Add files using upload-large-folder tool

README.md

@@ -0,0 +1,143 @@
+---
+library_name: transformers
+license: apache-2.0
+base_model: Qwen/Qwen3-1.7B
+tags:
+- generated_from_trainer
+datasets:
+- sumuks/essential-web-v1.0-sample-100M-with-cleaned-responses-sft
+model-index:
+- name: output/1.7B-Instruct-Tuned-New-Data
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
+<details><summary>See axolotl config</summary>
+
+axolotl version: `0.11.0`
+```yaml
+base_model: Qwen/Qwen3-1.7B
+
+# plugins:
+#   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+# plugins:
+#   - axolotl.integrations.liger.LigerPlugin
+
+# liger_rope: true
+# liger_rms_norm: true
+# liger_glu_activation: true
+# liger_layer_norm: true
+# liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: sumuks/essential-web-v1.0-sample-100M-with-cleaned-responses-sft
+    type: chat_template
+    field_messages: conversations
+    split: train
+val_set_size: 0.05
+dataset_prepared_path: dataset/prepared_dataset_1.7b
+
+train_on_inputs: false
+output_dir: ./output/1.7B-Instruct-Tuned-New-Data
+chat_template: qwen3
+sequence_len: 8192
+sample_packing: true
+eval_sample_packing: true
+# pad_to_sequence_len: true
+
+wandb_project: essential-web-sft
+wandb_name: qwen3-1.7b-sft-new-data
+
+gradient_accumulation_steps: 4
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+flash_attention: true
+micro_batch_size: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+num_epochs: 1
+
+load_best_model_at_end: true
+metric_for_best_model: loss
+greater_is_better: false
+
+early_stopping_patience: 3
+bf16: auto
+tf32: true
+
+logging_steps: 5
+
+deepspeed: ./configs_prod/zero3.json
+
+save_steps: 500
+eval_steps: 500
+
+warmup_ratio: 0.05
+# save_first_step: true
+```
+
+</details><br>
+
+# output/1.7B-Instruct-Tuned-New-Data
+
+This model is a fine-tuned version of [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) on the sumuks/essential-web-v1.0-sample-100M-with-cleaned-responses-sft dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.3669
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 2e-05
+- train_batch_size: 1
+- eval_batch_size: 1
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 2
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 8
+- total_eval_batch_size: 2
+- optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_steps: 164
+- training_steps: 3297
+
+### Training results
+
+| Training Loss | Epoch  | Step | Validation Loss |
+|:-------------:|:------:|:----:|:---------------:|
+| No log        | 0      | 0    | 0.8829          |
+| 0.3689        | 0.1517 | 500  | 0.4088          |
+| 0.3919        | 0.3033 | 1000 | 0.3952          |
+| 0.386         | 0.4550 | 1500 | 0.3839          |
+| 0.409         | 0.6066 | 2000 | 0.3755          |
+| 0.3473        | 0.7583 | 2500 | 0.3694          |
+| 0.3518        | 0.9099 | 3000 | 0.3669          |
+
+
+### Framework versions
+
+- Transformers 4.53.1
+- Pytorch 2.7.1+cu126
+- Datasets 3.6.0
+- Tokenizers 0.21.2

added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|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
+}

chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# 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' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first 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' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n\n' }}
+    {%- endif %}
+{%- endif %}

checkpoint-2000/added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoint-2000/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# 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' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first 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' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n\n' }}
+    {%- endif %}
+{%- endif %}

checkpoint-2000/config.json

@@ -0,0 +1,59 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.53.1",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-2000/generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "4.53.1"
+}

checkpoint-2000/latest

@@ -0,0 +1 @@
+global_step2000

checkpoint-2000/special_tokens_map.json

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

checkpoint-2000/tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "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
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-2000/trainer_state.json

@@ -0,0 +1,2883 @@
+{
+  "best_global_step": 2000,
+  "best_metric": 0.3754875957965851,
+  "best_model_checkpoint": "./output/1.7B-Instruct-Tuned-New-Data/checkpoint-2000",
+  "epoch": 0.6066120715802245,
+  "eval_steps": 500,
+  "global_step": 2000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0,
+      "eval_loss": 0.8828833103179932,
+      "eval_runtime": 183.2487,
+      "eval_samples_per_second": 48.759,
+      "eval_steps_per_second": 24.382,
+      "step": 0
+    },
+    {
+      "epoch": 0.001516530178950561,
+      "grad_norm": 24.921934804682877,
+      "learning_rate": 4.878048780487805e-07,
+      "loss": 0.8639,
+      "step": 5
+    },
+    {
+      "epoch": 0.003033060357901122,
+      "grad_norm": 22.993835600118015,
+      "learning_rate": 1.0975609756097562e-06,
+      "loss": 0.8551,
+      "step": 10
+    },
+    {
+      "epoch": 0.004549590536851683,
+      "grad_norm": 21.595580129534707,
+      "learning_rate": 1.707317073170732e-06,
+      "loss": 0.8283,
+      "step": 15
+    },
+    {
+      "epoch": 0.006066120715802244,
+      "grad_norm": 5.010673102150899,
+      "learning_rate": 2.317073170731708e-06,
+      "loss": 0.559,
+      "step": 20
+    },
+    {
+      "epoch": 0.0075826508947528055,
+      "grad_norm": 2.8725236445814146,
+      "learning_rate": 2.926829268292683e-06,
+      "loss": 0.5412,
+      "step": 25
+    },
+    {
+      "epoch": 0.009099181073703366,
+      "grad_norm": 2.187108009423291,
+      "learning_rate": 3.5365853658536588e-06,
+      "loss": 0.521,
+      "step": 30
+    },
+    {
+      "epoch": 0.010615711252653927,
+      "grad_norm": 1.4262904456959642,
+      "learning_rate": 4.146341463414634e-06,
+      "loss": 0.5138,
+      "step": 35
+    },
+    {
+      "epoch": 0.012132241431604488,
+      "grad_norm": 1.5492098631988767,
+      "learning_rate": 4.75609756097561e-06,
+      "loss": 0.5069,
+      "step": 40
+    },
+    {
+      "epoch": 0.01364877161055505,
+      "grad_norm": 1.2785619453952701,
+      "learning_rate": 5.365853658536586e-06,
+      "loss": 0.4792,
+      "step": 45
+    },
+    {
+      "epoch": 0.015165301789505611,
+      "grad_norm": 1.182973976230221,
+      "learning_rate": 5.9756097560975615e-06,
+      "loss": 0.4029,
+      "step": 50
+    },
+    {
+      "epoch": 0.01668183196845617,
+      "grad_norm": 1.1563458384082506,
+      "learning_rate": 6.585365853658538e-06,
+      "loss": 0.385,
+      "step": 55
+    },
+    {
+      "epoch": 0.018198362147406732,
+      "grad_norm": 1.3349292384420064,
+      "learning_rate": 7.1951219512195125e-06,
+      "loss": 0.4144,
+      "step": 60
+    },
+    {
+      "epoch": 0.019714892326357293,
+      "grad_norm": 0.9335126123634385,
+      "learning_rate": 7.804878048780489e-06,
+      "loss": 0.4261,
+      "step": 65
+    },
+    {
+      "epoch": 0.021231422505307854,
+      "grad_norm": 1.30792221561379,
+      "learning_rate": 8.414634146341464e-06,
+      "loss": 0.4606,
+      "step": 70
+    },
+    {
+      "epoch": 0.022747952684258416,
+      "grad_norm": 1.6097633219159984,
+      "learning_rate": 9.02439024390244e-06,
+      "loss": 0.4276,
+      "step": 75
+    },
+    {
+      "epoch": 0.024264482863208977,
+      "grad_norm": 1.1673685397111215,
+      "learning_rate": 9.634146341463415e-06,
+      "loss": 0.445,
+      "step": 80
+    },
+    {
+      "epoch": 0.025781013042159538,
+      "grad_norm": 1.3515026545633433,
+      "learning_rate": 1.024390243902439e-05,
+      "loss": 0.496,
+      "step": 85
+    },
+    {
+      "epoch": 0.0272975432211101,
+      "grad_norm": 1.5361335286977482,
+      "learning_rate": 1.0853658536585368e-05,
+      "loss": 0.4189,
+      "step": 90
+    },
+    {
+      "epoch": 0.02881407340006066,
+      "grad_norm": 1.3323687104870683,
+      "learning_rate": 1.1463414634146342e-05,
+      "loss": 0.4461,
+      "step": 95
+    },
+    {
+      "epoch": 0.030330603579011222,
+      "grad_norm": 1.3055093124799528,
+      "learning_rate": 1.2073170731707317e-05,
+      "loss": 0.4558,
+      "step": 100
+    },
+    {
+      "epoch": 0.03184713375796178,
+      "grad_norm": 1.4071077040506643,
+      "learning_rate": 1.2682926829268294e-05,
+      "loss": 0.4608,
+      "step": 105
+    },
+    {
+      "epoch": 0.03336366393691234,
+      "grad_norm": 1.140971116003336,
+      "learning_rate": 1.329268292682927e-05,
+      "loss": 0.4046,
+      "step": 110
+    },
+    {
+      "epoch": 0.034880194115862906,
+      "grad_norm": 1.3427459277886848,
+      "learning_rate": 1.3902439024390244e-05,
+      "loss": 0.3874,
+      "step": 115
+    },
+    {
+      "epoch": 0.036396724294813464,
+      "grad_norm": 1.1775148547297358,
+      "learning_rate": 1.451219512195122e-05,
+      "loss": 0.4605,
+      "step": 120
+    },
+    {
+      "epoch": 0.03791325447376403,
+      "grad_norm": 1.2616670504697265,
+      "learning_rate": 1.5121951219512196e-05,
+      "loss": 0.4781,
+      "step": 125
+    },
+    {
+      "epoch": 0.039429784652714586,
+      "grad_norm": 1.4344564952473762,
+      "learning_rate": 1.5731707317073173e-05,
+      "loss": 0.4337,
+      "step": 130
+    },
+    {
+      "epoch": 0.04094631483166515,
+      "grad_norm": 1.7662014691737036,
+      "learning_rate": 1.6341463414634145e-05,
+      "loss": 0.4636,
+      "step": 135
+    },
+    {
+      "epoch": 0.04246284501061571,
+      "grad_norm": 1.3917668573790878,
+      "learning_rate": 1.6951219512195124e-05,
+      "loss": 0.4651,
+      "step": 140
+    },
+    {
+      "epoch": 0.043979375189566274,
+      "grad_norm": 1.1345687800745432,
+      "learning_rate": 1.75609756097561e-05,
+      "loss": 0.3976,
+      "step": 145
+    },
+    {
+      "epoch": 0.04549590536851683,
+      "grad_norm": 1.0517487527616944,
+      "learning_rate": 1.8170731707317075e-05,
+      "loss": 0.4172,
+      "step": 150
+    },
+    {
+      "epoch": 0.047012435547467396,
+      "grad_norm": 1.0174150315487598,
+      "learning_rate": 1.878048780487805e-05,
+      "loss": 0.3947,
+      "step": 155
+    },
+    {
+      "epoch": 0.048528965726417954,
+      "grad_norm": 1.3496351705883955,
+      "learning_rate": 1.9390243902439026e-05,
+      "loss": 0.4429,
+      "step": 160
+    },
+    {
+      "epoch": 0.05004549590536852,
+      "grad_norm": 1.1321668963681037,
+      "learning_rate": 2e-05,
+      "loss": 0.4044,
+      "step": 165
+    },
+    {
+      "epoch": 0.051562026084319076,
+      "grad_norm": 1.0951786995673565,
+      "learning_rate": 1.999987431366603e-05,
+      "loss": 0.3749,
+      "step": 170
+    },
+    {
+      "epoch": 0.05307855626326964,
+      "grad_norm": 1.1089458391034641,
+      "learning_rate": 1.999949725782353e-05,
+      "loss": 0.4227,
+      "step": 175
+    },
+    {
+      "epoch": 0.0545950864422202,
+      "grad_norm": 1.1845203588062014,
+      "learning_rate": 1.9998868841950646e-05,
+      "loss": 0.3867,
+      "step": 180
+    },
+    {
+      "epoch": 0.056111616621170764,
+      "grad_norm": 0.8571685139076549,
+      "learning_rate": 1.9997989081844044e-05,
+      "loss": 0.4351,
+      "step": 185
+    },
+    {
+      "epoch": 0.05762814680012132,
+      "grad_norm": 0.9900250625278807,
+      "learning_rate": 1.999685799961849e-05,
+      "loss": 0.4125,
+      "step": 190
+    },
+    {
+      "epoch": 0.059144676979071886,
+      "grad_norm": 1.1898649968753983,
+      "learning_rate": 1.999547562370629e-05,
+      "loss": 0.4217,
+      "step": 195
+    },
+    {
+      "epoch": 0.060661207158022444,
+      "grad_norm": 1.1463755374202382,
+      "learning_rate": 1.99938419888566e-05,
+      "loss": 0.4564,
+      "step": 200
+    },
+    {
+      "epoch": 0.06217773733697301,
+      "grad_norm": 1.3479323790987279,
+      "learning_rate": 1.9991957136134542e-05,
+      "loss": 0.5052,
+      "step": 205
+    },
+    {
+      "epoch": 0.06369426751592357,
+      "grad_norm": 1.475205459236139,
+      "learning_rate": 1.9989821112920155e-05,
+      "loss": 0.38,
+      "step": 210
+    },
+    {
+      "epoch": 0.06521079769487413,
+      "grad_norm": 1.1629718970331968,
+      "learning_rate": 1.9987433972907225e-05,
+      "loss": 0.4202,
+      "step": 215
+    },
+    {
+      "epoch": 0.06672732787382468,
+      "grad_norm": 1.0931736708866155,
+      "learning_rate": 1.998479577610193e-05,
+      "loss": 0.3596,
+      "step": 220
+    },
+    {
+      "epoch": 0.06824385805277525,
+      "grad_norm": 1.0829052216446347,
+      "learning_rate": 1.9981906588821322e-05,
+      "loss": 0.4087,
+      "step": 225
+    },
+    {
+      "epoch": 0.06976038823172581,
+      "grad_norm": 1.247069136713069,
+      "learning_rate": 1.997876648369168e-05,
+      "loss": 0.4044,
+      "step": 230
+    },
+    {
+      "epoch": 0.07127691841067638,
+      "grad_norm": 1.0789938627982327,
+      "learning_rate": 1.9975375539646656e-05,
+      "loss": 0.3966,
+      "step": 235
+    },
+    {
+      "epoch": 0.07279344858962693,
+      "grad_norm": 1.2874248378984725,
+      "learning_rate": 1.997173384192532e-05,
+      "loss": 0.3919,
+      "step": 240
+    },
+    {
+      "epoch": 0.07430997876857749,
+      "grad_norm": 1.251206148462947,
+      "learning_rate": 1.9967841482070002e-05,
+      "loss": 0.4433,
+      "step": 245
+    },
+    {
+      "epoch": 0.07582650894752806,
+      "grad_norm": 1.2436949055413618,
+      "learning_rate": 1.996369855792398e-05,
+      "loss": 0.431,
+      "step": 250
+    },
+    {
+      "epoch": 0.07734303912647862,
+      "grad_norm": 1.2591168062725386,
+      "learning_rate": 1.9959305173629056e-05,
+      "loss": 0.4249,
+      "step": 255
+    },
+    {
+      "epoch": 0.07885956930542917,
+      "grad_norm": 1.1569304716868531,
+      "learning_rate": 1.9954661439622894e-05,
+      "loss": 0.4274,
+      "step": 260
+    },
+    {
+      "epoch": 0.08037609948437974,
+      "grad_norm": 1.1825334487051853,
+      "learning_rate": 1.994976747263628e-05,
+      "loss": 0.3794,
+      "step": 265
+    },
+    {
+      "epoch": 0.0818926296633303,
+      "grad_norm": 1.4809085739387609,
+      "learning_rate": 1.9944623395690162e-05,
+      "loss": 0.4105,
+      "step": 270
+    },
+    {
+      "epoch": 0.08340915984228087,
+      "grad_norm": 1.0467676858303625,
+      "learning_rate": 1.9939229338092584e-05,
+      "loss": 0.3861,
+      "step": 275
+    },
+    {
+      "epoch": 0.08492569002123142,
+      "grad_norm": 1.1992281733632604,
+      "learning_rate": 1.99335854354354e-05,
+      "loss": 0.3317,
+      "step": 280
+    },
+    {
+      "epoch": 0.08644222020018198,
+      "grad_norm": 1.3404254075615103,
+      "learning_rate": 1.9927691829590903e-05,
+      "loss": 0.4457,
+      "step": 285
+    },
+    {
+      "epoch": 0.08795875037913255,
+      "grad_norm": 1.3258249809215283,
+      "learning_rate": 1.992154866870824e-05,
+      "loss": 0.4478,
+      "step": 290
+    },
+    {
+      "epoch": 0.08947528055808311,
+      "grad_norm": 1.215305619652228,
+      "learning_rate": 1.9915156107209673e-05,
+      "loss": 0.4032,
+      "step": 295
+    },
+    {
+      "epoch": 0.09099181073703366,
+      "grad_norm": 1.1528812652880362,
+      "learning_rate": 1.9908514305786733e-05,
+      "loss": 0.3796,
+      "step": 300
+    },
+    {
+      "epoch": 0.09250834091598423,
+      "grad_norm": 1.1857942136064035,
+      "learning_rate": 1.990162343139616e-05,
+      "loss": 0.435,
+      "step": 305
+    },
+    {
+      "epoch": 0.09402487109493479,
+      "grad_norm": 1.220916675694686,
+      "learning_rate": 1.989448365725569e-05,
+      "loss": 0.3798,
+      "step": 310
+    },
+    {
+      "epoch": 0.09554140127388536,
+      "grad_norm": 0.985856274220688,
+      "learning_rate": 1.988709516283974e-05,
+      "loss": 0.3925,
+      "step": 315
+    },
+    {
+      "epoch": 0.09705793145283591,
+      "grad_norm": 1.0859800217625073,
+      "learning_rate": 1.987945813387486e-05,
+      "loss": 0.4437,
+      "step": 320
+    },
+    {
+      "epoch": 0.09857446163178647,
+      "grad_norm": 1.0765506678345211,
+      "learning_rate": 1.9871572762335085e-05,
+      "loss": 0.3847,
+      "step": 325
+    },
+    {
+      "epoch": 0.10009099181073704,
+      "grad_norm": 1.1464592931749376,
+      "learning_rate": 1.9863439246437108e-05,
+      "loss": 0.4129,
+      "step": 330
+    },
+    {
+      "epoch": 0.10160752198968759,
+      "grad_norm": 1.077561921028752,
+      "learning_rate": 1.985505779063528e-05,
+      "loss": 0.3808,
+      "step": 335
+    },
+    {
+      "epoch": 0.10312405216863815,
+      "grad_norm": 0.8205386721296154,
+      "learning_rate": 1.98464286056165e-05,
+      "loss": 0.3598,
+      "step": 340
+    },
+    {
+      "epoch": 0.10464058234758872,
+      "grad_norm": 1.1795375866850706,
+      "learning_rate": 1.9837551908294887e-05,
+      "loss": 0.3986,
+      "step": 345
+    },
+    {
+      "epoch": 0.10615711252653928,
+      "grad_norm": 1.2649777397213469,
+      "learning_rate": 1.9828427921806358e-05,
+      "loss": 0.4592,
+      "step": 350
+    },
+    {
+      "epoch": 0.10767364270548983,
+      "grad_norm": 1.2819761889608279,
+      "learning_rate": 1.9819056875502986e-05,
+      "loss": 0.423,
+      "step": 355
+    },
+    {
+      "epoch": 0.1091901728844404,
+      "grad_norm": 1.0892900872427516,
+      "learning_rate": 1.980943900494727e-05,
+      "loss": 0.4253,
+      "step": 360
+    },
+    {
+      "epoch": 0.11070670306339096,
+      "grad_norm": 1.0090566077966177,
+      "learning_rate": 1.979957455190618e-05,
+      "loss": 0.41,
+      "step": 365
+    },
+    {
+      "epoch": 0.11222323324234153,
+      "grad_norm": 1.1111913464704366,
+      "learning_rate": 1.9789463764345113e-05,
+      "loss": 0.4219,
+      "step": 370
+    },
+    {
+      "epoch": 0.11373976342129208,
+      "grad_norm": 0.9666799079863133,
+      "learning_rate": 1.9779106896421627e-05,
+      "loss": 0.4593,
+      "step": 375
+    },
+    {
+      "epoch": 0.11525629360024264,
+      "grad_norm": 1.2048376185649605,
+      "learning_rate": 1.9768504208479077e-05,
+      "loss": 0.4312,
+      "step": 380
+    },
+    {
+      "epoch": 0.11677282377919321,
+      "grad_norm": 1.3164491474278361,
+      "learning_rate": 1.975765596704006e-05,
+      "loss": 0.3745,
+      "step": 385
+    },
+    {
+      "epoch": 0.11828935395814377,
+      "grad_norm": 0.8809824695823545,
+      "learning_rate": 1.9746562444799712e-05,
+      "loss": 0.397,
+      "step": 390
+    },
+    {
+      "epoch": 0.11980588413709432,
+      "grad_norm": 1.270641853983644,
+      "learning_rate": 1.9735223920618857e-05,
+      "loss": 0.4513,
+      "step": 395
+    },
+    {
+      "epoch": 0.12132241431604489,
+      "grad_norm": 1.291191652244439,
+      "learning_rate": 1.9723640679517015e-05,
+      "loss": 0.4401,
+      "step": 400
+    },
+    {
+      "epoch": 0.12283894449499545,
+      "grad_norm": 1.0763253903156915,
+      "learning_rate": 1.9711813012665198e-05,
+      "loss": 0.4089,
+      "step": 405
+    },
+    {
+      "epoch": 0.12435547467394602,
+      "grad_norm": 1.0687463364798888,
+      "learning_rate": 1.9699741217378625e-05,
+      "loss": 0.424,
+      "step": 410
+    },
+    {
+      "epoch": 0.12587200485289657,
+      "grad_norm": 1.038145624225704,
+      "learning_rate": 1.9687425597109238e-05,
+      "loss": 0.3912,
+      "step": 415
+    },
+    {
+      "epoch": 0.12738853503184713,
+      "grad_norm": 0.9636211495905858,
+      "learning_rate": 1.9674866461438065e-05,
+      "loss": 0.3856,
+      "step": 420
+    },
+    {
+      "epoch": 0.1289050652107977,
+      "grad_norm": 1.2745603021068876,
+      "learning_rate": 1.966206412606745e-05,
+      "loss": 0.4363,
+      "step": 425
+    },
+    {
+      "epoch": 0.13042159538974826,
+      "grad_norm": 1.0775019059829327,
+      "learning_rate": 1.964901891281312e-05,
+      "loss": 0.3764,
+      "step": 430
+    },
+    {
+      "epoch": 0.13193812556869883,
+      "grad_norm": 0.9079407408157604,
+      "learning_rate": 1.9635731149596075e-05,
+      "loss": 0.4002,
+      "step": 435
+    },
+    {
+      "epoch": 0.13345465574764936,
+      "grad_norm": 1.0314514042029734,
+      "learning_rate": 1.962220117043436e-05,
+      "loss": 0.3707,
+      "step": 440
+    },
+    {
+      "epoch": 0.13497118592659993,
+      "grad_norm": 1.2219058440800525,
+      "learning_rate": 1.9608429315434683e-05,
+      "loss": 0.4086,
+      "step": 445
+    },
+    {
+      "epoch": 0.1364877161055505,
+      "grad_norm": 1.1625549762597092,
+      "learning_rate": 1.959441593078383e-05,
+      "loss": 0.3905,
+      "step": 450
+    },
+    {
+      "epoch": 0.13800424628450106,
+      "grad_norm": 1.1998854874780447,
+      "learning_rate": 1.9580161368739984e-05,
+      "loss": 0.3846,
+      "step": 455
+    },
+    {
+      "epoch": 0.13952077646345162,
+      "grad_norm": 1.0469274329601646,
+      "learning_rate": 1.956566598762388e-05,
+      "loss": 0.4229,
+      "step": 460
+    },
+    {
+      "epoch": 0.1410373066424022,
+      "grad_norm": 1.0010991478406723,
+      "learning_rate": 1.955093015180979e-05,
+      "loss": 0.3698,
+      "step": 465
+    },
+    {
+      "epoch": 0.14255383682135275,
+      "grad_norm": 1.315626156939816,
+      "learning_rate": 1.9535954231716334e-05,
+      "loss": 0.3568,
+      "step": 470
+    },
+    {
+      "epoch": 0.14407036700030332,
+      "grad_norm": 0.9897859436282127,
+      "learning_rate": 1.952073860379722e-05,
+      "loss": 0.3947,
+      "step": 475
+    },
+    {
+      "epoch": 0.14558689717925385,
+      "grad_norm": 1.0830619205943097,
+      "learning_rate": 1.950528365053174e-05,
+      "loss": 0.4243,
+      "step": 480
+    },
+    {
+      "epoch": 0.14710342735820442,
+      "grad_norm": 1.0427173223016224,
+      "learning_rate": 1.9489589760415186e-05,
+      "loss": 0.392,
+      "step": 485
+    },
+    {
+      "epoch": 0.14861995753715498,
+      "grad_norm": 0.8720070105985259,
+      "learning_rate": 1.9473657327949055e-05,
+      "loss": 0.3858,
+      "step": 490
+    },
+    {
+      "epoch": 0.15013648771610555,
+      "grad_norm": 1.0930260154511207,
+      "learning_rate": 1.9457486753631152e-05,
+      "loss": 0.4229,
+      "step": 495
+    },
+    {
+      "epoch": 0.1516530178950561,
+      "grad_norm": 1.0011322130603957,
+      "learning_rate": 1.9441078443945525e-05,
+      "loss": 0.3689,
+      "step": 500
+    },
+    {
+      "epoch": 0.1516530178950561,
+      "eval_loss": 0.4088059365749359,
+      "eval_runtime": 173.5642,
+      "eval_samples_per_second": 51.48,
+      "eval_steps_per_second": 25.743,
+      "step": 500
+    },
+    {
+      "epoch": 0.15316954807400668,
+      "grad_norm": 0.9905645606543211,
+      "learning_rate": 1.9424432811352224e-05,
+      "loss": 0.412,
+      "step": 505
+    },
+    {
+      "epoch": 0.15468607825295724,
+      "grad_norm": 1.3173648440740526,
+      "learning_rate": 1.940755027427696e-05,
+      "loss": 0.4601,
+      "step": 510
+    },
+    {
+      "epoch": 0.1562026084319078,
+      "grad_norm": 0.992923300776095,
+      "learning_rate": 1.939043125710057e-05,
+      "loss": 0.4151,
+      "step": 515
+    },
+    {
+      "epoch": 0.15771913861085834,
+      "grad_norm": 1.0345346572748841,
+      "learning_rate": 1.937307619014836e-05,
+      "loss": 0.4082,
+      "step": 520
+    },
+    {
+      "epoch": 0.1592356687898089,
+      "grad_norm": 1.0277113400586657,
+      "learning_rate": 1.9355485509679274e-05,
+      "loss": 0.3652,
+      "step": 525
+    },
+    {
+      "epoch": 0.16075219896875947,
+      "grad_norm": 1.1428063392505712,
+      "learning_rate": 1.9337659657874943e-05,
+      "loss": 0.4294,
+      "step": 530
+    },
+    {
+      "epoch": 0.16226872914771004,
+      "grad_norm": 0.818605782882043,
+      "learning_rate": 1.9319599082828554e-05,
+      "loss": 0.4165,
+      "step": 535
+    },
+    {
+      "epoch": 0.1637852593266606,
+      "grad_norm": 1.2266082158969924,
+      "learning_rate": 1.9301304238533608e-05,
+      "loss": 0.4088,
+      "step": 540
+    },
+    {
+      "epoch": 0.16530178950561117,
+      "grad_norm": 1.040073592097893,
+      "learning_rate": 1.9282775584872485e-05,
+      "loss": 0.4384,
+      "step": 545
+    },
+    {
+      "epoch": 0.16681831968456173,
+      "grad_norm": 1.0997516079813272,
+      "learning_rate": 1.926401358760489e-05,
+      "loss": 0.457,
+      "step": 550
+    },
+    {
+      "epoch": 0.16833484986351227,
+      "grad_norm": 1.2258344173220719,
+      "learning_rate": 1.924501871835616e-05,
+      "loss": 0.3966,
+      "step": 555
+    },
+    {
+      "epoch": 0.16985138004246284,
+      "grad_norm": 0.981904205499286,
+      "learning_rate": 1.9225791454605392e-05,
+      "loss": 0.3813,
+      "step": 560
+    },
+    {
+      "epoch": 0.1713679102214134,
+      "grad_norm": 1.084662004004182,
+      "learning_rate": 1.9206332279673437e-05,
+      "loss": 0.4593,
+      "step": 565
+    },
+    {
+      "epoch": 0.17288444040036396,
+      "grad_norm": 1.0398606480794417,
+      "learning_rate": 1.9186641682710774e-05,
+      "loss": 0.4093,
+      "step": 570
+    },
+    {
+      "epoch": 0.17440097057931453,
+      "grad_norm": 1.1672550355106228,
+      "learning_rate": 1.9166720158685187e-05,
+      "loss": 0.4052,
+      "step": 575
+    },
+    {
+      "epoch": 0.1759175007582651,
+      "grad_norm": 1.0087878638040124,
+      "learning_rate": 1.9146568208369346e-05,
+      "loss": 0.366,
+      "step": 580
+    },
+    {
+      "epoch": 0.17743403093721566,
+      "grad_norm": 1.2216691544147262,
+      "learning_rate": 1.91261863383282e-05,
+      "loss": 0.451,
+      "step": 585
+    },
+    {
+      "epoch": 0.17895056111616622,
+      "grad_norm": 1.3584786997202387,
+      "learning_rate": 1.9105575060906254e-05,
+      "loss": 0.436,
+      "step": 590
+    },
+    {
+      "epoch": 0.18046709129511676,
+      "grad_norm": 0.9189515390724051,
+      "learning_rate": 1.908473489421468e-05,
+      "loss": 0.3632,
+      "step": 595
+    },
+    {
+      "epoch": 0.18198362147406733,
+      "grad_norm": 1.0321679875275023,
+      "learning_rate": 1.9063666362118324e-05,
+      "loss": 0.3914,
+      "step": 600
+    },
+    {
+      "epoch": 0.1835001516530179,
+      "grad_norm": 1.039032214603117,
+      "learning_rate": 1.9042369994222487e-05,
+      "loss": 0.3524,
+      "step": 605
+    },
+    {
+      "epoch": 0.18501668183196845,
+      "grad_norm": 0.9559021915379782,
+      "learning_rate": 1.902084632585965e-05,
+      "loss": 0.436,
+      "step": 610
+    },
+    {
+      "epoch": 0.18653321201091902,
+      "grad_norm": 1.0082278999199734,
+      "learning_rate": 1.8999095898076012e-05,
+      "loss": 0.3322,
+      "step": 615
+    },
+    {
+      "epoch": 0.18804974218986958,
+      "grad_norm": 1.1936896778472303,
+      "learning_rate": 1.8977119257617878e-05,
+      "loss": 0.3529,
+      "step": 620
+    },
+    {
+      "epoch": 0.18956627236882015,
+      "grad_norm": 0.9718605188148373,
+      "learning_rate": 1.8954916956917922e-05,
+      "loss": 0.3966,
+      "step": 625
+    },
+    {
+      "epoch": 0.1910828025477707,
+      "grad_norm": 1.0441882225736692,
+      "learning_rate": 1.8932489554081295e-05,
+      "loss": 0.3683,
+      "step": 630
+    },
+    {
+      "epoch": 0.19259933272672125,
+      "grad_norm": 1.0302242941912132,
+      "learning_rate": 1.8909837612871615e-05,
+      "loss": 0.4015,
+      "step": 635
+    },
+    {
+      "epoch": 0.19411586290567182,
+      "grad_norm": 1.0668469700274468,
+      "learning_rate": 1.8886961702696765e-05,
+      "loss": 0.3682,
+      "step": 640
+    },
+    {
+      "epoch": 0.19563239308462238,
+      "grad_norm": 0.9494238968317048,
+      "learning_rate": 1.8863862398594606e-05,
+      "loss": 0.4081,
+      "step": 645
+    },
+    {
+      "epoch": 0.19714892326357294,
+      "grad_norm": 1.1243612355939872,
+      "learning_rate": 1.8840540281218506e-05,
+      "loss": 0.3883,
+      "step": 650
+    },
+    {
+      "epoch": 0.1986654534425235,
+      "grad_norm": 1.075535429789229,
+      "learning_rate": 1.881699593682275e-05,
+      "loss": 0.3991,
+      "step": 655
+    },
+    {
+      "epoch": 0.20018198362147407,
+      "grad_norm": 1.1424504644700166,
+      "learning_rate": 1.8793229957247808e-05,
+      "loss": 0.4006,
+      "step": 660
+    },
+    {
+      "epoch": 0.20169851380042464,
+      "grad_norm": 1.20832013347783,
+      "learning_rate": 1.8769242939905446e-05,
+      "loss": 0.3781,
+      "step": 665
+    },
+    {
+      "epoch": 0.20321504397937518,
+      "grad_norm": 0.9949915408461836,
+      "learning_rate": 1.874503548776372e-05,
+      "loss": 0.4073,
+      "step": 670
+    },
+    {
+      "epoch": 0.20473157415832574,
+      "grad_norm": 1.0184539078882147,
+      "learning_rate": 1.8720608209331813e-05,
+      "loss": 0.404,
+      "step": 675
+    },
+    {
+      "epoch": 0.2062481043372763,
+      "grad_norm": 1.0772016502245503,
+      "learning_rate": 1.8695961718644743e-05,
+      "loss": 0.4306,
+      "step": 680
+    },
+    {
+      "epoch": 0.20776463451622687,
+      "grad_norm": 1.0757591761900838,
+      "learning_rate": 1.8671096635247914e-05,
+      "loss": 0.3891,
+      "step": 685
+    },
+    {
+      "epoch": 0.20928116469517744,
+      "grad_norm": 1.134500144137449,
+      "learning_rate": 1.864601358418157e-05,
+      "loss": 0.3858,
+      "step": 690
+    },
+    {
+      "epoch": 0.210797694874128,
+      "grad_norm": 1.3231464100942203,
+      "learning_rate": 1.8620713195965052e-05,
+      "loss": 0.3896,
+      "step": 695
+    },
+    {
+      "epoch": 0.21231422505307856,
+      "grad_norm": 0.9378926668917021,
+      "learning_rate": 1.8595196106580973e-05,
+      "loss": 0.372,
+      "step": 700
+    },
+    {
+      "epoch": 0.21383075523202913,
+      "grad_norm": 1.156227424640605,
+      "learning_rate": 1.856946295745921e-05,
+      "loss": 0.4282,
+      "step": 705
+    },
+    {
+      "epoch": 0.21534728541097967,
+      "grad_norm": 1.2036566060604865,
+      "learning_rate": 1.8543514395460806e-05,
+      "loss": 0.4055,
+      "step": 710
+    },
+    {
+      "epoch": 0.21686381558993023,
+      "grad_norm": 1.0835648170732943,
+      "learning_rate": 1.8517351072861682e-05,
+      "loss": 0.3864,
+      "step": 715
+    },
+    {
+      "epoch": 0.2183803457688808,
+      "grad_norm": 1.0077498655883261,
+      "learning_rate": 1.8490973647336255e-05,
+      "loss": 0.4268,
+      "step": 720
+    },
+    {
+      "epoch": 0.21989687594783136,
+      "grad_norm": 0.9546564688055839,
+      "learning_rate": 1.8464382781940918e-05,
+      "loss": 0.3632,
+      "step": 725
+    },
+    {
+      "epoch": 0.22141340612678193,
+      "grad_norm": 0.8840597627792552,
+      "learning_rate": 1.843757914509734e-05,
+      "loss": 0.38,
+      "step": 730
+    },
+    {
+      "epoch": 0.2229299363057325,
+      "grad_norm": 1.158885040155564,
+      "learning_rate": 1.8410563410575696e-05,
+      "loss": 0.3923,
+      "step": 735
+    },
+    {
+      "epoch": 0.22444646648468305,
+      "grad_norm": 0.9247547809922593,
+      "learning_rate": 1.838333625747771e-05,
+      "loss": 0.3865,
+      "step": 740
+    },
+    {
+      "epoch": 0.22596299666363362,
+      "grad_norm": 1.0016406439137249,
+      "learning_rate": 1.835589837021959e-05,
+      "loss": 0.3776,
+      "step": 745
+    },
+    {
+      "epoch": 0.22747952684258416,
+      "grad_norm": 1.043626508615065,
+      "learning_rate": 1.8328250438514837e-05,
+      "loss": 0.3608,
+      "step": 750
+    },
+    {
+      "epoch": 0.22899605702153472,
+      "grad_norm": 0.9577907090027657,
+      "learning_rate": 1.830039315735688e-05,
+      "loss": 0.3894,
+      "step": 755
+    },
+    {
+      "epoch": 0.23051258720048529,
+      "grad_norm": 0.9580721684764322,
+      "learning_rate": 1.827232722700163e-05,
+      "loss": 0.3614,
+      "step": 760
+    },
+    {
+      "epoch": 0.23202911737943585,
+      "grad_norm": 0.9477520019373953,
+      "learning_rate": 1.8244053352949866e-05,
+      "loss": 0.3582,
+      "step": 765
+    },
+    {
+      "epoch": 0.23354564755838642,
+      "grad_norm": 0.9783981899125124,
+      "learning_rate": 1.82155722459295e-05,
+      "loss": 0.3686,
+      "step": 770
+    },
+    {
+      "epoch": 0.23506217773733698,
+      "grad_norm": 1.2531749551765516,
+      "learning_rate": 1.8186884621877726e-05,
+      "loss": 0.435,
+      "step": 775
+    },
+    {
+      "epoch": 0.23657870791628755,
+      "grad_norm": 1.0378676377391638,
+      "learning_rate": 1.815799120192299e-05,
+      "loss": 0.3919,
+      "step": 780
+    },
+    {
+      "epoch": 0.23809523809523808,
+      "grad_norm": 0.9221216623517181,
+      "learning_rate": 1.8128892712366916e-05,
+      "loss": 0.3703,
+      "step": 785
+    },
+    {
+      "epoch": 0.23961176827418865,
+      "grad_norm": 0.9471155975175013,
+      "learning_rate": 1.8099589884665986e-05,
+      "loss": 0.4248,
+      "step": 790
+    },
+    {
+      "epoch": 0.2411282984531392,
+      "grad_norm": 1.0217123358982474,
+      "learning_rate": 1.80700834554132e-05,
+      "loss": 0.3706,
+      "step": 795
+    },
+    {
+      "epoch": 0.24264482863208978,
+      "grad_norm": 1.1240824857344656,
+      "learning_rate": 1.804037416631954e-05,
+      "loss": 0.381,
+      "step": 800
+    },
+    {
+      "epoch": 0.24416135881104034,
+      "grad_norm": 0.9053272736194982,
+      "learning_rate": 1.801046276419534e-05,
+      "loss": 0.3612,
+      "step": 805
+    },
+    {
+      "epoch": 0.2456778889899909,
+      "grad_norm": 1.1627818653327249,
+      "learning_rate": 1.7980350000931494e-05,
+      "loss": 0.3961,
+      "step": 810
+    },
+    {
+      "epoch": 0.24719441916894147,
+      "grad_norm": 0.9805469968339321,
+      "learning_rate": 1.7950036633480557e-05,
+      "loss": 0.3902,
+      "step": 815
+    },
+    {
+      "epoch": 0.24871094934789204,
+      "grad_norm": 1.1280037352121748,
+      "learning_rate": 1.7919523423837743e-05,
+      "loss": 0.3845,
+      "step": 820
+    },
+    {
+      "epoch": 0.2502274795268426,
+      "grad_norm": 0.9614220888721607,
+      "learning_rate": 1.788881113902174e-05,
+      "loss": 0.3896,
+      "step": 825
+    },
+    {
+      "epoch": 0.25174400970579314,
+      "grad_norm": 0.9351755310616024,
+      "learning_rate": 1.7857900551055448e-05,
+      "loss": 0.39,
+      "step": 830
+    },
+    {
+      "epoch": 0.25326053988474373,
+      "grad_norm": 1.1017280153749585,
+      "learning_rate": 1.7826792436946562e-05,
+      "loss": 0.377,
+      "step": 835
+    },
+    {
+      "epoch": 0.25477707006369427,
+      "grad_norm": 0.9933284106135498,
+      "learning_rate": 1.779548757866804e-05,
+      "loss": 0.4235,
+      "step": 840
+    },
+    {
+      "epoch": 0.2562936002426448,
+      "grad_norm": 1.0616915093996913,
+      "learning_rate": 1.7763986763138467e-05,
+      "loss": 0.3812,
+      "step": 845
+    },
+    {
+      "epoch": 0.2578101304215954,
+      "grad_norm": 1.1439917497224341,
+      "learning_rate": 1.7732290782202244e-05,
+      "loss": 0.4254,
+      "step": 850
+    },
+    {
+      "epoch": 0.25932666060054593,
+      "grad_norm": 0.8582079212321349,
+      "learning_rate": 1.7700400432609695e-05,
+      "loss": 0.3699,
+      "step": 855
+    },
+    {
+      "epoch": 0.2608431907794965,
+      "grad_norm": 1.2134674630319888,
+      "learning_rate": 1.7668316515997047e-05,
+      "loss": 0.4338,
+      "step": 860
+    },
+    {
+      "epoch": 0.26235972095844706,
+      "grad_norm": 1.163884571927593,
+      "learning_rate": 1.7636039838866278e-05,
+      "loss": 0.3866,
+      "step": 865
+    },
+    {
+      "epoch": 0.26387625113739765,
+      "grad_norm": 1.0199336931679066,
+      "learning_rate": 1.760357121256482e-05,
+      "loss": 0.4029,
+      "step": 870
+    },
+    {
+      "epoch": 0.2653927813163482,
+      "grad_norm": 0.946151313424608,
+      "learning_rate": 1.757091145326521e-05,
+      "loss": 0.3595,
+      "step": 875
+    },
+    {
+      "epoch": 0.26690931149529873,
+      "grad_norm": 1.1932767975805696,
+      "learning_rate": 1.7538061381944524e-05,
+      "loss": 0.43,
+      "step": 880
+    },
+    {
+      "epoch": 0.2684258416742493,
+      "grad_norm": 0.9649699002383715,
+      "learning_rate": 1.7505021824363767e-05,
+      "loss": 0.3986,
+      "step": 885
+    },
+    {
+      "epoch": 0.26994237185319986,
+      "grad_norm": 0.9468284901526932,
+      "learning_rate": 1.7471793611047114e-05,
+      "loss": 0.3551,
+      "step": 890
+    },
+    {
+      "epoch": 0.27145890203215045,
+      "grad_norm": 1.026534302980973,
+      "learning_rate": 1.743837757726103e-05,
+      "loss": 0.3821,
+      "step": 895
+    },
+    {
+      "epoch": 0.272975432211101,
+      "grad_norm": 0.9999149595040162,
+      "learning_rate": 1.7404774562993268e-05,
+      "loss": 0.3755,
+      "step": 900
+    },
+    {
+      "epoch": 0.2744919623900516,
+      "grad_norm": 0.9887626409688292,
+      "learning_rate": 1.7370985412931766e-05,
+      "loss": 0.3929,
+      "step": 905
+    },
+    {
+      "epoch": 0.2760084925690021,
+      "grad_norm": 1.1520336639240936,
+      "learning_rate": 1.7337010976443404e-05,
+      "loss": 0.416,
+      "step": 910
+    },
+    {
+      "epoch": 0.2775250227479527,
+      "grad_norm": 1.2647475168427802,
+      "learning_rate": 1.730285210755265e-05,
+      "loss": 0.4117,
+      "step": 915
+    },
+    {
+      "epoch": 0.27904155292690325,
+      "grad_norm": 1.1162345721181772,
+      "learning_rate": 1.7268509664920115e-05,
+      "loss": 0.4133,
+      "step": 920
+    },
+    {
+      "epoch": 0.2805580831058538,
+      "grad_norm": 1.1083917949237179,
+      "learning_rate": 1.7233984511820937e-05,
+      "loss": 0.3991,
+      "step": 925
+    },
+    {
+      "epoch": 0.2820746132848044,
+      "grad_norm": 1.1415307366312324,
+      "learning_rate": 1.7199277516123098e-05,
+      "loss": 0.4018,
+      "step": 930
+    },
+    {
+      "epoch": 0.2835911434637549,
+      "grad_norm": 0.9966551314508225,
+      "learning_rate": 1.7164389550265607e-05,
+      "loss": 0.3764,
+      "step": 935
+    },
+    {
+      "epoch": 0.2851076736427055,
+      "grad_norm": 1.1395841818497014,
+      "learning_rate": 1.7129321491236578e-05,
+      "loss": 0.4094,
+      "step": 940
+    },
+    {
+      "epoch": 0.28662420382165604,
+      "grad_norm": 0.9290839897008805,
+      "learning_rate": 1.709407422055116e-05,
+      "loss": 0.3757,
+      "step": 945
+    },
+    {
+      "epoch": 0.28814073400060664,
+      "grad_norm": 0.9613783627220345,
+      "learning_rate": 1.70586486242294e-05,
+      "loss": 0.4067,
+      "step": 950
+    },
+    {
+      "epoch": 0.2896572641795572,
+      "grad_norm": 0.9710982407207631,
+      "learning_rate": 1.7023045592773968e-05,
+      "loss": 0.3521,
+      "step": 955
+    },
+    {
+      "epoch": 0.2911737943585077,
+      "grad_norm": 1.095728541816799,
+      "learning_rate": 1.6987266021147763e-05,
+      "loss": 0.3749,
+      "step": 960
+    },
+    {
+      "epoch": 0.2926903245374583,
+      "grad_norm": 1.0605469280918307,
+      "learning_rate": 1.695131080875142e-05,
+      "loss": 0.3562,
+      "step": 965
+    },
+    {
+      "epoch": 0.29420685471640884,
+      "grad_norm": 0.8408891042454621,
+      "learning_rate": 1.691518085940071e-05,
+      "loss": 0.3377,
+      "step": 970
+    },
+    {
+      "epoch": 0.29572338489535943,
+      "grad_norm": 0.9467466462817451,
+      "learning_rate": 1.6878877081303805e-05,
+      "loss": 0.3498,
+      "step": 975
+    },
+    {
+      "epoch": 0.29723991507430997,
+      "grad_norm": 1.1449510440698407,
+      "learning_rate": 1.6842400387038464e-05,
+      "loss": 0.3524,
+      "step": 980
+    },
+    {
+      "epoch": 0.29875644525326056,
+      "grad_norm": 0.948414328408764,
+      "learning_rate": 1.6805751693529083e-05,
+      "loss": 0.3784,
+      "step": 985
+    },
+    {
+      "epoch": 0.3002729754322111,
+      "grad_norm": 1.0675528322798367,
+      "learning_rate": 1.676893192202364e-05,
+      "loss": 0.3182,
+      "step": 990
+    },
+    {
+      "epoch": 0.30178950561116163,
+      "grad_norm": 1.129087718059458,
+      "learning_rate": 1.673194199807057e-05,
+      "loss": 0.3825,
+      "step": 995
+    },
+    {
+      "epoch": 0.3033060357901122,
+      "grad_norm": 1.0937265212418354,
+      "learning_rate": 1.6694782851495444e-05,
+      "loss": 0.3919,
+      "step": 1000
+    },
+    {
+      "epoch": 0.3033060357901122,
+      "eval_loss": 0.3952370584011078,
+      "eval_runtime": 181.3096,
+      "eval_samples_per_second": 49.28,
+      "eval_steps_per_second": 24.643,
+      "step": 1000
+    },
+    {
+      "epoch": 0.30482256596906276,
+      "grad_norm": 1.1557787421563186,
+      "learning_rate": 1.6657455416377654e-05,
+      "loss": 0.3941,
+      "step": 1005
+    },
+    {
+      "epoch": 0.30633909614801336,
+      "grad_norm": 1.1101059403333342,
+      "learning_rate": 1.661996063102689e-05,
+      "loss": 0.4257,
+      "step": 1010
+    },
+    {
+      "epoch": 0.3078556263269639,
+      "grad_norm": 0.8880119467039661,
+      "learning_rate": 1.6582299437959577e-05,
+      "loss": 0.4088,
+      "step": 1015
+    },
+    {
+      "epoch": 0.3093721565059145,
+      "grad_norm": 0.9624888079810805,
+      "learning_rate": 1.6544472783875173e-05,
+      "loss": 0.3875,
+      "step": 1020
+    },
+    {
+      "epoch": 0.310888686684865,
+      "grad_norm": 0.9935490083526604,
+      "learning_rate": 1.650648161963237e-05,
+      "loss": 0.3899,
+      "step": 1025
+    },
+    {
+      "epoch": 0.3124052168638156,
+      "grad_norm": 1.1083315990997267,
+      "learning_rate": 1.6468326900225204e-05,
+      "loss": 0.3855,
+      "step": 1030
+    },
+    {
+      "epoch": 0.31392174704276615,
+      "grad_norm": 0.9176932289492212,
+      "learning_rate": 1.6430009584759036e-05,
+      "loss": 0.3805,
+      "step": 1035
+    },
+    {
+      "epoch": 0.3154382772217167,
+      "grad_norm": 1.159934652109041,
+      "learning_rate": 1.6391530636426447e-05,
+      "loss": 0.4393,
+      "step": 1040
+    },
+    {
+      "epoch": 0.3169548074006673,
+      "grad_norm": 0.9549890829191309,
+      "learning_rate": 1.6352891022483025e-05,
+      "loss": 0.3853,
+      "step": 1045
+    },
+    {
+      "epoch": 0.3184713375796178,
+      "grad_norm": 0.8753544532645915,
+      "learning_rate": 1.631409171422306e-05,
+      "loss": 0.3783,
+      "step": 1050
+    },
+    {
+      "epoch": 0.3199878677585684,
+      "grad_norm": 1.145468835413513,
+      "learning_rate": 1.6275133686955107e-05,
+      "loss": 0.4251,
+      "step": 1055
+    },
+    {
+      "epoch": 0.32150439793751895,
+      "grad_norm": 1.0318505639013171,
+      "learning_rate": 1.6236017919977495e-05,
+      "loss": 0.4074,
+      "step": 1060
+    },
+    {
+      "epoch": 0.32302092811646954,
+      "grad_norm": 0.9665743048952797,
+      "learning_rate": 1.61967453965537e-05,
+      "loss": 0.3765,
+      "step": 1065
+    },
+    {
+      "epoch": 0.3245374582954201,
+      "grad_norm": 0.9170197925688426,
+      "learning_rate": 1.615731710388761e-05,
+      "loss": 0.3538,
+      "step": 1070
+    },
+    {
+      "epoch": 0.3260539884743706,
+      "grad_norm": 1.3931816470935723,
+      "learning_rate": 1.6117734033098744e-05,
+      "loss": 0.4235,
+      "step": 1075
+    },
+    {
+      "epoch": 0.3275705186533212,
+      "grad_norm": 1.07985655044954,
+      "learning_rate": 1.6077997179197314e-05,
+      "loss": 0.4324,
+      "step": 1080
+    },
+    {
+      "epoch": 0.32908704883227174,
+      "grad_norm": 0.8912377259192088,
+      "learning_rate": 1.6038107541059216e-05,
+      "loss": 0.3746,
+      "step": 1085
+    },
+    {
+      "epoch": 0.33060357901122234,
+      "grad_norm": 0.9135076769501183,
+      "learning_rate": 1.5998066121400925e-05,
+      "loss": 0.3453,
+      "step": 1090
+    },
+    {
+      "epoch": 0.3321201091901729,
+      "grad_norm": 1.0231218093240508,
+      "learning_rate": 1.5957873926754294e-05,
+      "loss": 0.3952,
+      "step": 1095
+    },
+    {
+      "epoch": 0.33363663936912347,
+      "grad_norm": 1.128624243072491,
+      "learning_rate": 1.5917531967441235e-05,
+      "loss": 0.3548,
+      "step": 1100
+    },
+    {
+      "epoch": 0.335153169548074,
+      "grad_norm": 0.8957025806927242,
+      "learning_rate": 1.587704125754835e-05,
+      "loss": 0.3723,
+      "step": 1105
+    },
+    {
+      "epoch": 0.33666969972702454,
+      "grad_norm": 1.1269942168579994,
+      "learning_rate": 1.583640281490141e-05,
+      "loss": 0.4201,
+      "step": 1110
+    },
+    {
+      "epoch": 0.33818622990597513,
+      "grad_norm": 1.133173361691646,
+      "learning_rate": 1.5795617661039794e-05,
+      "loss": 0.4232,
+      "step": 1115
+    },
+    {
+      "epoch": 0.33970276008492567,
+      "grad_norm": 0.8611310832137484,
+      "learning_rate": 1.5754686821190797e-05,
+      "loss": 0.3969,
+      "step": 1120
+    },
+    {
+      "epoch": 0.34121929026387626,
+      "grad_norm": 0.8942802074964451,
+      "learning_rate": 1.5713611324243858e-05,
+      "loss": 0.3453,
+      "step": 1125
+    },
+    {
+      "epoch": 0.3427358204428268,
+      "grad_norm": 1.0128940099871961,
+      "learning_rate": 1.5672392202724702e-05,
+      "loss": 0.41,
+      "step": 1130
+    },
+    {
+      "epoch": 0.3442523506217774,
+      "grad_norm": 0.8082512004172621,
+      "learning_rate": 1.5631030492769385e-05,
+      "loss": 0.3404,
+      "step": 1135
+    },
+    {
+      "epoch": 0.34576888080072793,
+      "grad_norm": 0.9097763994809697,
+      "learning_rate": 1.5589527234098247e-05,
+      "loss": 0.3542,
+      "step": 1140
+    },
+    {
+      "epoch": 0.3472854109796785,
+      "grad_norm": 0.8970233702221374,
+      "learning_rate": 1.5547883469989767e-05,
+      "loss": 0.3783,
+      "step": 1145
+    },
+    {
+      "epoch": 0.34880194115862906,
+      "grad_norm": 1.0373858768564683,
+      "learning_rate": 1.5506100247254363e-05,
+      "loss": 0.3424,
+      "step": 1150
+    },
+    {
+      "epoch": 0.3503184713375796,
+      "grad_norm": 0.9339798828268197,
+      "learning_rate": 1.5464178616208046e-05,
+      "loss": 0.4106,
+      "step": 1155
+    },
+    {
+      "epoch": 0.3518350015165302,
+      "grad_norm": 0.9599549402427533,
+      "learning_rate": 1.5422119630646043e-05,
+      "loss": 0.35,
+      "step": 1160
+    },
+    {
+      "epoch": 0.3533515316954807,
+      "grad_norm": 1.0567048479540737,
+      "learning_rate": 1.5379924347816296e-05,
+      "loss": 0.3996,
+      "step": 1165
+    },
+    {
+      "epoch": 0.3548680618744313,
+      "grad_norm": 1.1009203123989402,
+      "learning_rate": 1.533759382839288e-05,
+      "loss": 0.4392,
+      "step": 1170
+    },
+    {
+      "epoch": 0.35638459205338185,
+      "grad_norm": 0.9592441175249292,
+      "learning_rate": 1.5295129136449362e-05,
+      "loss": 0.3941,
+      "step": 1175
+    },
+    {
+      "epoch": 0.35790112223233245,
+      "grad_norm": 0.974862722178181,
+      "learning_rate": 1.5252531339432033e-05,
+      "loss": 0.3848,
+      "step": 1180
+    },
+    {
+      "epoch": 0.359417652411283,
+      "grad_norm": 0.800718600370757,
+      "learning_rate": 1.5209801508133077e-05,
+      "loss": 0.3603,
+      "step": 1185
+    },
+    {
+      "epoch": 0.3609341825902335,
+      "grad_norm": 1.044944939187582,
+      "learning_rate": 1.516694071666367e-05,
+      "loss": 0.4112,
+      "step": 1190
+    },
+    {
+      "epoch": 0.3624507127691841,
+      "grad_norm": 0.9861578107971376,
+      "learning_rate": 1.5123950042426958e-05,
+      "loss": 0.4303,
+      "step": 1195
+    },
+    {
+      "epoch": 0.36396724294813465,
+      "grad_norm": 0.9236514908326175,
+      "learning_rate": 1.5080830566090986e-05,
+      "loss": 0.3705,
+      "step": 1200
+    },
+    {
+      "epoch": 0.36548377312708524,
+      "grad_norm": 1.0547458629444628,
+      "learning_rate": 1.5037583371561538e-05,
+      "loss": 0.3558,
+      "step": 1205
+    },
+    {
+      "epoch": 0.3670003033060358,
+      "grad_norm": 1.0149155289164045,
+      "learning_rate": 1.4994209545954884e-05,
+      "loss": 0.3455,
+      "step": 1210
+    },
+    {
+      "epoch": 0.3685168334849864,
+      "grad_norm": 0.975465418286225,
+      "learning_rate": 1.4950710179570442e-05,
+      "loss": 0.3852,
+      "step": 1215
+    },
+    {
+      "epoch": 0.3700333636639369,
+      "grad_norm": 1.0918284001664949,
+      "learning_rate": 1.49070863658634e-05,
+      "loss": 0.4167,
+      "step": 1220
+    },
+    {
+      "epoch": 0.37154989384288745,
+      "grad_norm": 0.963559934243437,
+      "learning_rate": 1.4863339201417195e-05,
+      "loss": 0.4196,
+      "step": 1225
+    },
+    {
+      "epoch": 0.37306642402183804,
+      "grad_norm": 0.8857039743407772,
+      "learning_rate": 1.4819469785915972e-05,
+      "loss": 0.3796,
+      "step": 1230
+    },
+    {
+      "epoch": 0.3745829542007886,
+      "grad_norm": 0.8586023962029455,
+      "learning_rate": 1.4775479222116935e-05,
+      "loss": 0.4458,
+      "step": 1235
+    },
+    {
+      "epoch": 0.37609948437973917,
+      "grad_norm": 1.0620266861330567,
+      "learning_rate": 1.4731368615822623e-05,
+      "loss": 0.3992,
+      "step": 1240
+    },
+    {
+      "epoch": 0.3776160145586897,
+      "grad_norm": 0.9686222054309538,
+      "learning_rate": 1.468713907585311e-05,
+      "loss": 0.3897,
+      "step": 1245
+    },
+    {
+      "epoch": 0.3791325447376403,
+      "grad_norm": 0.9098741177020511,
+      "learning_rate": 1.4642791714018148e-05,
+      "loss": 0.3975,
+      "step": 1250
+    },
+    {
+      "epoch": 0.38064907491659083,
+      "grad_norm": 1.0748028384120583,
+      "learning_rate": 1.45983276450892e-05,
+      "loss": 0.3719,
+      "step": 1255
+    },
+    {
+      "epoch": 0.3821656050955414,
+      "grad_norm": 1.0534448032219224,
+      "learning_rate": 1.4553747986771426e-05,
+      "loss": 0.428,
+      "step": 1260
+    },
+    {
+      "epoch": 0.38368213527449196,
+      "grad_norm": 1.1494681374152054,
+      "learning_rate": 1.4509053859675601e-05,
+      "loss": 0.43,
+      "step": 1265
+    },
+    {
+      "epoch": 0.3851986654534425,
+      "grad_norm": 0.9573470062232139,
+      "learning_rate": 1.4464246387289913e-05,
+      "loss": 0.412,
+      "step": 1270
+    },
+    {
+      "epoch": 0.3867151956323931,
+      "grad_norm": 1.3662334486068808,
+      "learning_rate": 1.4419326695951752e-05,
+      "loss": 0.3896,
+      "step": 1275
+    },
+    {
+      "epoch": 0.38823172581134363,
+      "grad_norm": 1.0831328903659596,
+      "learning_rate": 1.4374295914819385e-05,
+      "loss": 0.3853,
+      "step": 1280
+    },
+    {
+      "epoch": 0.3897482559902942,
+      "grad_norm": 1.1259269630812838,
+      "learning_rate": 1.4329155175843572e-05,
+      "loss": 0.3424,
+      "step": 1285
+    },
+    {
+      "epoch": 0.39126478616924476,
+      "grad_norm": 1.0133166607814084,
+      "learning_rate": 1.4283905613739107e-05,
+      "loss": 0.3958,
+      "step": 1290
+    },
+    {
+      "epoch": 0.39278131634819535,
+      "grad_norm": 0.9526285779039391,
+      "learning_rate": 1.4238548365956308e-05,
+      "loss": 0.3465,
+      "step": 1295
+    },
+    {
+      "epoch": 0.3942978465271459,
+      "grad_norm": 0.8356690841228143,
+      "learning_rate": 1.4193084572652415e-05,
+      "loss": 0.3676,
+      "step": 1300
+    },
+    {
+      "epoch": 0.3958143767060964,
+      "grad_norm": 1.056339975220587,
+      "learning_rate": 1.4147515376662928e-05,
+      "loss": 0.3866,
+      "step": 1305
+    },
+    {
+      "epoch": 0.397330906885047,
+      "grad_norm": 1.093716550361687,
+      "learning_rate": 1.4101841923472885e-05,
+      "loss": 0.3623,
+      "step": 1310
+    },
+    {
+      "epoch": 0.39884743706399756,
+      "grad_norm": 1.1381025026138512,
+      "learning_rate": 1.4056065361188068e-05,
+      "loss": 0.3829,
+      "step": 1315
+    },
+    {
+      "epoch": 0.40036396724294815,
+      "grad_norm": 1.0195350418123983,
+      "learning_rate": 1.4010186840506123e-05,
+      "loss": 0.3487,
+      "step": 1320
+    },
+    {
+      "epoch": 0.4018804974218987,
+      "grad_norm": 0.8852855193645849,
+      "learning_rate": 1.396420751468768e-05,
+      "loss": 0.3533,
+      "step": 1325
+    },
+    {
+      "epoch": 0.4033970276008493,
+      "grad_norm": 0.993588466372261,
+      "learning_rate": 1.3918128539527312e-05,
+      "loss": 0.4471,
+      "step": 1330
+    },
+    {
+      "epoch": 0.4049135577797998,
+      "grad_norm": 0.9519991181443925,
+      "learning_rate": 1.3871951073324508e-05,
+      "loss": 0.3521,
+      "step": 1335
+    },
+    {
+      "epoch": 0.40643008795875035,
+      "grad_norm": 0.9493052976320027,
+      "learning_rate": 1.3825676276854563e-05,
+      "loss": 0.3804,
+      "step": 1340
+    },
+    {
+      "epoch": 0.40794661813770094,
+      "grad_norm": 0.8469034564877794,
+      "learning_rate": 1.377930531333938e-05,
+      "loss": 0.4021,
+      "step": 1345
+    },
+    {
+      "epoch": 0.4094631483166515,
+      "grad_norm": 1.1005168381310284,
+      "learning_rate": 1.3732839348418234e-05,
+      "loss": 0.328,
+      "step": 1350
+    },
+    {
+      "epoch": 0.4109796784956021,
+      "grad_norm": 1.0684696716820943,
+      "learning_rate": 1.3686279550118491e-05,
+      "loss": 0.377,
+      "step": 1355
+    },
+    {
+      "epoch": 0.4124962086745526,
+      "grad_norm": 1.0525285102148023,
+      "learning_rate": 1.3639627088826217e-05,
+      "loss": 0.3807,
+      "step": 1360
+    },
+    {
+      "epoch": 0.4140127388535032,
+      "grad_norm": 0.9870824418498249,
+      "learning_rate": 1.3592883137256776e-05,
+      "loss": 0.3821,
+      "step": 1365
+    },
+    {
+      "epoch": 0.41552926903245374,
+      "grad_norm": 0.7608715586792121,
+      "learning_rate": 1.3546048870425356e-05,
+      "loss": 0.3404,
+      "step": 1370
+    },
+    {
+      "epoch": 0.41704579921140433,
+      "grad_norm": 0.99595610767164,
+      "learning_rate": 1.3499125465617417e-05,
+      "loss": 0.3872,
+      "step": 1375
+    },
+    {
+      "epoch": 0.41856232939035487,
+      "grad_norm": 1.000735855327813,
+      "learning_rate": 1.34521141023591e-05,
+      "loss": 0.3428,
+      "step": 1380
+    },
+    {
+      "epoch": 0.4200788595693054,
+      "grad_norm": 0.9679808555916511,
+      "learning_rate": 1.3405015962387588e-05,
+      "loss": 0.3543,
+      "step": 1385
+    },
+    {
+      "epoch": 0.421595389748256,
+      "grad_norm": 0.9241312294010786,
+      "learning_rate": 1.3357832229621393e-05,
+      "loss": 0.4196,
+      "step": 1390
+    },
+    {
+      "epoch": 0.42311191992720654,
+      "grad_norm": 0.9695965917058144,
+      "learning_rate": 1.3310564090130588e-05,
+      "loss": 0.4028,
+      "step": 1395
+    },
+    {
+      "epoch": 0.42462845010615713,
+      "grad_norm": 0.9665874167024507,
+      "learning_rate": 1.3263212732107014e-05,
+      "loss": 0.4332,
+      "step": 1400
+    },
+    {
+      "epoch": 0.42614498028510767,
+      "grad_norm": 1.1010539769239436,
+      "learning_rate": 1.3215779345834385e-05,
+      "loss": 0.3641,
+      "step": 1405
+    },
+    {
+      "epoch": 0.42766151046405826,
+      "grad_norm": 0.738968771424043,
+      "learning_rate": 1.3168265123658386e-05,
+      "loss": 0.324,
+      "step": 1410
+    },
+    {
+      "epoch": 0.4291780406430088,
+      "grad_norm": 0.8453745632530565,
+      "learning_rate": 1.3120671259956699e-05,
+      "loss": 0.3628,
+      "step": 1415
+    },
+    {
+      "epoch": 0.43069457082195933,
+      "grad_norm": 1.0651118171429264,
+      "learning_rate": 1.3072998951108978e-05,
+      "loss": 0.3874,
+      "step": 1420
+    },
+    {
+      "epoch": 0.4322111010009099,
+      "grad_norm": 0.8241128807768539,
+      "learning_rate": 1.3025249395466758e-05,
+      "loss": 0.3672,
+      "step": 1425
+    },
+    {
+      "epoch": 0.43372763117986046,
+      "grad_norm": 0.8764574352955143,
+      "learning_rate": 1.297742379332337e-05,
+      "loss": 0.3784,
+      "step": 1430
+    },
+    {
+      "epoch": 0.43524416135881105,
+      "grad_norm": 0.9127221397188253,
+      "learning_rate": 1.292952334688373e-05,
+      "loss": 0.3538,
+      "step": 1435
+    },
+    {
+      "epoch": 0.4367606915377616,
+      "grad_norm": 0.9788821462735778,
+      "learning_rate": 1.2881549260234137e-05,
+      "loss": 0.3817,
+      "step": 1440
+    },
+    {
+      "epoch": 0.4382772217167122,
+      "grad_norm": 0.9336141701850212,
+      "learning_rate": 1.2833502739312009e-05,
+      "loss": 0.3911,
+      "step": 1445
+    },
+    {
+      "epoch": 0.4397937518956627,
+      "grad_norm": 1.1869329918853797,
+      "learning_rate": 1.2785384991875565e-05,
+      "loss": 0.3865,
+      "step": 1450
+    },
+    {
+      "epoch": 0.44131028207461326,
+      "grad_norm": 0.8526421566105249,
+      "learning_rate": 1.273719722747345e-05,
+      "loss": 0.3398,
+      "step": 1455
+    },
+    {
+      "epoch": 0.44282681225356385,
+      "grad_norm": 0.8605140943098712,
+      "learning_rate": 1.2688940657414362e-05,
+      "loss": 0.3591,
+      "step": 1460
+    },
+    {
+      "epoch": 0.4443433424325144,
+      "grad_norm": 1.1744022801387977,
+      "learning_rate": 1.264061649473657e-05,
+      "loss": 0.4196,
+      "step": 1465
+    },
+    {
+      "epoch": 0.445859872611465,
+      "grad_norm": 0.9522399346317854,
+      "learning_rate": 1.2592225954177453e-05,
+      "loss": 0.3397,
+      "step": 1470
+    },
+    {
+      "epoch": 0.4473764027904155,
+      "grad_norm": 0.9996838939742224,
+      "learning_rate": 1.2543770252142938e-05,
+      "loss": 0.4086,
+      "step": 1475
+    },
+    {
+      "epoch": 0.4488929329693661,
+      "grad_norm": 1.0879756329716435,
+      "learning_rate": 1.2495250606676927e-05,
+      "loss": 0.3837,
+      "step": 1480
+    },
+    {
+      "epoch": 0.45040946314831665,
+      "grad_norm": 1.174932056472148,
+      "learning_rate": 1.2446668237430697e-05,
+      "loss": 0.4314,
+      "step": 1485
+    },
+    {
+      "epoch": 0.45192599332726724,
+      "grad_norm": 0.8386548630342906,
+      "learning_rate": 1.2398024365632229e-05,
+      "loss": 0.3641,
+      "step": 1490
+    },
+    {
+      "epoch": 0.4534425235062178,
+      "grad_norm": 0.9535711206768751,
+      "learning_rate": 1.2349320214055502e-05,
+      "loss": 0.3634,
+      "step": 1495
+    },
+    {
+      "epoch": 0.4549590536851683,
+      "grad_norm": 0.9958036756892384,
+      "learning_rate": 1.2300557006989768e-05,
+      "loss": 0.386,
+      "step": 1500
+    },
+    {
+      "epoch": 0.4549590536851683,
+      "eval_loss": 0.3838871121406555,
+      "eval_runtime": 177.9377,
+      "eval_samples_per_second": 50.214,
+      "eval_steps_per_second": 25.11,
+      "step": 1500
+    },
+    {
+      "epoch": 0.4564755838641189,
+      "grad_norm": 1.1714023611493132,
+      "learning_rate": 1.2251735970208776e-05,
+      "loss": 0.4244,
+      "step": 1505
+    },
+    {
+      "epoch": 0.45799211404306944,
+      "grad_norm": 0.9816643204121793,
+      "learning_rate": 1.2202858330939946e-05,
+      "loss": 0.3582,
+      "step": 1510
+    },
+    {
+      "epoch": 0.45950864422202004,
+      "grad_norm": 1.2215461118765152,
+      "learning_rate": 1.2153925317833544e-05,
+      "loss": 0.3571,
+      "step": 1515
+    },
+    {
+      "epoch": 0.46102517440097057,
+      "grad_norm": 1.0705463427387603,
+      "learning_rate": 1.2104938160931775e-05,
+      "loss": 0.3729,
+      "step": 1520
+    },
+    {
+      "epoch": 0.46254170457992116,
+      "grad_norm": 1.0397410437607952,
+      "learning_rate": 1.2055898091637867e-05,
+      "loss": 0.3818,
+      "step": 1525
+    },
+    {
+      "epoch": 0.4640582347588717,
+      "grad_norm": 1.1007740207154797,
+      "learning_rate": 1.2006806342685127e-05,
+      "loss": 0.4075,
+      "step": 1530
+    },
+    {
+      "epoch": 0.46557476493782224,
+      "grad_norm": 1.0553911498111368,
+      "learning_rate": 1.195766414810595e-05,
+      "loss": 0.4049,
+      "step": 1535
+    },
+    {
+      "epoch": 0.46709129511677283,
+      "grad_norm": 1.048411929765153,
+      "learning_rate": 1.1908472743200787e-05,
+      "loss": 0.4239,
+      "step": 1540
+    },
+    {
+      "epoch": 0.46860782529572337,
+      "grad_norm": 0.9022404852741357,
+      "learning_rate": 1.1859233364507105e-05,
+      "loss": 0.4135,
+      "step": 1545
+    },
+    {
+      "epoch": 0.47012435547467396,
+      "grad_norm": 0.8399864038325862,
+      "learning_rate": 1.1809947249768312e-05,
+      "loss": 0.3431,
+      "step": 1550
+    },
+    {
+      "epoch": 0.4716408856536245,
+      "grad_norm": 0.9086969723025363,
+      "learning_rate": 1.1760615637902615e-05,
+      "loss": 0.4285,
+      "step": 1555
+    },
+    {
+      "epoch": 0.4731574158325751,
+      "grad_norm": 1.038569849088238,
+      "learning_rate": 1.1711239768971908e-05,
+      "loss": 0.3955,
+      "step": 1560
+    },
+    {
+      "epoch": 0.4746739460115256,
+      "grad_norm": 1.0391274519013625,
+      "learning_rate": 1.1661820884150577e-05,
+      "loss": 0.3598,
+      "step": 1565
+    },
+    {
+      "epoch": 0.47619047619047616,
+      "grad_norm": 1.1140866338724735,
+      "learning_rate": 1.1612360225694317e-05,
+      "loss": 0.419,
+      "step": 1570
+    },
+    {
+      "epoch": 0.47770700636942676,
+      "grad_norm": 1.1753789072266334,
+      "learning_rate": 1.1562859036908895e-05,
+      "loss": 0.4039,
+      "step": 1575
+    },
+    {
+      "epoch": 0.4792235365483773,
+      "grad_norm": 0.8669494063557552,
+      "learning_rate": 1.1513318562118902e-05,
+      "loss": 0.2988,
+      "step": 1580
+    },
+    {
+      "epoch": 0.4807400667273279,
+      "grad_norm": 1.1084096884130417,
+      "learning_rate": 1.1463740046636471e-05,
+      "loss": 0.3788,
+      "step": 1585
+    },
+    {
+      "epoch": 0.4822565969062784,
+      "grad_norm": 0.9636937678770835,
+      "learning_rate": 1.1414124736729966e-05,
+      "loss": 0.3891,
+      "step": 1590
+    },
+    {
+      "epoch": 0.483773127085229,
+      "grad_norm": 1.0650960432078391,
+      "learning_rate": 1.1364473879592674e-05,
+      "loss": 0.3815,
+      "step": 1595
+    },
+    {
+      "epoch": 0.48528965726417955,
+      "grad_norm": 0.8673741729500686,
+      "learning_rate": 1.1314788723311438e-05,
+      "loss": 0.402,
+      "step": 1600
+    },
+    {
+      "epoch": 0.48680618744313015,
+      "grad_norm": 0.8885954157154354,
+      "learning_rate": 1.1265070516835286e-05,
+      "loss": 0.3701,
+      "step": 1605
+    },
+    {
+      "epoch": 0.4883227176220807,
+      "grad_norm": 0.9978426248829919,
+      "learning_rate": 1.1215320509944038e-05,
+      "loss": 0.3451,
+      "step": 1610
+    },
+    {
+      "epoch": 0.4898392478010312,
+      "grad_norm": 1.0183575162282423,
+      "learning_rate": 1.1165539953216893e-05,
+      "loss": 0.3681,
+      "step": 1615
+    },
+    {
+      "epoch": 0.4913557779799818,
+      "grad_norm": 1.0579314109624234,
+      "learning_rate": 1.1115730098000982e-05,
+      "loss": 0.3972,
+      "step": 1620
+    },
+    {
+      "epoch": 0.49287230815893235,
+      "grad_norm": 1.032202979176664,
+      "learning_rate": 1.1065892196379928e-05,
+      "loss": 0.4024,
+      "step": 1625
+    },
+    {
+      "epoch": 0.49438883833788294,
+      "grad_norm": 1.1044778878734602,
+      "learning_rate": 1.101602750114236e-05,
+      "loss": 0.4044,
+      "step": 1630
+    },
+    {
+      "epoch": 0.4959053685168335,
+      "grad_norm": 0.8668001202243194,
+      "learning_rate": 1.0966137265750427e-05,
+      "loss": 0.3988,
+      "step": 1635
+    },
+    {
+      "epoch": 0.49742189869578407,
+      "grad_norm": 1.0209675003151268,
+      "learning_rate": 1.0916222744308285e-05,
+      "loss": 0.3875,
+      "step": 1640
+    },
+    {
+      "epoch": 0.4989384288747346,
+      "grad_norm": 1.098694287995219,
+      "learning_rate": 1.0866285191530572e-05,
+      "loss": 0.3787,
+      "step": 1645
+    },
+    {
+      "epoch": 0.5004549590536852,
+      "grad_norm": 1.0520492533718504,
+      "learning_rate": 1.0816325862710884e-05,
+      "loss": 0.367,
+      "step": 1650
+    },
+    {
+      "epoch": 0.5019714892326357,
+      "grad_norm": 0.8661615469589424,
+      "learning_rate": 1.0766346013690193e-05,
+      "loss": 0.3818,
+      "step": 1655
+    },
+    {
+      "epoch": 0.5034880194115863,
+      "grad_norm": 1.0448544478772042,
+      "learning_rate": 1.0716346900825298e-05,
+      "loss": 0.4066,
+      "step": 1660
+    },
+    {
+      "epoch": 0.5050045495905369,
+      "grad_norm": 0.9437379396406751,
+      "learning_rate": 1.066632978095724e-05,
+      "loss": 0.3612,
+      "step": 1665
+    },
+    {
+      "epoch": 0.5065210797694875,
+      "grad_norm": 1.0577638338010937,
+      "learning_rate": 1.0616295911379706e-05,
+      "loss": 0.3931,
+      "step": 1670
+    },
+    {
+      "epoch": 0.5080376099484379,
+      "grad_norm": 1.0224841150046953,
+      "learning_rate": 1.0566246549807424e-05,
+      "loss": 0.3568,
+      "step": 1675
+    },
+    {
+      "epoch": 0.5095541401273885,
+      "grad_norm": 1.0709827796628466,
+      "learning_rate": 1.0516182954344548e-05,
+      "loss": 0.3785,
+      "step": 1680
+    },
+    {
+      "epoch": 0.5110706703063391,
+      "grad_norm": 1.0408023532900905,
+      "learning_rate": 1.0466106383453033e-05,
+      "loss": 0.4197,
+      "step": 1685
+    },
+    {
+      "epoch": 0.5125872004852896,
+      "grad_norm": 0.9978865184694876,
+      "learning_rate": 1.0416018095921002e-05,
+      "loss": 0.361,
+      "step": 1690
+    },
+    {
+      "epoch": 0.5141037306642402,
+      "grad_norm": 0.9482830099197627,
+      "learning_rate": 1.0365919350831105e-05,
+      "loss": 0.3846,
+      "step": 1695
+    },
+    {
+      "epoch": 0.5156202608431908,
+      "grad_norm": 1.098252143243866,
+      "learning_rate": 1.031581140752886e-05,
+      "loss": 0.384,
+      "step": 1700
+    },
+    {
+      "epoch": 0.5171367910221414,
+      "grad_norm": 0.8273439653306288,
+      "learning_rate": 1.0265695525591003e-05,
+      "loss": 0.3457,
+      "step": 1705
+    },
+    {
+      "epoch": 0.5186533212010919,
+      "grad_norm": 0.8472856026513179,
+      "learning_rate": 1.0215572964793838e-05,
+      "loss": 0.3804,
+      "step": 1710
+    },
+    {
+      "epoch": 0.5201698513800425,
+      "grad_norm": 0.9358580008284806,
+      "learning_rate": 1.0165444985081543e-05,
+      "loss": 0.3508,
+      "step": 1715
+    },
+    {
+      "epoch": 0.521686381558993,
+      "grad_norm": 0.7254575613461925,
+      "learning_rate": 1.0115312846534518e-05,
+      "loss": 0.373,
+      "step": 1720
+    },
+    {
+      "epoch": 0.5232029117379435,
+      "grad_norm": 1.0896517684977018,
+      "learning_rate": 1.0065177809337703e-05,
+      "loss": 0.3391,
+      "step": 1725
+    },
+    {
+      "epoch": 0.5247194419168941,
+      "grad_norm": 0.9244440412828214,
+      "learning_rate": 1.0015041133748908e-05,
+      "loss": 0.3588,
+      "step": 1730
+    },
+    {
+      "epoch": 0.5262359720958447,
+      "grad_norm": 1.0159645169016112,
+      "learning_rate": 9.964904080067119e-06,
+      "loss": 0.3627,
+      "step": 1735
+    },
+    {
+      "epoch": 0.5277525022747953,
+      "grad_norm": 0.8376669598715493,
+      "learning_rate": 9.914767908600835e-06,
+      "loss": 0.3369,
+      "step": 1740
+    },
+    {
+      "epoch": 0.5292690324537458,
+      "grad_norm": 0.7978725648817389,
+      "learning_rate": 9.864633879636371e-06,
+      "loss": 0.326,
+      "step": 1745
+    },
+    {
+      "epoch": 0.5307855626326964,
+      "grad_norm": 0.9497158806314906,
+      "learning_rate": 9.814503253406188e-06,
+      "loss": 0.3761,
+      "step": 1750
+    },
+    {
+      "epoch": 0.532302092811647,
+      "grad_norm": 0.8666942612740441,
+      "learning_rate": 9.764377290057217e-06,
+      "loss": 0.3237,
+      "step": 1755
+    },
+    {
+      "epoch": 0.5338186229905975,
+      "grad_norm": 1.015724277407853,
+      "learning_rate": 9.714257249619166e-06,
+      "loss": 0.3851,
+      "step": 1760
+    },
+    {
+      "epoch": 0.535335153169548,
+      "grad_norm": 0.8370259755101838,
+      "learning_rate": 9.664144391972867e-06,
+      "loss": 0.3499,
+      "step": 1765
+    },
+    {
+      "epoch": 0.5368516833484986,
+      "grad_norm": 0.8982590989333401,
+      "learning_rate": 9.614039976818591e-06,
+      "loss": 0.3871,
+      "step": 1770
+    },
+    {
+      "epoch": 0.5383682135274492,
+      "grad_norm": 0.8871582462703803,
+      "learning_rate": 9.56394526364439e-06,
+      "loss": 0.3852,
+      "step": 1775
+    },
+    {
+      "epoch": 0.5398847437063997,
+      "grad_norm": 0.9733274161552575,
+      "learning_rate": 9.513861511694432e-06,
+      "loss": 0.3754,
+      "step": 1780
+    },
+    {
+      "epoch": 0.5414012738853503,
+      "grad_norm": 0.9575249396664576,
+      "learning_rate": 9.46378997993735e-06,
+      "loss": 0.3571,
+      "step": 1785
+    },
+    {
+      "epoch": 0.5429178040643009,
+      "grad_norm": 0.9607725237981556,
+      "learning_rate": 9.413731927034607e-06,
+      "loss": 0.3747,
+      "step": 1790
+    },
+    {
+      "epoch": 0.5444343342432515,
+      "grad_norm": 1.165305127561796,
+      "learning_rate": 9.363688611308825e-06,
+      "loss": 0.39,
+      "step": 1795
+    },
+    {
+      "epoch": 0.545950864422202,
+      "grad_norm": 1.2378284714450407,
+      "learning_rate": 9.313661290712182e-06,
+      "loss": 0.4023,
+      "step": 1800
+    },
+    {
+      "epoch": 0.5474673946011526,
+      "grad_norm": 0.9390807801563313,
+      "learning_rate": 9.26365122279479e-06,
+      "loss": 0.3655,
+      "step": 1805
+    },
+    {
+      "epoch": 0.5489839247801032,
+      "grad_norm": 0.8512608333919068,
+      "learning_rate": 9.213659664673063e-06,
+      "loss": 0.3833,
+      "step": 1810
+    },
+    {
+      "epoch": 0.5505004549590536,
+      "grad_norm": 0.9579623704318949,
+      "learning_rate": 9.163687872998134e-06,
+      "loss": 0.3565,
+      "step": 1815
+    },
+    {
+      "epoch": 0.5520169851380042,
+      "grad_norm": 0.9096584536659199,
+      "learning_rate": 9.113737103924266e-06,
+      "loss": 0.3946,
+      "step": 1820
+    },
+    {
+      "epoch": 0.5535335153169548,
+      "grad_norm": 0.8686976809327038,
+      "learning_rate": 9.063808613077265e-06,
+      "loss": 0.3416,
+      "step": 1825
+    },
+    {
+      "epoch": 0.5550500454959054,
+      "grad_norm": 1.04912717783663,
+      "learning_rate": 9.013903655522931e-06,
+      "loss": 0.4476,
+      "step": 1830
+    },
+    {
+      "epoch": 0.5565665756748559,
+      "grad_norm": 1.1364230419987023,
+      "learning_rate": 8.964023485735491e-06,
+      "loss": 0.3816,
+      "step": 1835
+    },
+    {
+      "epoch": 0.5580831058538065,
+      "grad_norm": 0.8842214833551543,
+      "learning_rate": 8.914169357566082e-06,
+      "loss": 0.3291,
+      "step": 1840
+    },
+    {
+      "epoch": 0.5595996360327571,
+      "grad_norm": 0.9895945394957149,
+      "learning_rate": 8.864342524211228e-06,
+      "loss": 0.3881,
+      "step": 1845
+    },
+    {
+      "epoch": 0.5611161662117076,
+      "grad_norm": 0.9031975360786588,
+      "learning_rate": 8.814544238181327e-06,
+      "loss": 0.407,
+      "step": 1850
+    },
+    {
+      "epoch": 0.5626326963906582,
+      "grad_norm": 1.011452507718733,
+      "learning_rate": 8.764775751269184e-06,
+      "loss": 0.3784,
+      "step": 1855
+    },
+    {
+      "epoch": 0.5641492265696088,
+      "grad_norm": 1.0249078931801974,
+      "learning_rate": 8.715038314518532e-06,
+      "loss": 0.387,
+      "step": 1860
+    },
+    {
+      "epoch": 0.5656657567485593,
+      "grad_norm": 1.0281856208543434,
+      "learning_rate": 8.66533317819259e-06,
+      "loss": 0.349,
+      "step": 1865
+    },
+    {
+      "epoch": 0.5671822869275098,
+      "grad_norm": 0.8713289058377761,
+      "learning_rate": 8.615661591742626e-06,
+      "loss": 0.3775,
+      "step": 1870
+    },
+    {
+      "epoch": 0.5686988171064604,
+      "grad_norm": 0.9270856112232262,
+      "learning_rate": 8.566024803776567e-06,
+      "loss": 0.403,
+      "step": 1875
+    },
+    {
+      "epoch": 0.570215347285411,
+      "grad_norm": 0.9868872288291042,
+      "learning_rate": 8.516424062027587e-06,
+      "loss": 0.3474,
+      "step": 1880
+    },
+    {
+      "epoch": 0.5717318774643615,
+      "grad_norm": 0.9141693333613865,
+      "learning_rate": 8.466860613322773e-06,
+      "loss": 0.3596,
+      "step": 1885
+    },
+    {
+      "epoch": 0.5732484076433121,
+      "grad_norm": 1.13028177709786,
+      "learning_rate": 8.417335703551753e-06,
+      "loss": 0.39,
+      "step": 1890
+    },
+    {
+      "epoch": 0.5747649378222627,
+      "grad_norm": 0.963375132739776,
+      "learning_rate": 8.3678505776354e-06,
+      "loss": 0.3582,
+      "step": 1895
+    },
+    {
+      "epoch": 0.5762814680012133,
+      "grad_norm": 0.8566314951014108,
+      "learning_rate": 8.318406479494526e-06,
+      "loss": 0.3499,
+      "step": 1900
+    },
+    {
+      "epoch": 0.5777979981801638,
+      "grad_norm": 0.871827865692706,
+      "learning_rate": 8.269004652018615e-06,
+      "loss": 0.3378,
+      "step": 1905
+    },
+    {
+      "epoch": 0.5793145283591143,
+      "grad_norm": 0.8864124825854094,
+      "learning_rate": 8.219646337034587e-06,
+      "loss": 0.3663,
+      "step": 1910
+    },
+    {
+      "epoch": 0.5808310585380649,
+      "grad_norm": 1.0414992512722574,
+      "learning_rate": 8.170332775275572e-06,
+      "loss": 0.3895,
+      "step": 1915
+    },
+    {
+      "epoch": 0.5823475887170154,
+      "grad_norm": 1.1285331789955597,
+      "learning_rate": 8.12106520634973e-06,
+      "loss": 0.391,
+      "step": 1920
+    },
+    {
+      "epoch": 0.583864118895966,
+      "grad_norm": 1.1338277664903524,
+      "learning_rate": 8.071844868709086e-06,
+      "loss": 0.4213,
+      "step": 1925
+    },
+    {
+      "epoch": 0.5853806490749166,
+      "grad_norm": 1.0011965373177365,
+      "learning_rate": 8.022672999618394e-06,
+      "loss": 0.3739,
+      "step": 1930
+    },
+    {
+      "epoch": 0.5868971792538672,
+      "grad_norm": 0.8353717953496863,
+      "learning_rate": 7.973550835124055e-06,
+      "loss": 0.3886,
+      "step": 1935
+    },
+    {
+      "epoch": 0.5884137094328177,
+      "grad_norm": 1.0274242422703623,
+      "learning_rate": 7.924479610023016e-06,
+      "loss": 0.3554,
+      "step": 1940
+    },
+    {
+      "epoch": 0.5899302396117683,
+      "grad_norm": 0.9984155763187357,
+      "learning_rate": 7.875460557831755e-06,
+      "loss": 0.3556,
+      "step": 1945
+    },
+    {
+      "epoch": 0.5914467697907189,
+      "grad_norm": 0.8717630327395798,
+      "learning_rate": 7.82649491075527e-06,
+      "loss": 0.3411,
+      "step": 1950
+    },
+    {
+      "epoch": 0.5929632999696693,
+      "grad_norm": 1.0351492468740082,
+      "learning_rate": 7.777583899656092e-06,
+      "loss": 0.3599,
+      "step": 1955
+    },
+    {
+      "epoch": 0.5944798301486199,
+      "grad_norm": 0.9357066998157261,
+      "learning_rate": 7.728728754023354e-06,
+      "loss": 0.3373,
+      "step": 1960
+    },
+    {
+      "epoch": 0.5959963603275705,
+      "grad_norm": 0.8738803213245617,
+      "learning_rate": 7.679930701941888e-06,
+      "loss": 0.3475,
+      "step": 1965
+    },
+    {
+      "epoch": 0.5975128905065211,
+      "grad_norm": 0.9262823621070619,
+      "learning_rate": 7.631190970061349e-06,
+      "loss": 0.3519,
+      "step": 1970
+    },
+    {
+      "epoch": 0.5990294206854716,
+      "grad_norm": 0.9004476821654784,
+      "learning_rate": 7.5825107835653814e-06,
+      "loss": 0.326,
+      "step": 1975
+    },
+    {
+      "epoch": 0.6005459508644222,
+      "grad_norm": 1.0358901873299116,
+      "learning_rate": 7.533891366140815e-06,
+      "loss": 0.3521,
+      "step": 1980
+    },
+    {
+      "epoch": 0.6020624810433728,
+      "grad_norm": 0.8420033070489784,
+      "learning_rate": 7.485333939946926e-06,
+      "loss": 0.3379,
+      "step": 1985
+    },
+    {
+      "epoch": 0.6035790112223233,
+      "grad_norm": 1.0494757959050431,
+      "learning_rate": 7.4368397255846845e-06,
+      "loss": 0.3945,
+      "step": 1990
+    },
+    {
+      "epoch": 0.6050955414012739,
+      "grad_norm": 0.9524555408133961,
+      "learning_rate": 7.388409942066099e-06,
+      "loss": 0.3596,
+      "step": 1995
+    },
+    {
+      "epoch": 0.6066120715802245,
+      "grad_norm": 1.0996405384525345,
+      "learning_rate": 7.340045806783559e-06,
+      "loss": 0.409,
+      "step": 2000
+    },
+    {
+      "epoch": 0.6066120715802245,
+      "eval_loss": 0.3754875957965851,
+      "eval_runtime": 173.9951,
+      "eval_samples_per_second": 51.352,
+      "eval_steps_per_second": 25.679,
+      "step": 2000
+    }
+  ],
+  "logging_steps": 5,
+  "max_steps": 3297,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "EarlyStoppingCallback": {
+      "args": {
+        "early_stopping_patience": 3,
+        "early_stopping_threshold": 0.0
+      },
+      "attributes": {
+        "early_stopping_patience_counter": 0
+      }
+    },
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 97417710010368.0,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-2000/zero_to_fp32.py

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

checkpoint-2500/added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoint-2500/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# 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' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first 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' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n\n' }}
+    {%- endif %}
+{%- endif %}

checkpoint-2500/config.json

@@ -0,0 +1,59 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.53.1",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-2500/generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "4.53.1"
+}

checkpoint-2500/latest

@@ -0,0 +1 @@
+global_step2500

checkpoint-2500/special_tokens_map.json

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

checkpoint-2500/tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "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
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-2500/zero_to_fp32.py

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

checkpoint-3000/added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoint-3000/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# 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' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first 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' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n\n' }}
+    {%- endif %}
+{%- endif %}

checkpoint-3000/config.json

@@ -0,0 +1,59 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.53.1",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-3000/latest

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

checkpoint-3000/special_tokens_map.json

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

checkpoint-3000/tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "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
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-3000/zero_to_fp32.py

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

checkpoint-3297/added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoint-3297/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# 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' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first 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' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n\n' }}
+    {%- endif %}
+{%- endif %}

checkpoint-3297/config.json

@@ -0,0 +1,59 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.53.1",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-3297/generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "4.53.1"
+}

checkpoint-3297/latest

@@ -0,0 +1 @@
+global_step3297

checkpoint-3297/special_tokens_map.json

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

checkpoint-3297/tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "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
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoint-3297/zero_to_fp32.py

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

config.json

@@ -0,0 +1,59 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.53.1",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "4.53.1"
+}

special_tokens_map.json

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

tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "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
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}