394770503044,
+      "learning_rate": 9.889951002925593e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962483876943589,
+      "num_tokens": 18723861.0,
+      "step": 950
+    },
+    {
+      "entropy": 0.0165386962890625,
+      "epoch": 0.2,
+      "grad_norm": 0.013212243095040321,
+      "learning_rate": 9.872416364497675e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9961817175149917,
+      "num_tokens": 19695091.0,
+      "step": 1000
+    },
+    {
+      "entropy": 0.017391357421875,
+      "epoch": 0.21,
+      "grad_norm": 0.014041919261217117,
+      "learning_rate": 9.853604019323301e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962570595741272,
+      "num_tokens": 20662255.0,
+      "step": 1050
+    },
+    {
+      "entropy": 0.016884765625,
+      "epoch": 0.22,
+      "grad_norm": 0.017174789682030678,
+      "learning_rate": 9.833518900613956e-05,
+      "loss": 0.0138,
+      "mean_token_accuracy": 0.9962636953592301,
+      "num_tokens": 21642092.0,
+      "step": 1100
+    },
+    {
+      "entropy": 0.019366455078125,
+      "epoch": 0.23,
+      "grad_norm": 0.017337976023554802,
+      "learning_rate": 9.812166275343917e-05,
+      "loss": 0.0143,
+      "mean_token_accuracy": 0.996159919500351,
+      "num_tokens": 22607505.0,
+      "step": 1150
+    },
+    {
+      "entropy": 0.0164886474609375,
+      "epoch": 0.24,
+      "grad_norm": 0.018486863002181053,
+      "learning_rate": 9.789551742869096e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963399863243103,
+      "num_tokens": 23584579.0,
+      "step": 1200
+    },
+    {
+      "entropy": 0.016282958984375,
+      "epoch": 0.25,
+      "grad_norm": 0.009195365011692047,
+      "learning_rate": 9.765681233458693e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9963367432355881,
+      "num_tokens": 24580285.0,
+      "step": 1250
+    },
+    {
+      "entropy": 0.0197235107421875,
+      "epoch": 0.26,
+      "grad_norm": 0.01205432415008545,
+      "learning_rate": 9.740561006740098e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.996049216389656,
+      "num_tokens": 25525513.0,
+      "step": 1300
+    },
+    {
+      "entropy": 0.0153009033203125,
+      "epoch": 0.27,
+      "grad_norm": 0.01174497976899147,
+      "learning_rate": 9.7141976500574e-05,
+      "loss": 0.0135,
+      "mean_token_accuracy": 0.9964245760440826,
+      "num_tokens": 26523426.0,
+      "step": 1350
+    },
+    {
+      "entropy": 0.0169378662109375,
+      "epoch": 0.28,
+      "grad_norm": 0.014242298901081085,
+      "learning_rate": 9.686598076743981e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962529343366623,
+      "num_tokens": 27493091.0,
+      "step": 1400
+    },
+    {
+      "entropy": 0.017811279296875,
+      "epoch": 0.29,
+      "grad_norm": 0.0160844549536705,
+      "learning_rate": 9.657769524309605e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.996181645989418,
+      "num_tokens": 28458553.0,
+      "step": 1450
+    },
+    {
+      "entropy": 0.0169232177734375,
+      "epoch": 0.3,
+      "grad_norm": 0.03241236135363579,
+      "learning_rate": 9.627719552542516e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9962622857093811,
+      "num_tokens": 29430497.0,
+      "step": 1500
+    },
+    {
+      "entropy": 0.0163720703125,
+      "epoch": 0.31,
+      "grad_norm": 0.01750756986439228,
+      "learning_rate": 9.596456041527001e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9964328509569168,
+      "num_tokens": 30414152.0,
+      "step": 1550
+    },
+    {
+      "entropy": 0.017760009765625,
+      "epoch": 0.32,
+      "grad_norm": 0.02182549238204956,
+      "learning_rate": 9.563987189576991e-05,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.9961263346672058,
+      "num_tokens": 31396417.0,
+      "step": 1600
+    },
+    {
+      "entropy": 0.0171954345703125,
+      "epoch": 0.33,
+      "grad_norm": 0.011301269754767418,
+      "learning_rate": 9.530321511086183e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962880289554596,
+      "num_tokens": 32364443.0,
+      "step": 1650
+    },
+    {
+      "entropy": 0.015126953125,
+      "epoch": 0.34,
+      "grad_norm": 0.009061276912689209,
+      "learning_rate": 9.495467834295291e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9964066076278687,
+      "num_tokens": 33374821.0,
+      "step": 1700
+    },
+    {
+      "entropy": 0.018043212890625,
+      "epoch": 0.35,
+      "grad_norm": 0.014127642847597599,
+      "learning_rate": 9.459435298976998e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963885354995727,
+      "num_tokens": 34338570.0,
+      "step": 1750
+    },
+    {
+      "entropy": 0.0155584716796875,
+      "epoch": 0.36,
+      "grad_norm": 0.00913177989423275,
+      "learning_rate": 9.422233354039198e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965262812376022,
+      "num_tokens": 35333253.0,
+      "step": 1800
+    },
+    {
+      "entropy": 0.0177911376953125,
+      "epoch": 0.37,
+      "grad_norm": 0.009629561565816402,
+      "learning_rate": 9.383871755047198e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9964758467674255,
+      "num_tokens": 36316255.0,
+      "step": 1850
+    },
+    {
+      "entropy": 0.0159539794921875,
+      "epoch": 0.38,
+      "grad_norm": 0.01395715307444334,
+      "learning_rate": 9.344360561665476e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964872688055039,
+      "num_tokens": 37303247.0,
+      "step": 1900
+    },
+    {
+      "entropy": 0.0152783203125,
+      "epoch": 0.39,
+      "grad_norm": 0.007368483114987612,
+      "learning_rate": 9.30371013501972e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965988719463348,
+      "num_tokens": 38287463.0,
+      "step": 1950
+    },
+    {
+      "entropy": 0.0158270263671875,
+      "epoch": 0.4,
+      "grad_norm": 0.01508031040430069,
+      "learning_rate": 9.261931134979791e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9965020543336869,
+      "num_tokens": 39273383.0,
+      "step": 2000
+    },
+    {
+      "entropy": 0.0167919921875,
+      "epoch": 0.41,
+      "grad_norm": 0.009929426945745945,
+      "learning_rate": 9.219034517364369e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965337771177292,
+      "num_tokens": 40251665.0,
+      "step": 2050
+    },
+    {
+      "entropy": 0.0153082275390625,
+      "epoch": 0.42,
+      "grad_norm": 0.007909618318080902,
+      "learning_rate": 9.17503153106797e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965605139732361,
+      "num_tokens": 41243731.0,
+      "step": 2100
+    },
+    {
+      "entropy": 0.0154461669921875,
+      "epoch": 0.43,
+      "grad_norm": 0.015921294689178467,
+      "learning_rate": 9.129933715111125e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9964988273382187,
+      "num_tokens": 42247346.0,
+      "step": 2150
+    },
+    {
+      "entropy": 0.01652587890625,
+      "epoch": 0.44,
+      "grad_norm": 0.006036434322595596,
+      "learning_rate": 9.083752895614464e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964230692386628,
+      "num_tokens": 43220001.0,
+      "step": 2200
+    },
+    {
+      "entropy": 0.015128173828125,
+      "epoch": 0.45,
+      "grad_norm": 0.009778963401913643,
+      "learning_rate": 9.03650118269753e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965469115972518,
+      "num_tokens": 44205197.0,
+      "step": 2250
+    },
+    {
+      "entropy": 0.015570068359375,
+      "epoch": 0.46,
+      "grad_norm": 0.014646470546722412,
+      "learning_rate": 8.988190967303101e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965356832742691,
+      "num_tokens": 45185181.0,
+      "step": 2300
+    },
+    {
+      "entropy": 0.01632568359375,
+      "epoch": 0.47,
+      "grad_norm": 0.006433142349123955,
+      "learning_rate": 8.938834917947889e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965211725234986,
+      "num_tokens": 46170545.0,
+      "step": 2350
+    },
+    {
+      "entropy": 0.01555419921875,
+      "epoch": 0.48,
+      "grad_norm": 0.007329673506319523,
+      "learning_rate": 8.888445977400435e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965089631080627,
+      "num_tokens": 47152165.0,
+      "step": 2400
+    },
+    {
+      "entropy": 0.0156890869140625,
+      "epoch": 0.49,
+      "grad_norm": 0.010034145787358284,
+      "learning_rate": 8.837037359287092e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965383523702621,
+      "num_tokens": 48122023.0,
+      "step": 2450
+    },
+    {
+      "entropy": 0.0163201904296875,
+      "epoch": 0.5,
+      "grad_norm": 0.008338450454175472,
+      "learning_rate": 8.784622544626985e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965311539173126,
+      "num_tokens": 49094218.0,
+      "step": 2500
+    },
+    {
+      "entropy": 0.0152789306640625,
+      "epoch": 0.51,
+      "grad_norm": 0.007771766744554043,
+      "learning_rate": 8.731215278296843e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965461915731431,
+      "num_tokens": 50090248.0,
+      "step": 2550
+    },
+    {
+      "entropy": 0.0171893310546875,
+      "epoch": 0.52,
+      "grad_norm": 0.010534558445215225,
+      "learning_rate": 8.676829565426646e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964153295755387,
+      "num_tokens": 51055771.0,
+      "step": 2600
+    },
+    {
+      "entropy": 0.0148272705078125,
+      "epoch": 0.53,
+      "grad_norm": 0.007072142791002989,
+      "learning_rate": 8.62147966772702e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966214698553085,
+      "num_tokens": 52051029.0,
+      "step": 2650
+    },
+    {
+      "entropy": 0.0146075439453125,
+      "epoch": 0.54,
+      "grad_norm": 0.007414950989186764,
+      "learning_rate": 8.565180099749355e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966325616836548,
+      "num_tokens": 53064519.0,
+      "step": 2700
+    },
+    {
+      "entropy": 0.0161309814453125,
+      "epoch": 0.55,
+      "grad_norm": 0.0070669627748429775,
+      "learning_rate": 8.50794562507961e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9964843678474427,
+      "num_tokens": 54043844.0,
+      "step": 2750
+    },
+    {
+      "entropy": 0.015706787109375,
+      "epoch": 0.56,
+      "grad_norm": 0.005744470749050379,
+      "learning_rate": 8.449791252466819e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965711253881454,
+      "num_tokens": 55026422.0,
+      "step": 2800
+    },
+    {
+      "entropy": 0.01641357421875,
+      "epoch": 0.57,
+      "grad_norm": 0.0069727362133562565,
+      "learning_rate": 8.390732231887314e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.996541822552681,
+      "num_tokens": 56009908.0,
+      "step": 2850
+    },
+    {
+      "entropy": 0.016571044921875,
+      "epoch": 0.58,
+      "grad_norm": 0.007201408036053181,
+      "learning_rate": 8.330784050545672e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964863443374634,
+      "num_tokens": 56948904.0,
+      "step": 2900
+    },
+    {
+      "entropy": 0.0156451416015625,
+      "epoch": 0.59,
+      "grad_norm": 0.005370237864553928,
+      "learning_rate": 8.269962428813474e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966201782226562,
+      "num_tokens": 57925745.0,
+      "step": 2950
+    },
+    {
+      "entropy": 0.01558837890625,
+      "epoch": 0.6,
+      "grad_norm": 0.005006297491490841,
+      "learning_rate": 8.208283316106902e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966268092393875,
+      "num_tokens": 58902798.0,
+      "step": 3000
+    },
+    {
+      "entropy": 0.015733642578125,
+      "epoch": 0.61,
+      "grad_norm": 0.06520090997219086,
+      "learning_rate": 8.145762886704286e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965688252449035,
+      "num_tokens": 59860890.0,
+      "step": 3050
+    },
+    {
+      "entropy": 0.015631103515625,
+      "epoch": 0.62,
+      "grad_norm": 0.017850181087851524,
+      "learning_rate": 8.082417535504683e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9963969469070435,
+      "num_tokens": 60845473.0,
+      "step": 3100
+    },
+    {
+      "entropy": 0.0175927734375,
+      "epoch": 0.63,
+      "grad_norm": 0.017514687031507492,
+      "learning_rate": 8.018263873728585e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9963253819942475,
+      "num_tokens": 61784192.0,
+      "step": 3150
+    },
+    {
+      "entropy": 0.0167742919921875,
+      "epoch": 0.64,
+      "grad_norm": 0.008665528148412704,
+      "learning_rate": 7.953318724561932e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964195656776428,
+      "num_tokens": 62766207.0,
+      "step": 3200
+    },
+    {
+      "entropy": 0.0157501220703125,
+      "epoch": 0.65,
+      "grad_norm": 0.019448991864919662,
+      "learning_rate": 7.887599118744509e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966711962223053,
+      "num_tokens": 63762073.0,
+      "step": 3250
+    },
+    {
+      "entropy": 0.0180609130859375,
+      "epoch": 0.66,
+      "grad_norm": 0.004935233388096094,
+      "learning_rate": 7.821122290103938e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964338165521621,
+      "num_tokens": 64710323.0,
+      "step": 3300
+    },
+    {
+      "entropy": 0.015982666015625,
+      "epoch": 0.67,
+      "grad_norm": 0.004395989701151848,
+      "learning_rate": 7.753905671036403e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966453295946122,
+      "num_tokens": 65703355.0,
+      "step": 3350
+    },
+    {
+      "entropy": 0.0160992431640625,
+      "epoch": 0.68,
+      "grad_norm": 0.01108239684253931,
+      "learning_rate": 7.685966887935309e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996572095155716,
+      "num_tokens": 66693270.0,
+      "step": 3400
+    },
+    {
+      "entropy": 0.0156439208984375,
+      "epoch": 0.69,
+      "grad_norm": 0.007867238484323025,
+      "learning_rate": 7.617323756569053e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965278363227844,
+      "num_tokens": 67662491.0,
+      "step": 3450
+    },
+    {
+      "entropy": 0.0155462646484375,
+      "epoch": 0.7,
+      "grad_norm": 0.006034135352820158,
+      "learning_rate": 7.547994277409168e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966409939527512,
+      "num_tokens": 68643197.0,
+      "step": 3500
+    },
+    {
+      "entropy": 0.015638427734375,
+      "epoch": 0.71,
+      "grad_norm": 0.007907884195446968,
+      "learning_rate": 7.477996630909994e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966798382997513,
+      "num_tokens": 69638359.0,
+      "step": 3550
+    },
+    {
+      "entropy": 0.0161322021484375,
+      "epoch": 0.72,
+      "grad_norm": 0.007897689007222652,
+      "learning_rate": 7.40734917274118e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966677987575531,
+      "num_tokens": 70621331.0,
+      "step": 3600
+    },
+    {
+      "entropy": 0.015335693359375,
+      "epoch": 0.73,
+      "grad_norm": 0.006248582154512405,
+      "learning_rate": 7.336070428974218e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966181468963623,
+      "num_tokens": 71609172.0,
+      "step": 3650
+    },
+    {
+      "entropy": 0.0166754150390625,
+      "epoch": 0.74,
+      "grad_norm": 0.009811230935156345,
+      "learning_rate": 7.26417909122431e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9965909200906754,
+      "num_tokens": 72589978.0,
+      "step": 3700
+    },
+    {
+      "entropy": 0.0143695068359375,
+      "epoch": 0.75,
+      "grad_norm": 0.012172735296189785,
+      "learning_rate": 7.191694011748818e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965182906389236,
+      "num_tokens": 73606622.0,
+      "step": 3750
+    },
+    {
+      "entropy": 0.0156829833984375,
+      "epoch": 0.76,
+      "grad_norm": 0.004879123531281948,
+      "learning_rate": 7.118634198503571e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965479630231857,
+      "num_tokens": 74626982.0,
+      "step": 3800
+    },
+    {
+      "entropy": 0.016346435546875,
+      "epoch": 0.77,
+      "grad_norm": 0.005593061912804842,
+      "learning_rate": 7.045018810158375e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966219502687454,
+      "num_tokens": 75622857.0,
+      "step": 3850
+    },
+    {
+      "entropy": 0.015889892578125,
+      "epoch": 0.78,
+      "grad_norm": 0.006423663813620806,
+      "learning_rate": 6.97086715107298e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966537064313888,
+      "num_tokens": 76601614.0,
+      "step": 3900
+    },
+    {
+      "entropy": 0.015166015625,
+      "epoch": 0.79,
+      "grad_norm": 0.009724115021526814,
+      "learning_rate": 6.896198666234833e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966633880138397,
+      "num_tokens": 77598059.0,
+      "step": 3950
+    },
+    {
+      "entropy": 0.0157177734375,
+      "epoch": 0.8,
+      "grad_norm": 0.014047837816178799,
+      "learning_rate": 6.821032936159986e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996613291501999,
+      "num_tokens": 78570180.0,
+      "step": 4000
+    },
+    {
+      "entropy": 0.0155902099609375,
+      "epoch": 0.81,
+      "grad_norm": 0.008200579322874546,
+      "learning_rate": 6.745389671758435e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9965583562850953,
+      "num_tokens": 79561943.0,
+      "step": 4050
+    },
+    {
+      "entropy": 0.014603271484375,
+      "epoch": 0.82,
+      "grad_norm": 0.0051344577223062515,
+      "learning_rate": 6.669288709165276e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967030560970307,
+      "num_tokens": 80576757.0,
+      "step": 4100
+    },
+    {
+      "entropy": 0.0161895751953125,
+      "epoch": 0.83,
+      "grad_norm": 0.00540529889985919,
+      "learning_rate": 6.59275000453902e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966316163539887,
+      "num_tokens": 81549255.0,
+      "step": 4150
+    },
+    {
+      "entropy": 0.01657470703125,
+      "epoch": 0.84,
+      "grad_norm": 0.005644636228680611,
+      "learning_rate": 6.515793628828447e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965110903978348,
+      "num_tokens": 82499683.0,
+      "step": 4200
+    },
+    {
+      "entropy": 0.0160162353515625,
+      "epoch": 0.85,
+      "grad_norm": 0.004190398845821619,
+      "learning_rate": 6.438439762509332e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966331452131272,
+      "num_tokens": 83475810.0,
+      "step": 4250
+    },
+    {
+      "entropy": 0.01598388671875,
+      "epoch": 0.86,
+      "grad_norm": 0.005400759633630514,
+      "learning_rate": 6.360708690292479e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966162234544754,
+      "num_tokens": 84432102.0,
+      "step": 4300
+    },
+    {
+      "entropy": 0.014876708984375,
+      "epoch": 0.87,
+      "grad_norm": 0.012628648430109024,
+      "learning_rate": 6.2826207958044e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967326593399047,
+      "num_tokens": 85429626.0,
+      "step": 4350
+    },
+    {
+      "entropy": 0.014339599609375,
+      "epoch": 0.88,
+      "grad_norm": 0.009649231098592281,
+      "learning_rate": 6.204196556242061e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967543077468872,
+      "num_tokens": 86450572.0,
+      "step": 4400
+    },
+    {
+      "entropy": 0.014298095703125,
+      "epoch": 0.89,
+      "grad_norm": 0.0049432734958827496,
+      "learning_rate": 6.125456537003095e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967639350891113,
+      "num_tokens": 87458843.0,
+      "step": 4450
+    },
+    {
+      "entropy": 0.015306396484375,
+      "epoch": 0.9,
+      "grad_norm": 0.004098173696547747,
+      "learning_rate": 6.046421386292887e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9966574615240097,
+      "num_tokens": 88453316.0,
+      "step": 4500
+    },
+    {
+      "entropy": 0.016375732421875,
+      "epoch": 0.91,
+      "grad_norm": 0.005302398465573788,
+      "learning_rate": 5.9671118297099274e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966228741407395,
+      "num_tokens": 89427450.0,
+      "step": 4550
+    },
+    {
+      "entropy": 0.015936279296875,
+      "epoch": 0.92,
+      "grad_norm": 0.004837509244680405,
+      "learning_rate": 5.887548664810896e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966443425416946,
+      "num_tokens": 90400676.0,
+      "step": 4600
+    },
+    {
+      "entropy": 0.0147259521484375,
+      "epoch": 0.93,
+      "grad_norm": 0.008831442333757877,
+      "learning_rate": 5.8077527556568614e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967519730329514,
+      "num_tokens": 91394757.0,
+      "step": 4650
+    },
+    {
+      "entropy": 0.01557373046875,
+      "epoch": 0.94,
+      "grad_norm": 0.0063737258315086365,
+      "learning_rate": 5.727745027342031e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966205650568009,
+      "num_tokens": 92386137.0,
+      "step": 4700
+    },
+    {
+      "entropy": 0.0151513671875,
+      "epoch": 0.95,
+      "grad_norm": 0.005380143877118826,
+      "learning_rate": 5.64754646050652e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9965577948093415,
+      "num_tokens": 93374756.0,
+      "step": 4750
+    },
+    {
+      "entropy": 0.0138897705078125,
+      "epoch": 0.96,
+      "grad_norm": 0.005837676115334034,
+      "learning_rate": 5.567178085834542e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967484200000762,
+      "num_tokens": 94389846.0,
+      "step": 4800
+    },
+    {
+      "entropy": 0.0154791259765625,
+      "epoch": 0.97,
+      "grad_norm": 0.008733139373362064,
+      "learning_rate": 5.486660978539468e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966599184274674,
+      "num_tokens": 95378637.0,
+      "step": 4850
+    },
+    {
+      "entropy": 0.015618896484375,
+      "epoch": 0.98,
+      "grad_norm": 0.016466081142425537,
+      "learning_rate": 5.406016252837219e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966346627473831,
+      "num_tokens": 96357363.0,
+      "step": 4900
+    },
+    {
+      "entropy": 0.0171966552734375,
+      "epoch": 0.99,
+      "grad_norm": 0.007259083911776543,
+      "learning_rate": 5.325265056409442e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965799397230148,
+      "num_tokens": 97295349.0,
+      "step": 4950
+    },
+    {
+      "entropy": 0.01597900390625,
+      "epoch": 1.0,
+      "grad_norm": 0.007410865277051926,
+      "learning_rate": 5.244428564857874e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.996621105670929,
+      "num_tokens": 98262502.0,
+      "step": 5000
+    },
+    {
+      "entropy": 0.0154595947265625,
+      "epoch": 1.01,
+      "grad_norm": 0.0048329527489840984,
+      "learning_rate": 5.163527976151428e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966327011585235,
+      "num_tokens": 99235771.0,
+      "step": 5050
+    },
+    {
+      "entropy": 0.0155859375,
+      "epoch": 1.02,
+      "grad_norm": 0.004490161780267954,
+      "learning_rate": 5.0825845050673834e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965754437446595,
+      "num_tokens": 100195990.0,
+      "step": 5100
+    },
+    {
+      "entropy": 0.01652099609375,
+      "epoch": 1.03,
+      "grad_norm": 0.00914011336863041,
+      "learning_rate": 5.0016193776281794e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9964404582977295,
+      "num_tokens": 101140160.0,
+      "step": 5150
+    },
+    {
+      "entropy": 0.015284423828125,
+      "epoch": 1.04,
+      "grad_norm": 0.0057498314417898655,
+      "learning_rate": 4.920653825535258e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.99665458381176,
+      "num_tokens": 102128346.0,
+      "step": 5200
+    },
+    {
+      "entropy": 0.015015869140625,
+      "epoch": 1.05,
+      "grad_norm": 0.004660275299102068,
+      "learning_rate": 4.839709080601419e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967097079753876,
+      "num_tokens": 103125113.0,
+      "step": 5250
+    },
+    {
+      "entropy": 0.0143218994140625,
+      "epoch": 1.06,
+      "grad_norm": 0.006082482635974884,
+      "learning_rate": 4.758806369183149e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967189061641694,
+      "num_tokens": 104136988.0,
+      "step": 5300
+    },
+    {
+      "entropy": 0.0155865478515625,
+      "epoch": 1.07,
+      "grad_norm": 0.00519182626157999,
+      "learning_rate": 4.6779669066143716e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966564351320266,
+      "num_tokens": 105122070.0,
+      "step": 5350
+    },
+    {
+      "entropy": 0.0145452880859375,
+      "epoch": 1.08,
+      "grad_norm": 0.005050841718912125,
+      "learning_rate": 4.597211891643093e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967609119415283,
+      "num_tokens": 106117109.0,
+      "step": 5400
+    },
+    {
+      "entropy": 0.0144915771484375,
+      "epoch": 1.09,
+      "grad_norm": 0.004848674405366182,
+      "learning_rate": 4.5165625008724035e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967008078098297,
+      "num_tokens": 107110319.0,
+      "step": 5450
+    },
+    {
+      "entropy": 0.0150311279296875,
+      "epoch": 1.1,
+      "grad_norm": 0.0036910909693688154,
+      "learning_rate": 4.4360398832072764e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967303448915481,
+      "num_tokens": 108104820.0,
+      "step": 5500
+    },
+    {
+      "entropy": 0.015078125,
+      "epoch": 1.11,
+      "grad_norm": 0.0053949966095387936,
+      "learning_rate": 4.3556651543086364e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966647911071778,
+      "num_tokens": 109091493.0,
+      "step": 5550
+    },
+    {
+      "entropy": 0.0147705078125,
+      "epoch": 1.12,
+      "grad_norm": 0.0037009185180068016,
+      "learning_rate": 4.275459391056142e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966964650154114,
+      "num_tokens": 110077375.0,
+      "step": 5600
+    },
+    {
+      "entropy": 0.0162713623046875,
+      "epoch": 1.13,
+      "grad_norm": 0.0031313870567828417,
+      "learning_rate": 4.195443626021139e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966827458143235,
+      "num_tokens": 111042295.0,
+      "step": 5650
+    },
+    {
+      "entropy": 0.0145147705078125,
+      "epoch": 1.1400000000000001,
+      "grad_norm": 0.004629215691238642,
+      "learning_rate": 4.1156388419512324e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966849237680435,
+      "num_tokens": 112033836.0,
+      "step": 5700
+    },
+    {
+      "entropy": 0.0152557373046875,
+      "epoch": 1.15,
+      "grad_norm": 0.00527191162109375,
+      "learning_rate": 4.0360659662679265e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.996674497127533,
+      "num_tokens": 113006848.0,
+      "step": 5750
+    },
+    {
+      "entropy": 0.0149151611328125,
+      "epoch": 1.16,
+      "grad_norm": 0.004986864514648914,
+      "learning_rate": 3.956745865578773e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966980391740798,
+      "num_tokens": 113998253.0,
+      "step": 5800
+    },
+    {
+      "entropy": 0.015072021484375,
+      "epoch": 1.17,
+      "grad_norm": 0.004452258348464966,
+      "learning_rate": 3.877699340205455e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966688090562821,
+      "num_tokens": 114979026.0,
+      "step": 5850
+    },
+    {
+      "entropy": 0.0157171630859375,
+      "epoch": 1.18,
+      "grad_norm": 0.0048246984370052814,
+      "learning_rate": 3.798947118729282e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966557210683823,
+      "num_tokens": 115947279.0,
+      "step": 5900
+    },
+    {
+      "entropy": 0.015772705078125,
+      "epoch": 1.19,
+      "grad_norm": 0.003676283173263073,
+      "learning_rate": 3.720509852555456e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966129893064499,
+      "num_tokens": 116911661.0,
+      "step": 5950
+    },
+    {
+      "entropy": 0.01443603515625,
+      "epoch": 1.2,
+      "grad_norm": 0.004810075741261244,
+      "learning_rate": 3.642408110497616e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966686731576919,
+      "num_tokens": 117890306.0,
+      "step": 6000
+    },
+    {
+      "entropy": 0.0159271240234375,
+      "epoch": 1.21,
+      "grad_norm": 0.003690144745633006,
+      "learning_rate": 3.5646623733840134e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966743797063827,
+      "num_tokens": 118844495.0,
+      "step": 6050
+    },
+    {
+      "entropy": 0.0148345947265625,
+      "epoch": 1.22,
+      "grad_norm": 0.0036596781574189663,
+      "learning_rate": 3.4872930286867763e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966356498003006,
+      "num_tokens": 119807321.0,
+      "step": 6100
+    },
+    {
+      "entropy": 0.0147174072265625,
+      "epoch": 1.23,
+      "grad_norm": 0.0046323081478476524,
+      "learning_rate": 3.4103203651756403e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967032814025879,
+      "num_tokens": 120794059.0,
+      "step": 6150
+    },
+    {
+      "entropy": 0.0155657958984375,
+      "epoch": 1.24,
+      "grad_norm": 0.005745012313127518,
+      "learning_rate": 3.333764567597579e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.996661651134491,
+      "num_tokens": 121765426.0,
+      "step": 6200
+    },
+    {
+      "entropy": 0.0151751708984375,
+      "epoch": 1.25,
+      "grad_norm": 0.00422937236726284,
+      "learning_rate": 3.257645711383699e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967177057266235,
+      "num_tokens": 122747186.0,
+      "step": 6250
+    },
+    {
+      "entropy": 0.014896240234375,
+      "epoch": 1.26,
+      "grad_norm": 0.0041396524757146835,
+      "learning_rate": 3.1819837573848055e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967236334085464,
+      "num_tokens": 123735045.0,
+      "step": 6300
+    },
+    {
+      "entropy": 0.014747314453125,
+      "epoch": 1.27,
+      "grad_norm": 0.0034245741553604603,
+      "learning_rate": 3.106798546637019e-05,
+      "loss": 0.012,
+      "mean_token_accuracy": 0.9968506091833115,
+      "num_tokens": 124748439.0,
+      "step": 6350
+    },
+    {
+      "entropy": 0.014637451171875,
+      "epoch": 1.28,
+      "grad_norm": 0.0044257547706365585,
+      "learning_rate": 3.0321097951588016e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.996695158481598,
+      "num_tokens": 125723365.0,
+      "step": 6400
+    },
+    {
+      "entropy": 0.0153790283203125,
+      "epoch": 1.29,
+      "grad_norm": 0.005239363294094801,
+      "learning_rate": 2.957937088780779e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966009676456451,
+      "num_tokens": 126678062.0,
+      "step": 6450
+    },
+    {
+      "entropy": 0.0145068359375,
+      "epoch": 1.3,
+      "grad_norm": 0.005349221173673868,
+      "learning_rate": 2.8842998780096896e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967689532041549,
+      "num_tokens": 127675372.0,
+      "step": 6500
+    },
+    {
+      "entropy": 0.0149395751953125,
+      "epoch": 1.31,
+      "grad_norm": 0.005388084799051285,
+      "learning_rate": 2.811217472927835e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967477923631668,
+      "num_tokens": 128665669.0,
+      "step": 6550
+    },
+    {
+      "entropy": 0.0159820556640625,
+      "epoch": 1.32,
+      "grad_norm": 0.006312755402177572,
+      "learning_rate": 2.7387090381293372e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966998076438904,
+      "num_tokens": 129622542.0,
+      "step": 6600
+    },
+    {
+      "entropy": 0.016163330078125,
+      "epoch": 1.33,
+      "grad_norm": 0.005128033459186554,
+      "learning_rate": 2.6667935876945582e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966605240106583,
+      "num_tokens": 130598823.0,
+      "step": 6650
+    },
+    {
+      "entropy": 0.017476806640625,
+      "epoch": 1.34,
+      "grad_norm": 0.005179600324481726,
+      "learning_rate": 2.59548998020399e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.996593764424324,
+      "num_tokens": 131552204.0,
+      "step": 6700
+    },
+    {
+      "entropy": 0.0151275634765625,
+      "epoch": 1.35,
+      "grad_norm": 0.005807869601994753,
+      "learning_rate": 2.5248169137929156e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966244864463806,
+      "num_tokens": 132526021.0,
+      "step": 6750
+    },
+    {
+      "entropy": 0.015770263671875,
+      "epoch": 1.3599999999999999,
+      "grad_norm": 0.005202575121074915,
+      "learning_rate": 2.4547929212481435e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966981512308121,
+      "num_tokens": 133483207.0,
+      "step": 6800
+    },
+    {
+      "entropy": 0.014925537109375,
+      "epoch": 1.37,
+      "grad_norm": 0.003400217741727829,
+      "learning_rate": 2.3854363651481194e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967307335138321,
+      "num_tokens": 134482802.0,
+      "step": 6850
+    },
+    {
+      "entropy": 0.0151629638671875,
+      "epoch": 1.38,
+      "grad_norm": 0.004814519081264734,
+      "learning_rate": 2.3167654330476412e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.996745794415474,
+      "num_tokens": 135484861.0,
+      "step": 6900
+    },
+    {
+      "entropy": 0.0138885498046875,
+      "epoch": 1.3900000000000001,
+      "grad_norm": 0.0049650054425001144,
+      "learning_rate": 2.2487981327084995e-05,
+      "loss": 0.012,
+      "mean_token_accuracy": 0.9967457884550095,
+      "num_tokens": 136499177.0,
+      "step": 6950
+    },
+    {
+      "entropy": 0.014207763671875,
+      "epoch": 1.4,
+      "grad_norm": 0.0068522971123456955,
+      "learning_rate": 2.1815522873772475e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967319291830062,
+      "num_tokens": 137503523.0,
+      "step": 7000
+    }
+  ],
+  "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": 2.388030314757751e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-7000/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_lr1e-4/checkpoint-7500/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": [
+    "up_proj",
+    "q_proj",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "v_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_lr1e-4/checkpoint-7500/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_lr1e-4/checkpoint-7500/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_lr1e-4/checkpoint-7500/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_lr1e-4/checkpoint-7500/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_lr1e-4/checkpoint-7500/trainer_state.json

@@ -0,0 +1,1534 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.5,
+  "eval_steps": 500,
+  "global_step": 7500,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "entropy": 0.166884765625,
+      "epoch": 0.01,
+      "grad_norm": 0.09714163839817047,
+      "learning_rate": 1.6333333333333335e-05,
+      "loss": 0.1512,
+      "mean_token_accuracy": 0.966504562497139,
+      "num_tokens": 973322.0,
+      "step": 50
+    },
+    {
+      "entropy": 0.048140869140625,
+      "epoch": 0.02,
+      "grad_norm": 0.057216521352529526,
+      "learning_rate": 3.3e-05,
+      "loss": 0.0424,
+      "mean_token_accuracy": 0.9896974349021912,
+      "num_tokens": 1965908.0,
+      "step": 100
+    },
+    {
+      "entropy": 0.029632568359375,
+      "epoch": 0.03,
+      "grad_norm": 0.054412033408880234,
+      "learning_rate": 4.966666666666667e-05,
+      "loss": 0.0248,
+      "mean_token_accuracy": 0.9936319047212601,
+      "num_tokens": 2931943.0,
+      "step": 150
+    },
+    {
+      "entropy": 0.02478515625,
+      "epoch": 0.04,
+      "grad_norm": 0.08800562471151352,
+      "learning_rate": 6.633333333333334e-05,
+      "loss": 0.0214,
+      "mean_token_accuracy": 0.9943305045366287,
+      "num_tokens": 3906979.0,
+      "step": 200
+    },
+    {
+      "entropy": 0.0224652099609375,
+      "epoch": 0.05,
+      "grad_norm": 0.041793134063482285,
+      "learning_rate": 8.3e-05,
+      "loss": 0.0191,
+      "mean_token_accuracy": 0.9948792272806167,
+      "num_tokens": 4891791.0,
+      "step": 250
+    },
+    {
+      "entropy": 0.01969482421875,
+      "epoch": 0.06,
+      "grad_norm": 0.04181526601314545,
+      "learning_rate": 9.966666666666667e-05,
+      "loss": 0.0173,
+      "mean_token_accuracy": 0.995435175895691,
+      "num_tokens": 5923811.0,
+      "step": 300
+    },
+    {
+      "entropy": 0.019368896484375,
+      "epoch": 0.07,
+      "grad_norm": 0.03995412588119507,
+      "learning_rate": 9.999370378817499e-05,
+      "loss": 0.0164,
+      "mean_token_accuracy": 0.9956372672319412,
+      "num_tokens": 6923169.0,
+      "step": 350
+    },
+    {
+      "entropy": 0.0204400634765625,
+      "epoch": 0.08,
+      "grad_norm": 0.06872096657752991,
+      "learning_rate": 9.997430021636957e-05,
+      "loss": 0.0163,
+      "mean_token_accuracy": 0.9956023478507996,
+      "num_tokens": 7901360.0,
+      "step": 400
+    },
+    {
+      "entropy": 0.0186175537109375,
+      "epoch": 0.09,
+      "grad_norm": 0.037068989127874374,
+      "learning_rate": 9.994179175045023e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958296239376068,
+      "num_tokens": 8899074.0,
+      "step": 450
+    },
+    {
+      "entropy": 0.0173492431640625,
+      "epoch": 0.1,
+      "grad_norm": 0.014651700854301453,
+      "learning_rate": 9.989618691519873e-05,
+      "loss": 0.015,
+      "mean_token_accuracy": 0.9959587389230729,
+      "num_tokens": 9909340.0,
+      "step": 500
+    },
+    {
+      "entropy": 0.0183203125,
+      "epoch": 0.11,
+      "grad_norm": 0.016151444986462593,
+      "learning_rate": 9.983749766969271e-05,
+      "loss": 0.0151,
+      "mean_token_accuracy": 0.9959770923852921,
+      "num_tokens": 10894801.0,
+      "step": 550
+    },
+    {
+      "entropy": 0.02047607421875,
+      "epoch": 0.12,
+      "grad_norm": 0.021302781999111176,
+      "learning_rate": 9.976573940416966e-05,
+      "loss": 0.0157,
+      "mean_token_accuracy": 0.9958055526018142,
+      "num_tokens": 11862148.0,
+      "step": 600
+    },
+    {
+      "entropy": 0.017862548828125,
+      "epoch": 0.13,
+      "grad_norm": 0.013205700553953648,
+      "learning_rate": 9.968093093599106e-05,
+      "loss": 0.0152,
+      "mean_token_accuracy": 0.9958789277076722,
+      "num_tokens": 12851463.0,
+      "step": 650
+    },
+    {
+      "entropy": 0.0182177734375,
+      "epoch": 0.14,
+      "grad_norm": 0.013120726682245731,
+      "learning_rate": 9.958309450470784e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9962116169929505,
+      "num_tokens": 13849228.0,
+      "step": 700
+    },
+    {
+      "entropy": 0.0179010009765625,
+      "epoch": 0.15,
+      "grad_norm": 0.014701537787914276,
+      "learning_rate": 9.947225576622847e-05,
+      "loss": 0.0146,
+      "mean_token_accuracy": 0.9961142331361771,
+      "num_tokens": 14812166.0,
+      "step": 750
+    },
+    {
+      "entropy": 0.0173638916015625,
+      "epoch": 0.16,
+      "grad_norm": 0.033504419028759,
+      "learning_rate": 9.934844378609117e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.9961155968904495,
+      "num_tokens": 15789120.0,
+      "step": 800
+    },
+    {
+      "entropy": 0.0183349609375,
+      "epoch": 0.17,
+      "grad_norm": 0.010470729321241379,
+      "learning_rate": 9.921169103184187e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.9961740493774414,
+      "num_tokens": 16764645.0,
+      "step": 850
+    },
+    {
+      "entropy": 0.016279296875,
+      "epoch": 0.18,
+      "grad_norm": 0.02781980112195015,
+      "learning_rate": 9.906203336452029e-05,
+      "loss": 0.0142,
+      "mean_token_accuracy": 0.9962105673551559,
+      "num_tokens": 17730003.0,
+      "step": 900
+    },
+    {
+      "entropy": 0.0166461181640625,
+      "epoch": 0.19,
+      "grad_norm": 0.020350394770503044,
+      "learning_rate": 9.889951002925593e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962483876943589,
+      "num_tokens": 18723861.0,
+      "step": 950
+    },
+    {
+      "entropy": 0.0165386962890625,
+      "epoch": 0.2,
+      "grad_norm": 0.013212243095040321,
+      "learning_rate": 9.872416364497675e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9961817175149917,
+      "num_tokens": 19695091.0,
+      "step": 1000
+    },
+    {
+      "entropy": 0.017391357421875,
+      "epoch": 0.21,
+      "grad_norm": 0.014041919261217117,
+      "learning_rate": 9.853604019323301e-05,
+      "loss": 0.0141,
+      "mean_token_accuracy": 0.9962570595741272,
+      "num_tokens": 20662255.0,
+      "step": 1050
+    },
+    {
+      "entropy": 0.016884765625,
+      "epoch": 0.22,
+      "grad_norm": 0.017174789682030678,
+      "learning_rate": 9.833518900613956e-05,
+      "loss": 0.0138,
+      "mean_token_accuracy": 0.9962636953592301,
+      "num_tokens": 21642092.0,
+      "step": 1100
+    },
+    {
+      "entropy": 0.019366455078125,
+      "epoch": 0.23,
+      "grad_norm": 0.017337976023554802,
+      "learning_rate": 9.812166275343917e-05,
+      "loss": 0.0143,
+      "mean_token_accuracy": 0.996159919500351,
+      "num_tokens": 22607505.0,
+      "step": 1150
+    },
+    {
+      "entropy": 0.0164886474609375,
+      "epoch": 0.24,
+      "grad_norm": 0.018486863002181053,
+      "learning_rate": 9.789551742869096e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963399863243103,
+      "num_tokens": 23584579.0,
+      "step": 1200
+    },
+    {
+      "entropy": 0.016282958984375,
+      "epoch": 0.25,
+      "grad_norm": 0.009195365011692047,
+      "learning_rate": 9.765681233458693e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9963367432355881,
+      "num_tokens": 24580285.0,
+      "step": 1250
+    },
+    {
+      "entropy": 0.0197235107421875,
+      "epoch": 0.26,
+      "grad_norm": 0.01205432415008545,
+      "learning_rate": 9.740561006740098e-05,
+      "loss": 0.0144,
+      "mean_token_accuracy": 0.996049216389656,
+      "num_tokens": 25525513.0,
+      "step": 1300
+    },
+    {
+      "entropy": 0.0153009033203125,
+      "epoch": 0.27,
+      "grad_norm": 0.01174497976899147,
+      "learning_rate": 9.7141976500574e-05,
+      "loss": 0.0135,
+      "mean_token_accuracy": 0.9964245760440826,
+      "num_tokens": 26523426.0,
+      "step": 1350
+    },
+    {
+      "entropy": 0.0169378662109375,
+      "epoch": 0.28,
+      "grad_norm": 0.014242298901081085,
+      "learning_rate": 9.686598076743981e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962529343366623,
+      "num_tokens": 27493091.0,
+      "step": 1400
+    },
+    {
+      "entropy": 0.017811279296875,
+      "epoch": 0.29,
+      "grad_norm": 0.0160844549536705,
+      "learning_rate": 9.657769524309605e-05,
+      "loss": 0.0145,
+      "mean_token_accuracy": 0.996181645989418,
+      "num_tokens": 28458553.0,
+      "step": 1450
+    },
+    {
+      "entropy": 0.0169232177734375,
+      "epoch": 0.3,
+      "grad_norm": 0.03241236135363579,
+      "learning_rate": 9.627719552542516e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9962622857093811,
+      "num_tokens": 29430497.0,
+      "step": 1500
+    },
+    {
+      "entropy": 0.0163720703125,
+      "epoch": 0.31,
+      "grad_norm": 0.01750756986439228,
+      "learning_rate": 9.596456041527001e-05,
+      "loss": 0.0136,
+      "mean_token_accuracy": 0.9964328509569168,
+      "num_tokens": 30414152.0,
+      "step": 1550
+    },
+    {
+      "entropy": 0.017760009765625,
+      "epoch": 0.32,
+      "grad_norm": 0.02182549238204956,
+      "learning_rate": 9.563987189576991e-05,
+      "loss": 0.0147,
+      "mean_token_accuracy": 0.9961263346672058,
+      "num_tokens": 31396417.0,
+      "step": 1600
+    },
+    {
+      "entropy": 0.0171954345703125,
+      "epoch": 0.33,
+      "grad_norm": 0.011301269754767418,
+      "learning_rate": 9.530321511086183e-05,
+      "loss": 0.0139,
+      "mean_token_accuracy": 0.9962880289554596,
+      "num_tokens": 32364443.0,
+      "step": 1650
+    },
+    {
+      "entropy": 0.015126953125,
+      "epoch": 0.34,
+      "grad_norm": 0.009061276912689209,
+      "learning_rate": 9.495467834295291e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9964066076278687,
+      "num_tokens": 33374821.0,
+      "step": 1700
+    },
+    {
+      "entropy": 0.018043212890625,
+      "epoch": 0.35,
+      "grad_norm": 0.014127642847597599,
+      "learning_rate": 9.459435298976998e-05,
+      "loss": 0.0137,
+      "mean_token_accuracy": 0.9963885354995727,
+      "num_tokens": 34338570.0,
+      "step": 1750
+    },
+    {
+      "entropy": 0.0155584716796875,
+      "epoch": 0.36,
+      "grad_norm": 0.00913177989423275,
+      "learning_rate": 9.422233354039198e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965262812376022,
+      "num_tokens": 35333253.0,
+      "step": 1800
+    },
+    {
+      "entropy": 0.0177911376953125,
+      "epoch": 0.37,
+      "grad_norm": 0.009629561565816402,
+      "learning_rate": 9.383871755047198e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9964758467674255,
+      "num_tokens": 36316255.0,
+      "step": 1850
+    },
+    {
+      "entropy": 0.0159539794921875,
+      "epoch": 0.38,
+      "grad_norm": 0.01395715307444334,
+      "learning_rate": 9.344360561665476e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964872688055039,
+      "num_tokens": 37303247.0,
+      "step": 1900
+    },
+    {
+      "entropy": 0.0152783203125,
+      "epoch": 0.39,
+      "grad_norm": 0.007368483114987612,
+      "learning_rate": 9.30371013501972e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965988719463348,
+      "num_tokens": 38287463.0,
+      "step": 1950
+    },
+    {
+      "entropy": 0.0158270263671875,
+      "epoch": 0.4,
+      "grad_norm": 0.01508031040430069,
+      "learning_rate": 9.261931134979791e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9965020543336869,
+      "num_tokens": 39273383.0,
+      "step": 2000
+    },
+    {
+      "entropy": 0.0167919921875,
+      "epoch": 0.41,
+      "grad_norm": 0.009929426945745945,
+      "learning_rate": 9.219034517364369e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965337771177292,
+      "num_tokens": 40251665.0,
+      "step": 2050
+    },
+    {
+      "entropy": 0.0153082275390625,
+      "epoch": 0.42,
+      "grad_norm": 0.007909618318080902,
+      "learning_rate": 9.17503153106797e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965605139732361,
+      "num_tokens": 41243731.0,
+      "step": 2100
+    },
+    {
+      "entropy": 0.0154461669921875,
+      "epoch": 0.43,
+      "grad_norm": 0.015921294689178467,
+      "learning_rate": 9.129933715111125e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9964988273382187,
+      "num_tokens": 42247346.0,
+      "step": 2150
+    },
+    {
+      "entropy": 0.01652587890625,
+      "epoch": 0.44,
+      "grad_norm": 0.006036434322595596,
+      "learning_rate": 9.083752895614464e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964230692386628,
+      "num_tokens": 43220001.0,
+      "step": 2200
+    },
+    {
+      "entropy": 0.015128173828125,
+      "epoch": 0.45,
+      "grad_norm": 0.009778963401913643,
+      "learning_rate": 9.03650118269753e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965469115972518,
+      "num_tokens": 44205197.0,
+      "step": 2250
+    },
+    {
+      "entropy": 0.015570068359375,
+      "epoch": 0.46,
+      "grad_norm": 0.014646470546722412,
+      "learning_rate": 8.988190967303101e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965356832742691,
+      "num_tokens": 45185181.0,
+      "step": 2300
+    },
+    {
+      "entropy": 0.01632568359375,
+      "epoch": 0.47,
+      "grad_norm": 0.006433142349123955,
+      "learning_rate": 8.938834917947889e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9965211725234986,
+      "num_tokens": 46170545.0,
+      "step": 2350
+    },
+    {
+      "entropy": 0.01555419921875,
+      "epoch": 0.48,
+      "grad_norm": 0.007329673506319523,
+      "learning_rate": 8.888445977400435e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965089631080627,
+      "num_tokens": 47152165.0,
+      "step": 2400
+    },
+    {
+      "entropy": 0.0156890869140625,
+      "epoch": 0.49,
+      "grad_norm": 0.010034145787358284,
+      "learning_rate": 8.837037359287092e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965383523702621,
+      "num_tokens": 48122023.0,
+      "step": 2450
+    },
+    {
+      "entropy": 0.0163201904296875,
+      "epoch": 0.5,
+      "grad_norm": 0.008338450454175472,
+      "learning_rate": 8.784622544626985e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965311539173126,
+      "num_tokens": 49094218.0,
+      "step": 2500
+    },
+    {
+      "entropy": 0.0152789306640625,
+      "epoch": 0.51,
+      "grad_norm": 0.007771766744554043,
+      "learning_rate": 8.731215278296843e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9965461915731431,
+      "num_tokens": 50090248.0,
+      "step": 2550
+    },
+    {
+      "entropy": 0.0171893310546875,
+      "epoch": 0.52,
+      "grad_norm": 0.010534558445215225,
+      "learning_rate": 8.676829565426646e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964153295755387,
+      "num_tokens": 51055771.0,
+      "step": 2600
+    },
+    {
+      "entropy": 0.0148272705078125,
+      "epoch": 0.53,
+      "grad_norm": 0.007072142791002989,
+      "learning_rate": 8.62147966772702e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966214698553085,
+      "num_tokens": 52051029.0,
+      "step": 2650
+    },
+    {
+      "entropy": 0.0146075439453125,
+      "epoch": 0.54,
+      "grad_norm": 0.007414950989186764,
+      "learning_rate": 8.565180099749355e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966325616836548,
+      "num_tokens": 53064519.0,
+      "step": 2700
+    },
+    {
+      "entropy": 0.0161309814453125,
+      "epoch": 0.55,
+      "grad_norm": 0.0070669627748429775,
+      "learning_rate": 8.50794562507961e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9964843678474427,
+      "num_tokens": 54043844.0,
+      "step": 2750
+    },
+    {
+      "entropy": 0.015706787109375,
+      "epoch": 0.56,
+      "grad_norm": 0.005744470749050379,
+      "learning_rate": 8.449791252466819e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965711253881454,
+      "num_tokens": 55026422.0,
+      "step": 2800
+    },
+    {
+      "entropy": 0.01641357421875,
+      "epoch": 0.57,
+      "grad_norm": 0.0069727362133562565,
+      "learning_rate": 8.390732231887314e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.996541822552681,
+      "num_tokens": 56009908.0,
+      "step": 2850
+    },
+    {
+      "entropy": 0.016571044921875,
+      "epoch": 0.58,
+      "grad_norm": 0.007201408036053181,
+      "learning_rate": 8.330784050545672e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964863443374634,
+      "num_tokens": 56948904.0,
+      "step": 2900
+    },
+    {
+      "entropy": 0.0156451416015625,
+      "epoch": 0.59,
+      "grad_norm": 0.005370237864553928,
+      "learning_rate": 8.269962428813474e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966201782226562,
+      "num_tokens": 57925745.0,
+      "step": 2950
+    },
+    {
+      "entropy": 0.01558837890625,
+      "epoch": 0.6,
+      "grad_norm": 0.005006297491490841,
+      "learning_rate": 8.208283316106902e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966268092393875,
+      "num_tokens": 58902798.0,
+      "step": 3000
+    },
+    {
+      "entropy": 0.015733642578125,
+      "epoch": 0.61,
+      "grad_norm": 0.06520090997219086,
+      "learning_rate": 8.145762886704286e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965688252449035,
+      "num_tokens": 59860890.0,
+      "step": 3050
+    },
+    {
+      "entropy": 0.015631103515625,
+      "epoch": 0.62,
+      "grad_norm": 0.017850181087851524,
+      "learning_rate": 8.082417535504683e-05,
+      "loss": 0.0134,
+      "mean_token_accuracy": 0.9963969469070435,
+      "num_tokens": 60845473.0,
+      "step": 3100
+    },
+    {
+      "entropy": 0.0175927734375,
+      "epoch": 0.63,
+      "grad_norm": 0.017514687031507492,
+      "learning_rate": 8.018263873728585e-05,
+      "loss": 0.014,
+      "mean_token_accuracy": 0.9963253819942475,
+      "num_tokens": 61784192.0,
+      "step": 3150
+    },
+    {
+      "entropy": 0.0167742919921875,
+      "epoch": 0.64,
+      "grad_norm": 0.008665528148412704,
+      "learning_rate": 7.953318724561932e-05,
+      "loss": 0.0132,
+      "mean_token_accuracy": 0.9964195656776428,
+      "num_tokens": 62766207.0,
+      "step": 3200
+    },
+    {
+      "entropy": 0.0157501220703125,
+      "epoch": 0.65,
+      "grad_norm": 0.019448991864919662,
+      "learning_rate": 7.887599118744509e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966711962223053,
+      "num_tokens": 63762073.0,
+      "step": 3250
+    },
+    {
+      "entropy": 0.0180609130859375,
+      "epoch": 0.66,
+      "grad_norm": 0.004935233388096094,
+      "learning_rate": 7.821122290103938e-05,
+      "loss": 0.0133,
+      "mean_token_accuracy": 0.9964338165521621,
+      "num_tokens": 64710323.0,
+      "step": 3300
+    },
+    {
+      "entropy": 0.015982666015625,
+      "epoch": 0.67,
+      "grad_norm": 0.004395989701151848,
+      "learning_rate": 7.753905671036403e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966453295946122,
+      "num_tokens": 65703355.0,
+      "step": 3350
+    },
+    {
+      "entropy": 0.0160992431640625,
+      "epoch": 0.68,
+      "grad_norm": 0.01108239684253931,
+      "learning_rate": 7.685966887935309e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996572095155716,
+      "num_tokens": 66693270.0,
+      "step": 3400
+    },
+    {
+      "entropy": 0.0156439208984375,
+      "epoch": 0.69,
+      "grad_norm": 0.007867238484323025,
+      "learning_rate": 7.617323756569053e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9965278363227844,
+      "num_tokens": 67662491.0,
+      "step": 3450
+    },
+    {
+      "entropy": 0.0155462646484375,
+      "epoch": 0.7,
+      "grad_norm": 0.006034135352820158,
+      "learning_rate": 7.547994277409168e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966409939527512,
+      "num_tokens": 68643197.0,
+      "step": 3500
+    },
+    {
+      "entropy": 0.015638427734375,
+      "epoch": 0.71,
+      "grad_norm": 0.007907884195446968,
+      "learning_rate": 7.477996630909994e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966798382997513,
+      "num_tokens": 69638359.0,
+      "step": 3550
+    },
+    {
+      "entropy": 0.0161322021484375,
+      "epoch": 0.72,
+      "grad_norm": 0.007897689007222652,
+      "learning_rate": 7.40734917274118e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966677987575531,
+      "num_tokens": 70621331.0,
+      "step": 3600
+    },
+    {
+      "entropy": 0.015335693359375,
+      "epoch": 0.73,
+      "grad_norm": 0.006248582154512405,
+      "learning_rate": 7.336070428974218e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966181468963623,
+      "num_tokens": 71609172.0,
+      "step": 3650
+    },
+    {
+      "entropy": 0.0166754150390625,
+      "epoch": 0.74,
+      "grad_norm": 0.009811230935156345,
+      "learning_rate": 7.26417909122431e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9965909200906754,
+      "num_tokens": 72589978.0,
+      "step": 3700
+    },
+    {
+      "entropy": 0.0143695068359375,
+      "epoch": 0.75,
+      "grad_norm": 0.012172735296189785,
+      "learning_rate": 7.191694011748818e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965182906389236,
+      "num_tokens": 73606622.0,
+      "step": 3750
+    },
+    {
+      "entropy": 0.0156829833984375,
+      "epoch": 0.76,
+      "grad_norm": 0.004879123531281948,
+      "learning_rate": 7.118634198503571e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965479630231857,
+      "num_tokens": 74626982.0,
+      "step": 3800
+    },
+    {
+      "entropy": 0.016346435546875,
+      "epoch": 0.77,
+      "grad_norm": 0.005593061912804842,
+      "learning_rate": 7.045018810158375e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966219502687454,
+      "num_tokens": 75622857.0,
+      "step": 3850
+    },
+    {
+      "entropy": 0.015889892578125,
+      "epoch": 0.78,
+      "grad_norm": 0.006423663813620806,
+      "learning_rate": 6.97086715107298e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.9966537064313888,
+      "num_tokens": 76601614.0,
+      "step": 3900
+    },
+    {
+      "entropy": 0.015166015625,
+      "epoch": 0.79,
+      "grad_norm": 0.009724115021526814,
+      "learning_rate": 6.896198666234833e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966633880138397,
+      "num_tokens": 77598059.0,
+      "step": 3950
+    },
+    {
+      "entropy": 0.0157177734375,
+      "epoch": 0.8,
+      "grad_norm": 0.014047837816178799,
+      "learning_rate": 6.821032936159986e-05,
+      "loss": 0.0128,
+      "mean_token_accuracy": 0.996613291501999,
+      "num_tokens": 78570180.0,
+      "step": 4000
+    },
+    {
+      "entropy": 0.0155902099609375,
+      "epoch": 0.81,
+      "grad_norm": 0.008200579322874546,
+      "learning_rate": 6.745389671758435e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9965583562850953,
+      "num_tokens": 79561943.0,
+      "step": 4050
+    },
+    {
+      "entropy": 0.014603271484375,
+      "epoch": 0.82,
+      "grad_norm": 0.0051344577223062515,
+      "learning_rate": 6.669288709165276e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967030560970307,
+      "num_tokens": 80576757.0,
+      "step": 4100
+    },
+    {
+      "entropy": 0.0161895751953125,
+      "epoch": 0.83,
+      "grad_norm": 0.00540529889985919,
+      "learning_rate": 6.59275000453902e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966316163539887,
+      "num_tokens": 81549255.0,
+      "step": 4150
+    },
+    {
+      "entropy": 0.01657470703125,
+      "epoch": 0.84,
+      "grad_norm": 0.005644636228680611,
+      "learning_rate": 6.515793628828447e-05,
+      "loss": 0.0131,
+      "mean_token_accuracy": 0.9965110903978348,
+      "num_tokens": 82499683.0,
+      "step": 4200
+    },
+    {
+      "entropy": 0.0160162353515625,
+      "epoch": 0.85,
+      "grad_norm": 0.004190398845821619,
+      "learning_rate": 6.438439762509332e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966331452131272,
+      "num_tokens": 83475810.0,
+      "step": 4250
+    },
+    {
+      "entropy": 0.01598388671875,
+      "epoch": 0.86,
+      "grad_norm": 0.005400759633630514,
+      "learning_rate": 6.360708690292479e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9966162234544754,
+      "num_tokens": 84432102.0,
+      "step": 4300
+    },
+    {
+      "entropy": 0.014876708984375,
+      "epoch": 0.87,
+      "grad_norm": 0.012628648430109024,
+      "learning_rate": 6.2826207958044e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967326593399047,
+      "num_tokens": 85429626.0,
+      "step": 4350
+    },
+    {
+      "entropy": 0.014339599609375,
+      "epoch": 0.88,
+      "grad_norm": 0.009649231098592281,
+      "learning_rate": 6.204196556242061e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967543077468872,
+      "num_tokens": 86450572.0,
+      "step": 4400
+    },
+    {
+      "entropy": 0.014298095703125,
+      "epoch": 0.89,
+      "grad_norm": 0.0049432734958827496,
+      "learning_rate": 6.125456537003095e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967639350891113,
+      "num_tokens": 87458843.0,
+      "step": 4450
+    },
+    {
+      "entropy": 0.015306396484375,
+      "epoch": 0.9,
+      "grad_norm": 0.004098173696547747,
+      "learning_rate": 6.046421386292887e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9966574615240097,
+      "num_tokens": 88453316.0,
+      "step": 4500
+    },
+    {
+      "entropy": 0.016375732421875,
+      "epoch": 0.91,
+      "grad_norm": 0.005302398465573788,
+      "learning_rate": 5.9671118297099274e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966228741407395,
+      "num_tokens": 89427450.0,
+      "step": 4550
+    },
+    {
+      "entropy": 0.015936279296875,
+      "epoch": 0.92,
+      "grad_norm": 0.004837509244680405,
+      "learning_rate": 5.887548664810896e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966443425416946,
+      "num_tokens": 90400676.0,
+      "step": 4600
+    },
+    {
+      "entropy": 0.0147259521484375,
+      "epoch": 0.93,
+      "grad_norm": 0.008831442333757877,
+      "learning_rate": 5.8077527556568614e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967519730329514,
+      "num_tokens": 91394757.0,
+      "step": 4650
+    },
+    {
+      "entropy": 0.01557373046875,
+      "epoch": 0.94,
+      "grad_norm": 0.0063737258315086365,
+      "learning_rate": 5.727745027342031e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966205650568009,
+      "num_tokens": 92386137.0,
+      "step": 4700
+    },
+    {
+      "entropy": 0.0151513671875,
+      "epoch": 0.95,
+      "grad_norm": 0.005380143877118826,
+      "learning_rate": 5.64754646050652e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9965577948093415,
+      "num_tokens": 93374756.0,
+      "step": 4750
+    },
+    {
+      "entropy": 0.0138897705078125,
+      "epoch": 0.96,
+      "grad_norm": 0.005837676115334034,
+      "learning_rate": 5.567178085834542e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967484200000762,
+      "num_tokens": 94389846.0,
+      "step": 4800
+    },
+    {
+      "entropy": 0.0154791259765625,
+      "epoch": 0.97,
+      "grad_norm": 0.008733139373362064,
+      "learning_rate": 5.486660978539468e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966599184274674,
+      "num_tokens": 95378637.0,
+      "step": 4850
+    },
+    {
+      "entropy": 0.015618896484375,
+      "epoch": 0.98,
+      "grad_norm": 0.016466081142425537,
+      "learning_rate": 5.406016252837219e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966346627473831,
+      "num_tokens": 96357363.0,
+      "step": 4900
+    },
+    {
+      "entropy": 0.0171966552734375,
+      "epoch": 0.99,
+      "grad_norm": 0.007259083911776543,
+      "learning_rate": 5.325265056409442e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965799397230148,
+      "num_tokens": 97295349.0,
+      "step": 4950
+    },
+    {
+      "entropy": 0.01597900390625,
+      "epoch": 1.0,
+      "grad_norm": 0.007410865277051926,
+      "learning_rate": 5.244428564857874e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.996621105670929,
+      "num_tokens": 98262502.0,
+      "step": 5000
+    },
+    {
+      "entropy": 0.0154595947265625,
+      "epoch": 1.01,
+      "grad_norm": 0.0048329527489840984,
+      "learning_rate": 5.163527976151428e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966327011585235,
+      "num_tokens": 99235771.0,
+      "step": 5050
+    },
+    {
+      "entropy": 0.0155859375,
+      "epoch": 1.02,
+      "grad_norm": 0.004490161780267954,
+      "learning_rate": 5.0825845050673834e-05,
+      "loss": 0.0129,
+      "mean_token_accuracy": 0.9965754437446595,
+      "num_tokens": 100195990.0,
+      "step": 5100
+    },
+    {
+      "entropy": 0.01652099609375,
+      "epoch": 1.03,
+      "grad_norm": 0.00914011336863041,
+      "learning_rate": 5.0016193776281794e-05,
+      "loss": 0.013,
+      "mean_token_accuracy": 0.9964404582977295,
+      "num_tokens": 101140160.0,
+      "step": 5150
+    },
+    {
+      "entropy": 0.015284423828125,
+      "epoch": 1.04,
+      "grad_norm": 0.0057498314417898655,
+      "learning_rate": 4.920653825535258e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.99665458381176,
+      "num_tokens": 102128346.0,
+      "step": 5200
+    },
+    {
+      "entropy": 0.015015869140625,
+      "epoch": 1.05,
+      "grad_norm": 0.004660275299102068,
+      "learning_rate": 4.839709080601419e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967097079753876,
+      "num_tokens": 103125113.0,
+      "step": 5250
+    },
+    {
+      "entropy": 0.0143218994140625,
+      "epoch": 1.06,
+      "grad_norm": 0.006082482635974884,
+      "learning_rate": 4.758806369183149e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967189061641694,
+      "num_tokens": 104136988.0,
+      "step": 5300
+    },
+    {
+      "entropy": 0.0155865478515625,
+      "epoch": 1.07,
+      "grad_norm": 0.00519182626157999,
+      "learning_rate": 4.6779669066143716e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966564351320266,
+      "num_tokens": 105122070.0,
+      "step": 5350
+    },
+    {
+      "entropy": 0.0145452880859375,
+      "epoch": 1.08,
+      "grad_norm": 0.005050841718912125,
+      "learning_rate": 4.597211891643093e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967609119415283,
+      "num_tokens": 106117109.0,
+      "step": 5400
+    },
+    {
+      "entropy": 0.0144915771484375,
+      "epoch": 1.09,
+      "grad_norm": 0.004848674405366182,
+      "learning_rate": 4.5165625008724035e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967008078098297,
+      "num_tokens": 107110319.0,
+      "step": 5450
+    },
+    {
+      "entropy": 0.0150311279296875,
+      "epoch": 1.1,
+      "grad_norm": 0.0036910909693688154,
+      "learning_rate": 4.4360398832072764e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967303448915481,
+      "num_tokens": 108104820.0,
+      "step": 5500
+    },
+    {
+      "entropy": 0.015078125,
+      "epoch": 1.11,
+      "grad_norm": 0.0053949966095387936,
+      "learning_rate": 4.3556651543086364e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966647911071778,
+      "num_tokens": 109091493.0,
+      "step": 5550
+    },
+    {
+      "entropy": 0.0147705078125,
+      "epoch": 1.12,
+      "grad_norm": 0.0037009185180068016,
+      "learning_rate": 4.275459391056142e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966964650154114,
+      "num_tokens": 110077375.0,
+      "step": 5600
+    },
+    {
+      "entropy": 0.0162713623046875,
+      "epoch": 1.13,
+      "grad_norm": 0.0031313870567828417,
+      "learning_rate": 4.195443626021139e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966827458143235,
+      "num_tokens": 111042295.0,
+      "step": 5650
+    },
+    {
+      "entropy": 0.0145147705078125,
+      "epoch": 1.1400000000000001,
+      "grad_norm": 0.004629215691238642,
+      "learning_rate": 4.1156388419512324e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966849237680435,
+      "num_tokens": 112033836.0,
+      "step": 5700
+    },
+    {
+      "entropy": 0.0152557373046875,
+      "epoch": 1.15,
+      "grad_norm": 0.00527191162109375,
+      "learning_rate": 4.0360659662679265e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.996674497127533,
+      "num_tokens": 113006848.0,
+      "step": 5750
+    },
+    {
+      "entropy": 0.0149151611328125,
+      "epoch": 1.16,
+      "grad_norm": 0.004986864514648914,
+      "learning_rate": 3.956745865578773e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966980391740798,
+      "num_tokens": 113998253.0,
+      "step": 5800
+    },
+    {
+      "entropy": 0.015072021484375,
+      "epoch": 1.17,
+      "grad_norm": 0.004452258348464966,
+      "learning_rate": 3.877699340205455e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966688090562821,
+      "num_tokens": 114979026.0,
+      "step": 5850
+    },
+    {
+      "entropy": 0.0157171630859375,
+      "epoch": 1.18,
+      "grad_norm": 0.0048246984370052814,
+      "learning_rate": 3.798947118729282e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966557210683823,
+      "num_tokens": 115947279.0,
+      "step": 5900
+    },
+    {
+      "entropy": 0.015772705078125,
+      "epoch": 1.19,
+      "grad_norm": 0.003676283173263073,
+      "learning_rate": 3.720509852555456e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966129893064499,
+      "num_tokens": 116911661.0,
+      "step": 5950
+    },
+    {
+      "entropy": 0.01443603515625,
+      "epoch": 1.2,
+      "grad_norm": 0.004810075741261244,
+      "learning_rate": 3.642408110497616e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966686731576919,
+      "num_tokens": 117890306.0,
+      "step": 6000
+    },
+    {
+      "entropy": 0.0159271240234375,
+      "epoch": 1.21,
+      "grad_norm": 0.003690144745633006,
+      "learning_rate": 3.5646623733840134e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966743797063827,
+      "num_tokens": 118844495.0,
+      "step": 6050
+    },
+    {
+      "entropy": 0.0148345947265625,
+      "epoch": 1.22,
+      "grad_norm": 0.0036596781574189663,
+      "learning_rate": 3.4872930286867763e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966356498003006,
+      "num_tokens": 119807321.0,
+      "step": 6100
+    },
+    {
+      "entropy": 0.0147174072265625,
+      "epoch": 1.23,
+      "grad_norm": 0.0046323081478476524,
+      "learning_rate": 3.4103203651756403e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967032814025879,
+      "num_tokens": 120794059.0,
+      "step": 6150
+    },
+    {
+      "entropy": 0.0155657958984375,
+      "epoch": 1.24,
+      "grad_norm": 0.005745012313127518,
+      "learning_rate": 3.333764567597579e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.996661651134491,
+      "num_tokens": 121765426.0,
+      "step": 6200
+    },
+    {
+      "entropy": 0.0151751708984375,
+      "epoch": 1.25,
+      "grad_norm": 0.00422937236726284,
+      "learning_rate": 3.257645711383699e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9967177057266235,
+      "num_tokens": 122747186.0,
+      "step": 6250
+    },
+    {
+      "entropy": 0.014896240234375,
+      "epoch": 1.26,
+      "grad_norm": 0.0041396524757146835,
+      "learning_rate": 3.1819837573848055e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967236334085464,
+      "num_tokens": 123735045.0,
+      "step": 6300
+    },
+    {
+      "entropy": 0.014747314453125,
+      "epoch": 1.27,
+      "grad_norm": 0.0034245741553604603,
+      "learning_rate": 3.106798546637019e-05,
+      "loss": 0.012,
+      "mean_token_accuracy": 0.9968506091833115,
+      "num_tokens": 124748439.0,
+      "step": 6350
+    },
+    {
+      "entropy": 0.014637451171875,
+      "epoch": 1.28,
+      "grad_norm": 0.0044257547706365585,
+      "learning_rate": 3.0321097951588016e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.996695158481598,
+      "num_tokens": 125723365.0,
+      "step": 6400
+    },
+    {
+      "entropy": 0.0153790283203125,
+      "epoch": 1.29,
+      "grad_norm": 0.005239363294094801,
+      "learning_rate": 2.957937088780779e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966009676456451,
+      "num_tokens": 126678062.0,
+      "step": 6450
+    },
+    {
+      "entropy": 0.0145068359375,
+      "epoch": 1.3,
+      "grad_norm": 0.005349221173673868,
+      "learning_rate": 2.8842998780096896e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967689532041549,
+      "num_tokens": 127675372.0,
+      "step": 6500
+    },
+    {
+      "entropy": 0.0149395751953125,
+      "epoch": 1.31,
+      "grad_norm": 0.005388084799051285,
+      "learning_rate": 2.811217472927835e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967477923631668,
+      "num_tokens": 128665669.0,
+      "step": 6550
+    },
+    {
+      "entropy": 0.0159820556640625,
+      "epoch": 1.32,
+      "grad_norm": 0.006312755402177572,
+      "learning_rate": 2.7387090381293372e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966998076438904,
+      "num_tokens": 129622542.0,
+      "step": 6600
+    },
+    {
+      "entropy": 0.016163330078125,
+      "epoch": 1.33,
+      "grad_norm": 0.005128033459186554,
+      "learning_rate": 2.6667935876945582e-05,
+      "loss": 0.0124,
+      "mean_token_accuracy": 0.9966605240106583,
+      "num_tokens": 130598823.0,
+      "step": 6650
+    },
+    {
+      "entropy": 0.017476806640625,
+      "epoch": 1.34,
+      "grad_norm": 0.005179600324481726,
+      "learning_rate": 2.59548998020399e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.996593764424324,
+      "num_tokens": 131552204.0,
+      "step": 6700
+    },
+    {
+      "entropy": 0.0151275634765625,
+      "epoch": 1.35,
+      "grad_norm": 0.005807869601994753,
+      "learning_rate": 2.5248169137929156e-05,
+      "loss": 0.0126,
+      "mean_token_accuracy": 0.9966244864463806,
+      "num_tokens": 132526021.0,
+      "step": 6750
+    },
+    {
+      "entropy": 0.015770263671875,
+      "epoch": 1.3599999999999999,
+      "grad_norm": 0.005202575121074915,
+      "learning_rate": 2.4547929212481435e-05,
+      "loss": 0.0127,
+      "mean_token_accuracy": 0.9966981512308121,
+      "num_tokens": 133483207.0,
+      "step": 6800
+    },
+    {
+      "entropy": 0.014925537109375,
+      "epoch": 1.37,
+      "grad_norm": 0.003400217741727829,
+      "learning_rate": 2.3854363651481194e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967307335138321,
+      "num_tokens": 134482802.0,
+      "step": 6850
+    },
+    {
+      "entropy": 0.0151629638671875,
+      "epoch": 1.38,
+      "grad_norm": 0.004814519081264734,
+      "learning_rate": 2.3167654330476412e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.996745794415474,
+      "num_tokens": 135484861.0,
+      "step": 6900
+    },
+    {
+      "entropy": 0.0138885498046875,
+      "epoch": 1.3900000000000001,
+      "grad_norm": 0.0049650054425001144,
+      "learning_rate": 2.2487981327084995e-05,
+      "loss": 0.012,
+      "mean_token_accuracy": 0.9967457884550095,
+      "num_tokens": 136499177.0,
+      "step": 6950
+    },
+    {
+      "entropy": 0.014207763671875,
+      "epoch": 1.4,
+      "grad_norm": 0.0068522971123456955,
+      "learning_rate": 2.1815522873772475e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967319291830062,
+      "num_tokens": 137503523.0,
+      "step": 7000
+    },
+    {
+      "entropy": 0.0145257568359375,
+      "epoch": 1.41,
+      "grad_norm": 0.00624112319201231,
+      "learning_rate": 2.1150455311113708e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9966984778642655,
+      "num_tokens": 138497821.0,
+      "step": 7050
+    },
+    {
+      "entropy": 0.0151031494140625,
+      "epoch": 1.42,
+      "grad_norm": 0.00566850695759058,
+      "learning_rate": 2.0492953041550483e-05,
+      "loss": 0.0123,
+      "mean_token_accuracy": 0.9967049795389176,
+      "num_tokens": 139480692.0,
+      "step": 7100
+    },
+    {
+      "entropy": 0.014796142578125,
+      "epoch": 1.43,
+      "grad_norm": 0.00434489268809557,
+      "learning_rate": 1.9843188483657697e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967473483085633,
+      "num_tokens": 140475746.0,
+      "step": 7150
+    },
+    {
+      "entropy": 0.0156658935546875,
+      "epoch": 1.44,
+      "grad_norm": 0.004734216723591089,
+      "learning_rate": 1.9201332026929396e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966971826553345,
+      "num_tokens": 141443323.0,
+      "step": 7200
+    },
+    {
+      "entropy": 0.0151654052734375,
+      "epoch": 1.45,
+      "grad_norm": 0.004659540485590696,
+      "learning_rate": 1.8567551987097188e-05,
+      "loss": 0.0121,
+      "mean_token_accuracy": 0.9967643207311631,
+      "num_tokens": 142442496.0,
+      "step": 7250
+    },
+    {
+      "entropy": 0.0146234130859375,
+      "epoch": 1.46,
+      "grad_norm": 0.0037653304170817137,
+      "learning_rate": 1.794201456199231e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967849618196487,
+      "num_tokens": 143436493.0,
+      "step": 7300
+    },
+    {
+      "entropy": 0.016064453125,
+      "epoch": 1.47,
+      "grad_norm": 0.004325070418417454,
+      "learning_rate": 1.732488378796311e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9967092323303223,
+      "num_tokens": 144400422.0,
+      "step": 7350
+    },
+    {
+      "entropy": 0.0147503662109375,
+      "epoch": 1.48,
+      "grad_norm": 0.0051532830111682415,
+      "learning_rate": 1.671632149685943e-05,
+      "loss": 0.012,
+      "mean_token_accuracy": 0.9967712318897247,
+      "num_tokens": 145397669.0,
+      "step": 7400
+    },
+    {
+      "entropy": 0.0142218017578125,
+      "epoch": 1.49,
+      "grad_norm": 0.01885133981704712,
+      "learning_rate": 1.611648727359498e-05,
+      "loss": 0.0122,
+      "mean_token_accuracy": 0.9967398834228516,
+      "num_tokens": 146393328.0,
+      "step": 7450
+    },
+    {
+      "entropy": 0.015579833984375,
+      "epoch": 1.5,
+      "grad_norm": 0.00416885782033205,
+      "learning_rate": 1.5525538414298925e-05,
+      "loss": 0.0125,
+      "mean_token_accuracy": 0.9966780996322632,
+      "num_tokens": 147355946.0,
+      "step": 7500
+    }
+  ],
+  "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": 2.559137821805773e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

qwen2.5-coder-3B_r64_lr1e-4/checkpoint-7500/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_lr1e-4/checkpoint-8000/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
+  }
+}