diff --git a/.gitattributes b/.gitattributes
index a6344aac8c09253b3b630fb776ae94478aa0275b..c149555b5474b181cd0ce18afd369315ab44c2ae 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -33,3 +33,18 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+qwen3/amid/checkpoint-1256/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/amid/checkpoint-1570/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/amid/checkpoint-314/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/amid/checkpoint-628/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/amid/checkpoint-942/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/amid/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/csd/checkpoint-1256/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/csd/checkpoint-1570/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/csd/checkpoint-314/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/csd/checkpoint-628/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/csd/checkpoint-942/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/csd/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/distillm_adaptive_sfkl/checkpoint-1256/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/distillm_adaptive_sfkl/checkpoint-1570/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+qwen3/distillm_adaptive_sfkl/checkpoint-314/tokenizer.json filter=lfs diff=lfs merge=lfs -text
diff --git a/qwen3/amid/README.md b/qwen3/amid/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..6b12ad5993fadf3afe3a4cea738a0ca2aed10f6c
--- /dev/null
+++ b/qwen3/amid/README.md
@@ -0,0 +1,79 @@
+---
+library_name: peft
+license: other
+base_model: Qwen/Qwen3-0.6B
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+pipeline_tag: text-generation
+model-index:
+- name: amid
+  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. -->
+
+# amid
+
+This model is a fine-tuned version of [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) on the cypher_prepared_train dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.0040
+- Selector Count: 1712.0
+- Generator Count: 1712.0
+- Selector Accuracy: 98.6565
+- Generator Exact Match: 96.028
+- Generator Rouge1: 99.8671
+- Generator Rouge2: 99.8011
+- Generator Rougel: 99.6739
+
+## 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: 0.0001
+- train_batch_size: 2
+- eval_batch_size: 16
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 2
+- gradient_accumulation_steps: 8
+- total_train_batch_size: 32
+- total_eval_batch_size: 32
+- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine_with_min_lr
+- num_epochs: 5
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss | Selector Count | Generator Count | Selector Accuracy | Generator Exact Match | Generator Rouge1 | Generator Rouge2 | Generator Rougel |
+|:-------------:|:-----:|:----:|:---------------:|:--------------:|:---------------:|:-----------------:|:---------------------:|:----------------:|:----------------:|:----------------:|
+| 0.0068        | 1.0   | 314  | 0.0128          | 1712.0         | 1712.0          | 90.1869           | 87.3832               | 99.2069          | 98.8793          | 98.7553          |
+| 0.0036        | 2.0   | 628  | 0.0069          | 1712.0         | 1712.0          | 96.9042           | 93.5164               | 99.7294          | 99.5697          | 99.5152          |
+| 0.0006        | 3.0   | 942  | 0.0057          | 1712.0         | 1712.0          | 96.5537           | 95.3855               | 99.8982          | 99.8084          | 99.6633          |
+| 0.0003        | 4.0   | 1256 | 0.0043          | 1712.0         | 1712.0          | 98.1893           | 95.8528               | 99.9186          | 99.8428          | 99.7108          |
+| 0.0004        | 5.0   | 1570 | 0.0040          | 1712.0         | 1712.0          | 98.6565           | 95.736                | 99.9082          | 99.8294          | 99.7068          |
+
+
+### Framework versions
+
+- PEFT 0.18.1
+- Transformers 5.8.0
+- Pytorch 2.9.1+cu128
+- Datasets 4.0.0
+- Tokenizers 0.22.2
\ No newline at end of file
diff --git a/qwen3/amid/adapter_config.json b/qwen3/amid/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..16362f47190537658d6b04519de786417a5f6c2f
--- /dev/null
+++ b/qwen3/amid/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/amid/adapter_model.safetensors b/qwen3/amid/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..93524c4b229da26352140faae8b8d1bed15559bc
--- /dev/null
+++ b/qwen3/amid/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db85fbee8a5b8671bb7eea278430126deabd15db5ca476d10a9fd242caf6c334
+size 40422560
diff --git a/qwen3/amid/all_results.json b/qwen3/amid/all_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..c884ac87bac73cacf29b75b03b5e83eaceb2b530
--- /dev/null
+++ b/qwen3/amid/all_results.json
@@ -0,0 +1,19 @@
+{
+    "epoch": 5.0,
+    "eval_generator_count": 1712.0,
+    "eval_generator_exact_match": 96.028,
+    "eval_generator_rouge1": 99.8671,
+    "eval_generator_rouge2": 99.8011,
+    "eval_generator_rougeL": 99.6739,
+    "eval_loss": 0.004025007598102093,
+    "eval_runtime": 579.6459,
+    "eval_samples_per_second": 5.89,
+    "eval_selector_accuracy": 98.6565,
+    "eval_selector_count": 1712.0,
+    "eval_steps_per_second": 0.185,
+    "total_flos": 5.871444840310374e+16,
+    "train_loss": 0.01864182253792336,
+    "train_runtime": 7621.8556,
+    "train_samples_per_second": 6.578,
+    "train_steps_per_second": 0.206
+}
\ No newline at end of file
diff --git a/qwen3/amid/chat_template.jinja b/qwen3/amid/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/amid/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1256/README.md b/qwen3/amid/checkpoint-1256/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1256/adapter_config.json b/qwen3/amid/checkpoint-1256/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..16362f47190537658d6b04519de786417a5f6c2f
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1256/adapter_model.safetensors b/qwen3/amid/checkpoint-1256/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..18d33b0d0510c794c3dbe3b609c61fa1166bff4b
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e895f0f483699995b38d2518b8ecce916cfba3f34b558ab5d80a2533af2989f8
+size 40422560
diff --git a/qwen3/amid/checkpoint-1256/chat_template.jinja b/qwen3/amid/checkpoint-1256/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..67dc43ff15fb981ccd75531c19301c37143f964e
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:daf81cbb6f746277d70839536a20bb3b6caef9b74a015978d16aef31a5c91d3e
+size 121135557
diff --git a/qwen3/amid/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..d98497b39bf0def47ebd28a351286ed47a8fa75a
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44369c0a57c589843da109a7f8fd845aebfbf5d3a285d0ed07f57adac8935df3
+size 121135749
diff --git a/qwen3/amid/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt b/qwen3/amid/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..bf82f007d10484089fdfa7d383764f985e0f7d1e
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7cf6368e603b46a1c1678c78c926cd574cebc8901f75d8b9ce6c471073a294d
+size 351841765
diff --git a/qwen3/amid/checkpoint-1256/latest b/qwen3/amid/checkpoint-1256/latest
new file mode 100644
index 0000000000000000000000000000000000000000..74abb3d23ea83ebdac59918752db5d1782834ca7
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/latest
@@ -0,0 +1 @@
+global_step1256
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1256/rng_state_0.pth b/qwen3/amid/checkpoint-1256/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..2b7f039f457e66dff06dde7241eed0ae7045f59a
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7530a82318bfc63bf9efe3909e0cfb53991fac12078349c3cd390c37b6717a9c
+size 14917
diff --git a/qwen3/amid/checkpoint-1256/rng_state_1.pth b/qwen3/amid/checkpoint-1256/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..069e0c29d1ba8eb599dc83543dbe5bcff209ded7
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82240e85efab23fc7f6753a25eb53f58033ffdc19e106b8c4ca745f9571b2f29
+size 14917
diff --git a/qwen3/amid/checkpoint-1256/scheduler.pt b/qwen3/amid/checkpoint-1256/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..40e863d07cfed22cd912014eb3ba3fd2b16b173f
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1775297f14e75f7463ae51983a66296f81f4b2e2e5d7234f1a6d295abfa4cf23
+size 1465
diff --git a/qwen3/amid/checkpoint-1256/tokenizer.json b/qwen3/amid/checkpoint-1256/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/amid/checkpoint-1256/tokenizer_config.json b/qwen3/amid/checkpoint-1256/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/amid/checkpoint-1256/trainer_state.json b/qwen3/amid/checkpoint-1256/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..765791289ac887e7428011b543fd937ff1430919
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/trainer_state.json
@@ -0,0 +1,528 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 4.0,
+  "eval_steps": 500,
+  "global_step": 1256,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 0.38384687900543213,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.27438180446624755,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.38263317942619324,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.13717052936553956,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.08620006591081619,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.13135391473770142,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.15773959457874298,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.1076702356338501,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.22962816059589386,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.08698999881744385,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.12232966721057892,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.1128507375717163,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.15597161650657654,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.1277535080909729,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.7329031825065613,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.08885469436645507,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.45483648777008057,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.07881734371185303,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.16407009959220886,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.06113555431365967,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.07238732278347015,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.06768973469734192,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.4741341471672058,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.04917148947715759,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.36410900950431824,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.010153122991323472,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.11222824454307556,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.00552816241979599,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.47910866141319275,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.006760746240615845,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 87.3832,
+      "eval_generator_rouge1": 99.2069,
+      "eval_generator_rouge2": 98.8793,
+      "eval_generator_rougeL": 98.7553,
+      "eval_loss": 0.012839075177907944,
+      "eval_runtime": 592.0147,
+      "eval_samples_per_second": 5.767,
+      "eval_selector_accuracy": 90.1869,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.27295932173728943,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.009849193692207336,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.060826949775218964,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.005673223733901977,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 0.04862668365240097,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.005297792330384255,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 0.05064748600125313,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": 0.0021365636959671972,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.05359279736876488,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": 0.00429258793592453,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.11369121074676514,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0034508116543292998,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.01604766771197319,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": 0.0028350137174129484,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.15401995182037354,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": 0.005546276643872261,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 0.5877118706703186,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": 0.0044526983052492145,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.10250785946846008,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": 0.0056363165378570555,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.043251167982816696,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": 0.0032581716775894167,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.2089434117078781,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": 0.004207362234592438,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.14037534594535828,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": 0.0019151581451296805,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 0.22862067818641663,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": 0.004427855834364891,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.03684071823954582,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": 0.0017833203077316285,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.0672929510474205,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": 0.0035866759717464447,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 93.5164,
+      "eval_generator_rouge1": 99.7294,
+      "eval_generator_rouge2": 99.5697,
+      "eval_generator_rougeL": 99.5152,
+      "eval_loss": 0.006910035386681557,
+      "eval_runtime": 591.6591,
+      "eval_samples_per_second": 5.77,
+      "eval_selector_accuracy": 96.9042,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.029134005308151245,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": 0.0020816387608647347,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.17367292940616608,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": 0.0019141122698783874,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.016447540372610092,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": 0.0009251200594007969,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.015570643357932568,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": 0.0017806822434067726,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.010291235521435738,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": 0.0010542120784521104,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.012873738072812557,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": 0.0009571929462254047,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.009928264655172825,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": 0.0007098949514329434,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.014300181530416012,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": 0.0007610605098307133,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.03284010663628578,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": 0.002758556976914406,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.19087237119674683,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": 0.0011194748803973198,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.02937050722539425,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": 0.0014943076297640801,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.031206471845507622,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": 0.0027773819863796236,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.019134322181344032,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": 0.0018997505307197572,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.2537505030632019,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": 0.0010082274675369263,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.0017131563508883119,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": 0.0012075364589691161,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.03925931081175804,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": 0.000642749946564436,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.3855,
+      "eval_generator_rouge1": 99.8982,
+      "eval_generator_rouge2": 99.8084,
+      "eval_generator_rougeL": 99.6633,
+      "eval_loss": 0.0056763943284749985,
+      "eval_runtime": 590.3317,
+      "eval_samples_per_second": 5.783,
+      "eval_selector_accuracy": 96.5537,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.0036900911945849657,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": 0.0012762822210788728,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.028549807146191597,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": 0.0006767791230231524,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.007299487479031086,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": 0.0012082518078386784,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.03702578693628311,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": 0.0014168912544846534,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.21273940801620483,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": 0.0012415657751262189,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.0033840558025985956,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": 0.0009985800832509994,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.020664699375629425,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": 0.0011398554779589176,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.0015457016415894032,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": 0.0016313906759023666,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.011757194995880127,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": 0.0010684640146791935,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.007703748065978289,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": 0.0006688186898827553,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.04895896837115288,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": 0.0016563836485147475,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.04278242960572243,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": 0.0006196124479174614,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.004100222606211901,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": 0.0009099396876990796,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.06636715680360794,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": 0.0012004765681922435,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.005992025602608919,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": 0.0003460923908278346,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.8528,
+      "eval_generator_rouge1": 99.9186,
+      "eval_generator_rouge2": 99.8428,
+      "eval_generator_rougeL": 99.7108,
+      "eval_loss": 0.004311855416744947,
+      "eval_runtime": 587.5171,
+      "eval_samples_per_second": 5.811,
+      "eval_selector_accuracy": 98.1893,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.182,
+      "step": 1256
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 4.697643190059008e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/amid/checkpoint-1256/training_args.bin b/qwen3/amid/checkpoint-1256/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..571568fd8ab61d629f2a731739ee4c9e1309cca4
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e67133ff7a80f6f166c03bd5f62d9c580548b0f6840a1ce3891dee623991e47a
+size 7185
diff --git a/qwen3/amid/checkpoint-1256/zero_to_fp32.py b/qwen3/amid/checkpoint-1256/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/amid/checkpoint-1256/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/amid/checkpoint-1570/README.md b/qwen3/amid/checkpoint-1570/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1570/adapter_config.json b/qwen3/amid/checkpoint-1570/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..16362f47190537658d6b04519de786417a5f6c2f
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1570/adapter_model.safetensors b/qwen3/amid/checkpoint-1570/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..93524c4b229da26352140faae8b8d1bed15559bc
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db85fbee8a5b8671bb7eea278430126deabd15db5ca476d10a9fd242caf6c334
+size 40422560
diff --git a/qwen3/amid/checkpoint-1570/chat_template.jinja b/qwen3/amid/checkpoint-1570/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..b2a17da452982fbffb2c4120f5d2c208bc177b3f
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c8a2fb2070f7e001bc99481db2203d5aa9bf9e5ffcb975a0d649c8b3afd2e3d
+size 121135557
diff --git a/qwen3/amid/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..69f98096e3ef7d0022322f2b8be4afcfadb6e9ab
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14e01dfbc51732351c410136faeca403b1f7b4bf8d7dc4d3fed0e8fe44352467
+size 121135749
diff --git a/qwen3/amid/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt b/qwen3/amid/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..49eda819fd2374e2ceb2c262558ba7272fffab7a
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30a51ab7a4e9a5c35e16652def3c25f866a3914f6bd6c7f31e27993036a6619c
+size 351841765
diff --git a/qwen3/amid/checkpoint-1570/latest b/qwen3/amid/checkpoint-1570/latest
new file mode 100644
index 0000000000000000000000000000000000000000..18c2fcff58fa6daa0a2e7c3a015552a69085778b
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/latest
@@ -0,0 +1 @@
+global_step1570
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-1570/rng_state_0.pth b/qwen3/amid/checkpoint-1570/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..928e2f15274bc26236f1b69d09740ff4421e3412
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd57b79bb2ca202aa5c978ef0a2bf65f56add247a12c8c298d32cf0c38ee2b3c
+size 14917
diff --git a/qwen3/amid/checkpoint-1570/rng_state_1.pth b/qwen3/amid/checkpoint-1570/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..0ed396f7be83535575cdd77fea793c4b9f9d0a18
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1703a50a85e61da94a321183a031f61260e7309d0f19e4f03215e48c454d4087
+size 14917
diff --git a/qwen3/amid/checkpoint-1570/scheduler.pt b/qwen3/amid/checkpoint-1570/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..e22954463492d8a070ae427c95418df61893c54c
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6332fac17ae99a07597512c705658da2fd8429d1c39bb776b30e107c2cb7aae
+size 1465
diff --git a/qwen3/amid/checkpoint-1570/tokenizer.json b/qwen3/amid/checkpoint-1570/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/amid/checkpoint-1570/tokenizer_config.json b/qwen3/amid/checkpoint-1570/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/amid/checkpoint-1570/trainer_state.json b/qwen3/amid/checkpoint-1570/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..2698a13a67fd52a3f87941f89b547407e0d10a72
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/trainer_state.json
@@ -0,0 +1,655 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 5.0,
+  "eval_steps": 500,
+  "global_step": 1570,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 0.38384687900543213,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.27438180446624755,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.38263317942619324,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.13717052936553956,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.08620006591081619,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.13135391473770142,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.15773959457874298,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.1076702356338501,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.22962816059589386,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.08698999881744385,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.12232966721057892,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.1128507375717163,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.15597161650657654,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.1277535080909729,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.7329031825065613,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.08885469436645507,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.45483648777008057,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.07881734371185303,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.16407009959220886,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.06113555431365967,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.07238732278347015,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.06768973469734192,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.4741341471672058,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.04917148947715759,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.36410900950431824,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.010153122991323472,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.11222824454307556,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.00552816241979599,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.47910866141319275,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.006760746240615845,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 87.3832,
+      "eval_generator_rouge1": 99.2069,
+      "eval_generator_rouge2": 98.8793,
+      "eval_generator_rougeL": 98.7553,
+      "eval_loss": 0.012839075177907944,
+      "eval_runtime": 592.0147,
+      "eval_samples_per_second": 5.767,
+      "eval_selector_accuracy": 90.1869,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.27295932173728943,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.009849193692207336,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.060826949775218964,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.005673223733901977,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 0.04862668365240097,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.005297792330384255,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 0.05064748600125313,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": 0.0021365636959671972,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.05359279736876488,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": 0.00429258793592453,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.11369121074676514,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0034508116543292998,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.01604766771197319,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": 0.0028350137174129484,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.15401995182037354,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": 0.005546276643872261,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 0.5877118706703186,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": 0.0044526983052492145,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.10250785946846008,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": 0.0056363165378570555,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.043251167982816696,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": 0.0032581716775894167,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.2089434117078781,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": 0.004207362234592438,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.14037534594535828,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": 0.0019151581451296805,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 0.22862067818641663,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": 0.004427855834364891,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.03684071823954582,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": 0.0017833203077316285,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.0672929510474205,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": 0.0035866759717464447,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 93.5164,
+      "eval_generator_rouge1": 99.7294,
+      "eval_generator_rouge2": 99.5697,
+      "eval_generator_rougeL": 99.5152,
+      "eval_loss": 0.006910035386681557,
+      "eval_runtime": 591.6591,
+      "eval_samples_per_second": 5.77,
+      "eval_selector_accuracy": 96.9042,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.029134005308151245,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": 0.0020816387608647347,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.17367292940616608,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": 0.0019141122698783874,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.016447540372610092,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": 0.0009251200594007969,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.015570643357932568,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": 0.0017806822434067726,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.010291235521435738,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": 0.0010542120784521104,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.012873738072812557,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": 0.0009571929462254047,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.009928264655172825,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": 0.0007098949514329434,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.014300181530416012,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": 0.0007610605098307133,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.03284010663628578,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": 0.002758556976914406,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.19087237119674683,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": 0.0011194748803973198,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.02937050722539425,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": 0.0014943076297640801,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.031206471845507622,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": 0.0027773819863796236,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.019134322181344032,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": 0.0018997505307197572,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.2537505030632019,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": 0.0010082274675369263,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.0017131563508883119,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": 0.0012075364589691161,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.03925931081175804,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": 0.000642749946564436,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.3855,
+      "eval_generator_rouge1": 99.8982,
+      "eval_generator_rouge2": 99.8084,
+      "eval_generator_rougeL": 99.6633,
+      "eval_loss": 0.0056763943284749985,
+      "eval_runtime": 590.3317,
+      "eval_samples_per_second": 5.783,
+      "eval_selector_accuracy": 96.5537,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.0036900911945849657,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": 0.0012762822210788728,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.028549807146191597,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": 0.0006767791230231524,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.007299487479031086,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": 0.0012082518078386784,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.03702578693628311,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": 0.0014168912544846534,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.21273940801620483,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": 0.0012415657751262189,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.0033840558025985956,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": 0.0009985800832509994,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.020664699375629425,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": 0.0011398554779589176,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.0015457016415894032,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": 0.0016313906759023666,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.011757194995880127,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": 0.0010684640146791935,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.007703748065978289,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": 0.0006688186898827553,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.04895896837115288,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": 0.0016563836485147475,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.04278242960572243,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": 0.0006196124479174614,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.004100222606211901,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": 0.0009099396876990796,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.06636715680360794,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": 0.0012004765681922435,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.005992025602608919,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": 0.0003460923908278346,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.8528,
+      "eval_generator_rouge1": 99.9186,
+      "eval_generator_rouge2": 99.8428,
+      "eval_generator_rougeL": 99.7108,
+      "eval_loss": 0.004311855416744947,
+      "eval_runtime": 587.5171,
+      "eval_samples_per_second": 5.811,
+      "eval_selector_accuracy": 98.1893,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.182,
+      "step": 1256
+    },
+    {
+      "epoch": 4.0127642600717985,
+      "grad_norm": 0.12402024120092392,
+      "learning_rate": 9.464081721960592e-06,
+      "loss": 0.0005224554799497128,
+      "step": 1260
+    },
+    {
+      "epoch": 4.076585560430794,
+      "grad_norm": 0.032988350838422775,
+      "learning_rate": 8.331628936475818e-06,
+      "loss": 0.00022895974107086658,
+      "step": 1280
+    },
+    {
+      "epoch": 4.140406860789788,
+      "grad_norm": 0.002514337655156851,
+      "learning_rate": 7.2659843568748385e-06,
+      "loss": 0.00028093236032873393,
+      "step": 1300
+    },
+    {
+      "epoch": 4.204228161148784,
+      "grad_norm": 0.013952493667602539,
+      "learning_rate": 6.268854516770008e-06,
+      "loss": 0.0007838346064090729,
+      "step": 1320
+    },
+    {
+      "epoch": 4.268049461507778,
+      "grad_norm": 0.00595195172354579,
+      "learning_rate": 5.341836229602331e-06,
+      "loss": 0.0010836148634552956,
+      "step": 1340
+    },
+    {
+      "epoch": 4.331870761866773,
+      "grad_norm": 0.24550670385360718,
+      "learning_rate": 4.4864140314888216e-06,
+      "loss": 0.001040823757648468,
+      "step": 1360
+    },
+    {
+      "epoch": 4.395692062225768,
+      "grad_norm": 0.009249023161828518,
+      "learning_rate": 3.703957803871932e-06,
+      "loss": 0.0006210384890437126,
+      "step": 1380
+    },
+    {
+      "epoch": 4.4595133625847625,
+      "grad_norm": 0.010251481086015701,
+      "learning_rate": 2.9957205797780495e-06,
+      "loss": 0.0005129187367856503,
+      "step": 1400
+    },
+    {
+      "epoch": 4.523334662943758,
+      "grad_norm": 0.024013608694076538,
+      "learning_rate": 2.362836537198199e-06,
+      "loss": 0.0004933405201882124,
+      "step": 1420
+    },
+    {
+      "epoch": 4.587155963302752,
+      "grad_norm": 0.028239818289875984,
+      "learning_rate": 1.8063191828043703e-06,
+      "loss": 0.0005304558202624321,
+      "step": 1440
+    },
+    {
+      "epoch": 4.650977263661747,
+      "grad_norm": 0.01863635517656803,
+      "learning_rate": 1.3270597289100437e-06,
+      "loss": 0.00033774527255445717,
+      "step": 1460
+    },
+    {
+      "epoch": 4.714798564020742,
+      "grad_norm": 0.0036166764330118895,
+      "learning_rate": 9.258256662741205e-07,
+      "loss": 0.0002381577854976058,
+      "step": 1480
+    },
+    {
+      "epoch": 4.778619864379737,
+      "grad_norm": 0.016647396609187126,
+      "learning_rate": 6.032595350337341e-07,
+      "loss": 0.0005346011370420456,
+      "step": 1500
+    },
+    {
+      "epoch": 4.842441164738732,
+      "grad_norm": 0.14798367023468018,
+      "learning_rate": 3.5987789573421356e-07,
+      "loss": 0.0015266412869095803,
+      "step": 1520
+    },
+    {
+      "epoch": 4.9062624650977265,
+      "grad_norm": 0.1593816727399826,
+      "learning_rate": 1.9607050210401192e-07,
+      "loss": 0.0003545734565705061,
+      "step": 1540
+    },
+    {
+      "epoch": 4.970083765456721,
+      "grad_norm": 0.01692381501197815,
+      "learning_rate": 1.1209967689925832e-07,
+      "loss": 0.0004393013659864664,
+      "step": 1560
+    },
+    {
+      "epoch": 5.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.736,
+      "eval_generator_rouge1": 99.9082,
+      "eval_generator_rouge2": 99.8294,
+      "eval_generator_rougeL": 99.7068,
+      "eval_loss": 0.004025007598102093,
+      "eval_runtime": 591.5846,
+      "eval_samples_per_second": 5.771,
+      "eval_selector_accuracy": 98.6565,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 1570
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.871444840310374e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/amid/checkpoint-1570/training_args.bin b/qwen3/amid/checkpoint-1570/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..571568fd8ab61d629f2a731739ee4c9e1309cca4
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e67133ff7a80f6f166c03bd5f62d9c580548b0f6840a1ce3891dee623991e47a
+size 7185
diff --git a/qwen3/amid/checkpoint-1570/zero_to_fp32.py b/qwen3/amid/checkpoint-1570/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/amid/checkpoint-1570/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/amid/checkpoint-314/README.md b/qwen3/amid/checkpoint-314/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-314/adapter_config.json b/qwen3/amid/checkpoint-314/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..16362f47190537658d6b04519de786417a5f6c2f
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-314/adapter_model.safetensors b/qwen3/amid/checkpoint-314/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..b23a97434be3be41e630afe1a0af9e491893d213
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:819b56a96acbcec984ed0fb2f70a3c64d727524acc5ee849ced515dec088b2d5
+size 40422560
diff --git a/qwen3/amid/checkpoint-314/chat_template.jinja b/qwen3/amid/checkpoint-314/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..672fc2344ac3fce9a37b71054471c9f755cb4e3e
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cf04e8382c2e8daba7ce0547f28dc0bba6cbd2ab62fda8b602564525f4b12c3
+size 121135557
diff --git a/qwen3/amid/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..29a56298d21a5c7c0d6a4d0c4c44c688439cfd2e
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80b1a0a2b5fa241dbf640d8ef1bb61ffe49a12a3719b2f5832ef3ade0f0444c1
+size 121135749
diff --git a/qwen3/amid/checkpoint-314/global_step314/mp_rank_00_model_states.pt b/qwen3/amid/checkpoint-314/global_step314/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..92f5c2d640fd5581bdd0addfc56693fca2bfaad8
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/global_step314/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51f7326388d5f9abc2bd8ef5dc008c578b2cf58cd5a1b02c8cb29ce1608c1d2b
+size 351841765
diff --git a/qwen3/amid/checkpoint-314/latest b/qwen3/amid/checkpoint-314/latest
new file mode 100644
index 0000000000000000000000000000000000000000..a9786513b1a514db2c3a9da7408844635663d55d
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/latest
@@ -0,0 +1 @@
+global_step314
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-314/rng_state_0.pth b/qwen3/amid/checkpoint-314/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..c83b1c0fbf28c44631daf9c84d4dca2490a24c36
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdba92889c38478a7a838b8eb355ab882c464bbba5c4d3c73edec3f82d60d0a2
+size 14917
diff --git a/qwen3/amid/checkpoint-314/rng_state_1.pth b/qwen3/amid/checkpoint-314/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..9d3d426bd9b5bf50e3a60467650048b9c0db604d
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e89a8da6898755602708e3be195153442a313c7b82548f2de733720343c1812e
+size 14917
diff --git a/qwen3/amid/checkpoint-314/scheduler.pt b/qwen3/amid/checkpoint-314/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..a3cb460802cda03bf0833a98a6f1428a3fd4f585
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17e254b5642bb75983fd5105bec79b35376cfbc99d29702287668b8767d46149
+size 1465
diff --git a/qwen3/amid/checkpoint-314/tokenizer.json b/qwen3/amid/checkpoint-314/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/amid/checkpoint-314/tokenizer_config.json b/qwen3/amid/checkpoint-314/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/amid/checkpoint-314/trainer_state.json b/qwen3/amid/checkpoint-314/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..eff54052fb7f9d2c53302dc45c83e4ff4abee4d0
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/trainer_state.json
@@ -0,0 +1,154 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.0,
+  "eval_steps": 500,
+  "global_step": 314,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 0.38384687900543213,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.27438180446624755,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.38263317942619324,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.13717052936553956,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.08620006591081619,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.13135391473770142,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.15773959457874298,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.1076702356338501,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.22962816059589386,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.08698999881744385,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.12232966721057892,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.1128507375717163,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.15597161650657654,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.1277535080909729,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.7329031825065613,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.08885469436645507,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.45483648777008057,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.07881734371185303,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.16407009959220886,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.06113555431365967,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.07238732278347015,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.06768973469734192,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.4741341471672058,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.04917148947715759,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.36410900950431824,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.010153122991323472,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.11222824454307556,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.00552816241979599,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.47910866141319275,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.006760746240615845,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 87.3832,
+      "eval_generator_rouge1": 99.2069,
+      "eval_generator_rouge2": 98.8793,
+      "eval_generator_rougeL": 98.7553,
+      "eval_loss": 0.012839075177907944,
+      "eval_runtime": 592.0147,
+      "eval_samples_per_second": 5.767,
+      "eval_selector_accuracy": 90.1869,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 314
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.172842003562496e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/amid/checkpoint-314/training_args.bin b/qwen3/amid/checkpoint-314/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..571568fd8ab61d629f2a731739ee4c9e1309cca4
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e67133ff7a80f6f166c03bd5f62d9c580548b0f6840a1ce3891dee623991e47a
+size 7185
diff --git a/qwen3/amid/checkpoint-314/zero_to_fp32.py b/qwen3/amid/checkpoint-314/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/amid/checkpoint-314/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/amid/checkpoint-628/README.md b/qwen3/amid/checkpoint-628/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-628/adapter_config.json b/qwen3/amid/checkpoint-628/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..16362f47190537658d6b04519de786417a5f6c2f
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-628/adapter_model.safetensors b/qwen3/amid/checkpoint-628/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..eff5c2f18672c814bd1e6945688d53b0ba456d98
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c389c80927963f79f7ce42996bf58ac7aae228480a2b389697bf4ec6f1b1ca8
+size 40422560
diff --git a/qwen3/amid/checkpoint-628/chat_template.jinja b/qwen3/amid/checkpoint-628/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..b9fe3375336715538c345cf6a6785a3b7bb158fb
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f1e23dd6dcb516558f88b2f93393cbdf049e7fd6f9856cbdf98e8bd50ef685e
+size 121135557
diff --git a/qwen3/amid/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..621ea987db06a11b70b35757c04bf9e7adef162f
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8069a1dac0872d05ac2f54e1ed805c47b45310bbcad4e7ba63e22e58906a0d11
+size 121135749
diff --git a/qwen3/amid/checkpoint-628/global_step628/mp_rank_00_model_states.pt b/qwen3/amid/checkpoint-628/global_step628/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..605b010f654e21ee5503b6dc6402faf3910acec8
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/global_step628/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbf2f57dcb83df3e0f068c6e0b4767c8d1ef6a5e02ce0f6b0b5ebedb4ad97c47
+size 351841765
diff --git a/qwen3/amid/checkpoint-628/latest b/qwen3/amid/checkpoint-628/latest
new file mode 100644
index 0000000000000000000000000000000000000000..af055c3db05dcb2af541005f22308dbd946aba65
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/latest
@@ -0,0 +1 @@
+global_step628
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-628/rng_state_0.pth b/qwen3/amid/checkpoint-628/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..f3ea7d402f1624f296420a075d283430bc36cae8
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bf9f12c4ef51f269865246a370b77796f13cfe2423d7c104ba089255731e5e0
+size 14917
diff --git a/qwen3/amid/checkpoint-628/rng_state_1.pth b/qwen3/amid/checkpoint-628/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..9581325e77eefdaebabc9a563e5ac399f9155ffb
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b6d50fa9ea1e53c22b6a8e937368cbc670dadbe5385fa480760ee21b0618315
+size 14917
diff --git a/qwen3/amid/checkpoint-628/scheduler.pt b/qwen3/amid/checkpoint-628/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..96c0cabcd99be4945ffbd7897ecc5367ac4beb28
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb823a2c51d7fb42dfe3b693a7159784285dbf38c5a797e6e73292096123ab7a
+size 1465
diff --git a/qwen3/amid/checkpoint-628/tokenizer.json b/qwen3/amid/checkpoint-628/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/amid/checkpoint-628/tokenizer_config.json b/qwen3/amid/checkpoint-628/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/amid/checkpoint-628/trainer_state.json b/qwen3/amid/checkpoint-628/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..2a3ac10923096e3ff393a11e2d3f600154815aa4
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/trainer_state.json
@@ -0,0 +1,281 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.0,
+  "eval_steps": 500,
+  "global_step": 628,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 0.38384687900543213,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.27438180446624755,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.38263317942619324,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.13717052936553956,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.08620006591081619,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.13135391473770142,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.15773959457874298,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.1076702356338501,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.22962816059589386,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.08698999881744385,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.12232966721057892,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.1128507375717163,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.15597161650657654,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.1277535080909729,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.7329031825065613,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.08885469436645507,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.45483648777008057,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.07881734371185303,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.16407009959220886,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.06113555431365967,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.07238732278347015,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.06768973469734192,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.4741341471672058,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.04917148947715759,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.36410900950431824,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.010153122991323472,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.11222824454307556,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.00552816241979599,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.47910866141319275,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.006760746240615845,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 87.3832,
+      "eval_generator_rouge1": 99.2069,
+      "eval_generator_rouge2": 98.8793,
+      "eval_generator_rougeL": 98.7553,
+      "eval_loss": 0.012839075177907944,
+      "eval_runtime": 592.0147,
+      "eval_samples_per_second": 5.767,
+      "eval_selector_accuracy": 90.1869,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.27295932173728943,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.009849193692207336,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.060826949775218964,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.005673223733901977,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 0.04862668365240097,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.005297792330384255,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 0.05064748600125313,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": 0.0021365636959671972,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.05359279736876488,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": 0.00429258793592453,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.11369121074676514,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0034508116543292998,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.01604766771197319,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": 0.0028350137174129484,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.15401995182037354,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": 0.005546276643872261,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 0.5877118706703186,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": 0.0044526983052492145,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.10250785946846008,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": 0.0056363165378570555,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.043251167982816696,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": 0.0032581716775894167,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.2089434117078781,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": 0.004207362234592438,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.14037534594535828,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": 0.0019151581451296805,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 0.22862067818641663,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": 0.004427855834364891,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.03684071823954582,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": 0.0017833203077316285,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.0672929510474205,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": 0.0035866759717464447,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 93.5164,
+      "eval_generator_rouge1": 99.7294,
+      "eval_generator_rouge2": 99.5697,
+      "eval_generator_rougeL": 99.5152,
+      "eval_loss": 0.006910035386681557,
+      "eval_runtime": 591.6591,
+      "eval_samples_per_second": 5.77,
+      "eval_selector_accuracy": 96.9042,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 628
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.346351770520781e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/amid/checkpoint-628/training_args.bin b/qwen3/amid/checkpoint-628/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..571568fd8ab61d629f2a731739ee4c9e1309cca4
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e67133ff7a80f6f166c03bd5f62d9c580548b0f6840a1ce3891dee623991e47a
+size 7185
diff --git a/qwen3/amid/checkpoint-628/zero_to_fp32.py b/qwen3/amid/checkpoint-628/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/amid/checkpoint-628/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/amid/checkpoint-942/README.md b/qwen3/amid/checkpoint-942/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-942/adapter_config.json b/qwen3/amid/checkpoint-942/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..16362f47190537658d6b04519de786417a5f6c2f
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "k_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-942/adapter_model.safetensors b/qwen3/amid/checkpoint-942/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d82ae009b056b3cada54c54d53e197cf57018fc1
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42a66b61ac882bf27ec3608465ee839ade315ab4d05e90c6b29feafe1a79e596
+size 40422560
diff --git a/qwen3/amid/checkpoint-942/chat_template.jinja b/qwen3/amid/checkpoint-942/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-942/global_step942/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-942/global_step942/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..430aff02c35959ec02efb03036320b751b969159
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/global_step942/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3366587ffc75b4047977e6e15443096b8b57493fda0e89a96b367f05499adeb
+size 121135557
diff --git a/qwen3/amid/checkpoint-942/global_step942/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/amid/checkpoint-942/global_step942/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..087981fdf7d55983b4e15fe9ccaec7d77f73a91e
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/global_step942/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4858a135049b98b9aa364311c5fbf77e482bdf035726ae5962dd3f4b021d3ffa
+size 121135749
diff --git a/qwen3/amid/checkpoint-942/global_step942/mp_rank_00_model_states.pt b/qwen3/amid/checkpoint-942/global_step942/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..5122c81878b1d13a21fdbe6fb35a872f7b98fa8a
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/global_step942/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45a146f178cca0f1049469fae5f2a961745eabf87ecd1fea5549d3f80cc853e0
+size 351841765
diff --git a/qwen3/amid/checkpoint-942/latest b/qwen3/amid/checkpoint-942/latest
new file mode 100644
index 0000000000000000000000000000000000000000..4e9a68e4460aca593c480611a486d2a4cd73ef42
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/latest
@@ -0,0 +1 @@
+global_step942
\ No newline at end of file
diff --git a/qwen3/amid/checkpoint-942/rng_state_0.pth b/qwen3/amid/checkpoint-942/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..54f43bf1f663704b3073af0807e101057868ead8
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dfb868039ddf2110ae2286a4ee4bea74f00ea64076a85da4fb4086aec87ba5d
+size 14917
diff --git a/qwen3/amid/checkpoint-942/rng_state_1.pth b/qwen3/amid/checkpoint-942/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..f2fe299f679ce3632aaa6df223a4f01d59730243
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0eeb29092a9d04903a10dcd0c0f8397a7cade5f85c1d5bf69b75e46dd3a3ae3e
+size 14917
diff --git a/qwen3/amid/checkpoint-942/scheduler.pt b/qwen3/amid/checkpoint-942/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..6b1f8af0480bf83ab728abfeff11bc3288214349
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8076da6f897951c5bdf8649e4fba596e8c49819927746017442ffd85f40b18a3
+size 1465
diff --git a/qwen3/amid/checkpoint-942/tokenizer.json b/qwen3/amid/checkpoint-942/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/amid/checkpoint-942/tokenizer_config.json b/qwen3/amid/checkpoint-942/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/amid/checkpoint-942/trainer_state.json b/qwen3/amid/checkpoint-942/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..03fbdef9401c3d148fff75f43120c6c0c995e8a0
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/trainer_state.json
@@ -0,0 +1,408 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 3.0,
+  "eval_steps": 500,
+  "global_step": 942,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 0.38384687900543213,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.27438180446624755,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.38263317942619324,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.13717052936553956,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.08620006591081619,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.13135391473770142,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.15773959457874298,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.1076702356338501,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.22962816059589386,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.08698999881744385,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.12232966721057892,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.1128507375717163,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.15597161650657654,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.1277535080909729,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.7329031825065613,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.08885469436645507,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.45483648777008057,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.07881734371185303,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.16407009959220886,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.06113555431365967,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.07238732278347015,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.06768973469734192,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.4741341471672058,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.04917148947715759,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.36410900950431824,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.010153122991323472,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.11222824454307556,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.00552816241979599,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.47910866141319275,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.006760746240615845,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 87.3832,
+      "eval_generator_rouge1": 99.2069,
+      "eval_generator_rouge2": 98.8793,
+      "eval_generator_rougeL": 98.7553,
+      "eval_loss": 0.012839075177907944,
+      "eval_runtime": 592.0147,
+      "eval_samples_per_second": 5.767,
+      "eval_selector_accuracy": 90.1869,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.27295932173728943,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.009849193692207336,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.060826949775218964,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.005673223733901977,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 0.04862668365240097,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.005297792330384255,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 0.05064748600125313,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": 0.0021365636959671972,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.05359279736876488,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": 0.00429258793592453,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.11369121074676514,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0034508116543292998,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.01604766771197319,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": 0.0028350137174129484,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.15401995182037354,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": 0.005546276643872261,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 0.5877118706703186,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": 0.0044526983052492145,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.10250785946846008,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": 0.0056363165378570555,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.043251167982816696,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": 0.0032581716775894167,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.2089434117078781,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": 0.004207362234592438,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.14037534594535828,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": 0.0019151581451296805,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 0.22862067818641663,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": 0.004427855834364891,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.03684071823954582,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": 0.0017833203077316285,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.0672929510474205,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": 0.0035866759717464447,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 93.5164,
+      "eval_generator_rouge1": 99.7294,
+      "eval_generator_rouge2": 99.5697,
+      "eval_generator_rougeL": 99.5152,
+      "eval_loss": 0.006910035386681557,
+      "eval_runtime": 591.6591,
+      "eval_samples_per_second": 5.77,
+      "eval_selector_accuracy": 96.9042,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.029134005308151245,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": 0.0020816387608647347,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.17367292940616608,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": 0.0019141122698783874,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.016447540372610092,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": 0.0009251200594007969,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.015570643357932568,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": 0.0017806822434067726,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.010291235521435738,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": 0.0010542120784521104,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.012873738072812557,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": 0.0009571929462254047,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.009928264655172825,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": 0.0007098949514329434,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.014300181530416012,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": 0.0007610605098307133,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.03284010663628578,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": 0.002758556976914406,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.19087237119674683,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": 0.0011194748803973198,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.02937050722539425,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": 0.0014943076297640801,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.031206471845507622,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": 0.0027773819863796236,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.019134322181344032,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": 0.0018997505307197572,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.2537505030632019,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": 0.0010082274675369263,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.0017131563508883119,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": 0.0012075364589691161,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.03925931081175804,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": 0.000642749946564436,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.3855,
+      "eval_generator_rouge1": 99.8982,
+      "eval_generator_rouge2": 99.8084,
+      "eval_generator_rougeL": 99.6633,
+      "eval_loss": 0.0056763943284749985,
+      "eval_runtime": 590.3317,
+      "eval_samples_per_second": 5.783,
+      "eval_selector_accuracy": 96.5537,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 942
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 3.524157743353037e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/amid/checkpoint-942/training_args.bin b/qwen3/amid/checkpoint-942/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..571568fd8ab61d629f2a731739ee4c9e1309cca4
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e67133ff7a80f6f166c03bd5f62d9c580548b0f6840a1ce3891dee623991e47a
+size 7185
diff --git a/qwen3/amid/checkpoint-942/zero_to_fp32.py b/qwen3/amid/checkpoint-942/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/amid/checkpoint-942/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/amid/eval_results.json b/qwen3/amid/eval_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..000c38aea42d8d2cd1a4372ff0a248dbdcdfdb29
--- /dev/null
+++ b/qwen3/amid/eval_results.json
@@ -0,0 +1,14 @@
+{
+    "epoch": 5.0,
+    "eval_generator_count": 1712.0,
+    "eval_generator_exact_match": 96.028,
+    "eval_generator_rouge1": 99.8671,
+    "eval_generator_rouge2": 99.8011,
+    "eval_generator_rougeL": 99.6739,
+    "eval_loss": 0.004025007598102093,
+    "eval_runtime": 579.6459,
+    "eval_samples_per_second": 5.89,
+    "eval_selector_accuracy": 98.6565,
+    "eval_selector_count": 1712.0,
+    "eval_steps_per_second": 0.185
+}
\ No newline at end of file
diff --git a/qwen3/amid/tokenizer.json b/qwen3/amid/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/amid/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/amid/tokenizer_config.json b/qwen3/amid/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/amid/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/amid/train_results.json b/qwen3/amid/train_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..cb24c11958e8a5666c3d9051ba3435bc54cc5d4b
--- /dev/null
+++ b/qwen3/amid/train_results.json
@@ -0,0 +1,8 @@
+{
+    "epoch": 5.0,
+    "total_flos": 5.871444840310374e+16,
+    "train_loss": 0.01864182253792336,
+    "train_runtime": 7621.8556,
+    "train_samples_per_second": 6.578,
+    "train_steps_per_second": 0.206
+}
\ No newline at end of file
diff --git a/qwen3/amid/trainer_state.json b/qwen3/amid/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..2cca169a60a34350c1c45267e33165ea7b674a70
--- /dev/null
+++ b/qwen3/amid/trainer_state.json
@@ -0,0 +1,664 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 5.0,
+  "eval_steps": 500,
+  "global_step": 1570,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 0.38384687900543213,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.27438180446624755,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.38263317942619324,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.13717052936553956,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.08620006591081619,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.13135391473770142,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.15773959457874298,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.1076702356338501,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.22962816059589386,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.08698999881744385,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.12232966721057892,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.1128507375717163,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.15597161650657654,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.1277535080909729,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.7329031825065613,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.08885469436645507,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.45483648777008057,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.07881734371185303,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.16407009959220886,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.06113555431365967,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.07238732278347015,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.06768973469734192,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.4741341471672058,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.04917148947715759,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.36410900950431824,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.010153122991323472,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.11222824454307556,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.00552816241979599,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.47910866141319275,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.006760746240615845,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 87.3832,
+      "eval_generator_rouge1": 99.2069,
+      "eval_generator_rouge2": 98.8793,
+      "eval_generator_rougeL": 98.7553,
+      "eval_loss": 0.012839075177907944,
+      "eval_runtime": 592.0147,
+      "eval_samples_per_second": 5.767,
+      "eval_selector_accuracy": 90.1869,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.27295932173728943,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.009849193692207336,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.060826949775218964,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.005673223733901977,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 0.04862668365240097,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.005297792330384255,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 0.05064748600125313,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": 0.0021365636959671972,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.05359279736876488,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": 0.00429258793592453,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.11369121074676514,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0034508116543292998,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.01604766771197319,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": 0.0028350137174129484,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.15401995182037354,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": 0.005546276643872261,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 0.5877118706703186,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": 0.0044526983052492145,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.10250785946846008,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": 0.0056363165378570555,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.043251167982816696,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": 0.0032581716775894167,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.2089434117078781,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": 0.004207362234592438,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.14037534594535828,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": 0.0019151581451296805,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 0.22862067818641663,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": 0.004427855834364891,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.03684071823954582,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": 0.0017833203077316285,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.0672929510474205,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": 0.0035866759717464447,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 93.5164,
+      "eval_generator_rouge1": 99.7294,
+      "eval_generator_rouge2": 99.5697,
+      "eval_generator_rougeL": 99.5152,
+      "eval_loss": 0.006910035386681557,
+      "eval_runtime": 591.6591,
+      "eval_samples_per_second": 5.77,
+      "eval_selector_accuracy": 96.9042,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.029134005308151245,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": 0.0020816387608647347,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.17367292940616608,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": 0.0019141122698783874,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.016447540372610092,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": 0.0009251200594007969,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.015570643357932568,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": 0.0017806822434067726,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.010291235521435738,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": 0.0010542120784521104,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.012873738072812557,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": 0.0009571929462254047,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.009928264655172825,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": 0.0007098949514329434,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.014300181530416012,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": 0.0007610605098307133,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.03284010663628578,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": 0.002758556976914406,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.19087237119674683,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": 0.0011194748803973198,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.02937050722539425,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": 0.0014943076297640801,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.031206471845507622,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": 0.0027773819863796236,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.019134322181344032,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": 0.0018997505307197572,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.2537505030632019,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": 0.0010082274675369263,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.0017131563508883119,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": 0.0012075364589691161,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.03925931081175804,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": 0.000642749946564436,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.3855,
+      "eval_generator_rouge1": 99.8982,
+      "eval_generator_rouge2": 99.8084,
+      "eval_generator_rougeL": 99.6633,
+      "eval_loss": 0.0056763943284749985,
+      "eval_runtime": 590.3317,
+      "eval_samples_per_second": 5.783,
+      "eval_selector_accuracy": 96.5537,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.0036900911945849657,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": 0.0012762822210788728,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.028549807146191597,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": 0.0006767791230231524,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.007299487479031086,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": 0.0012082518078386784,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.03702578693628311,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": 0.0014168912544846534,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.21273940801620483,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": 0.0012415657751262189,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.0033840558025985956,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": 0.0009985800832509994,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.020664699375629425,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": 0.0011398554779589176,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.0015457016415894032,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": 0.0016313906759023666,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.011757194995880127,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": 0.0010684640146791935,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.007703748065978289,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": 0.0006688186898827553,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.04895896837115288,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": 0.0016563836485147475,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.04278242960572243,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": 0.0006196124479174614,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.004100222606211901,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": 0.0009099396876990796,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.06636715680360794,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": 0.0012004765681922435,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.005992025602608919,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": 0.0003460923908278346,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.8528,
+      "eval_generator_rouge1": 99.9186,
+      "eval_generator_rouge2": 99.8428,
+      "eval_generator_rougeL": 99.7108,
+      "eval_loss": 0.004311855416744947,
+      "eval_runtime": 587.5171,
+      "eval_samples_per_second": 5.811,
+      "eval_selector_accuracy": 98.1893,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.182,
+      "step": 1256
+    },
+    {
+      "epoch": 4.0127642600717985,
+      "grad_norm": 0.12402024120092392,
+      "learning_rate": 9.464081721960592e-06,
+      "loss": 0.0005224554799497128,
+      "step": 1260
+    },
+    {
+      "epoch": 4.076585560430794,
+      "grad_norm": 0.032988350838422775,
+      "learning_rate": 8.331628936475818e-06,
+      "loss": 0.00022895974107086658,
+      "step": 1280
+    },
+    {
+      "epoch": 4.140406860789788,
+      "grad_norm": 0.002514337655156851,
+      "learning_rate": 7.2659843568748385e-06,
+      "loss": 0.00028093236032873393,
+      "step": 1300
+    },
+    {
+      "epoch": 4.204228161148784,
+      "grad_norm": 0.013952493667602539,
+      "learning_rate": 6.268854516770008e-06,
+      "loss": 0.0007838346064090729,
+      "step": 1320
+    },
+    {
+      "epoch": 4.268049461507778,
+      "grad_norm": 0.00595195172354579,
+      "learning_rate": 5.341836229602331e-06,
+      "loss": 0.0010836148634552956,
+      "step": 1340
+    },
+    {
+      "epoch": 4.331870761866773,
+      "grad_norm": 0.24550670385360718,
+      "learning_rate": 4.4864140314888216e-06,
+      "loss": 0.001040823757648468,
+      "step": 1360
+    },
+    {
+      "epoch": 4.395692062225768,
+      "grad_norm": 0.009249023161828518,
+      "learning_rate": 3.703957803871932e-06,
+      "loss": 0.0006210384890437126,
+      "step": 1380
+    },
+    {
+      "epoch": 4.4595133625847625,
+      "grad_norm": 0.010251481086015701,
+      "learning_rate": 2.9957205797780495e-06,
+      "loss": 0.0005129187367856503,
+      "step": 1400
+    },
+    {
+      "epoch": 4.523334662943758,
+      "grad_norm": 0.024013608694076538,
+      "learning_rate": 2.362836537198199e-06,
+      "loss": 0.0004933405201882124,
+      "step": 1420
+    },
+    {
+      "epoch": 4.587155963302752,
+      "grad_norm": 0.028239818289875984,
+      "learning_rate": 1.8063191828043703e-06,
+      "loss": 0.0005304558202624321,
+      "step": 1440
+    },
+    {
+      "epoch": 4.650977263661747,
+      "grad_norm": 0.01863635517656803,
+      "learning_rate": 1.3270597289100437e-06,
+      "loss": 0.00033774527255445717,
+      "step": 1460
+    },
+    {
+      "epoch": 4.714798564020742,
+      "grad_norm": 0.0036166764330118895,
+      "learning_rate": 9.258256662741205e-07,
+      "loss": 0.0002381577854976058,
+      "step": 1480
+    },
+    {
+      "epoch": 4.778619864379737,
+      "grad_norm": 0.016647396609187126,
+      "learning_rate": 6.032595350337341e-07,
+      "loss": 0.0005346011370420456,
+      "step": 1500
+    },
+    {
+      "epoch": 4.842441164738732,
+      "grad_norm": 0.14798367023468018,
+      "learning_rate": 3.5987789573421356e-07,
+      "loss": 0.0015266412869095803,
+      "step": 1520
+    },
+    {
+      "epoch": 4.9062624650977265,
+      "grad_norm": 0.1593816727399826,
+      "learning_rate": 1.9607050210401192e-07,
+      "loss": 0.0003545734565705061,
+      "step": 1540
+    },
+    {
+      "epoch": 4.970083765456721,
+      "grad_norm": 0.01692381501197815,
+      "learning_rate": 1.1209967689925832e-07,
+      "loss": 0.0004393013659864664,
+      "step": 1560
+    },
+    {
+      "epoch": 5.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.736,
+      "eval_generator_rouge1": 99.9082,
+      "eval_generator_rouge2": 99.8294,
+      "eval_generator_rougeL": 99.7068,
+      "eval_loss": 0.004025007598102093,
+      "eval_runtime": 591.5846,
+      "eval_samples_per_second": 5.771,
+      "eval_selector_accuracy": 98.6565,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 1570
+    },
+    {
+      "epoch": 5.0,
+      "step": 1570,
+      "total_flos": 5.871444840310374e+16,
+      "train_loss": 0.01864182253792336,
+      "train_runtime": 7621.8556,
+      "train_samples_per_second": 6.578,
+      "train_steps_per_second": 0.206
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.871444840310374e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/amid/training_args.bin b/qwen3/amid/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..571568fd8ab61d629f2a731739ee4c9e1309cca4
--- /dev/null
+++ b/qwen3/amid/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e67133ff7a80f6f166c03bd5f62d9c580548b0f6840a1ce3891dee623991e47a
+size 7185
diff --git a/qwen3/csd/README.md b/qwen3/csd/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..aecfa20e01ac21ca72d6d48d31f51cc23fde71fe
--- /dev/null
+++ b/qwen3/csd/README.md
@@ -0,0 +1,79 @@
+---
+library_name: peft
+license: other
+base_model: Qwen/Qwen3-0.6B
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+pipeline_tag: text-generation
+model-index:
+- name: csd
+  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. -->
+
+# csd
+
+This model is a fine-tuned version of [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) on the cypher_prepared_train dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.0041
+- Selector Count: 1712.0
+- Generator Count: 1712.0
+- Selector Accuracy: 98.5981
+- Generator Exact Match: 96.6121
+- Generator Rouge1: 99.9371
+- Generator Rouge2: 99.8891
+- Generator Rougel: 99.7571
+
+## 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: 0.0001
+- train_batch_size: 2
+- eval_batch_size: 16
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 2
+- gradient_accumulation_steps: 8
+- total_train_batch_size: 32
+- total_eval_batch_size: 32
+- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine_with_min_lr
+- num_epochs: 5
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss | Selector Count | Generator Count | Selector Accuracy | Generator Exact Match | Generator Rouge1 | Generator Rouge2 | Generator Rougel |
+|:-------------:|:-----:|:----:|:---------------:|:--------------:|:---------------:|:-----------------:|:---------------------:|:----------------:|:----------------:|:----------------:|
+| 0.0050        | 1.0   | 314  | 0.0192          | 1712.0         | 1712.0          | 93.3995           | 86.2734               | 99.1343          | 98.8297          | 98.6074          |
+| -0.0061       | 2.0   | 628  | 0.0068          | 1712.0         | 1712.0          | 97.8388           | 94.6846               | 99.7652          | 99.667           | 99.5471          |
+| -0.0188       | 3.0   | 942  | 0.0053          | 1712.0         | 1712.0          | 97.722            | 95.736                | 99.9465          | 99.8866          | 99.7132          |
+| -0.0361       | 4.0   | 1256 | 0.0042          | 1712.0         | 1712.0          | 98.3645           | 96.2617               | 99.8812          | 99.8259          | 99.6849          |
+| -0.0416       | 5.0   | 1570 | 0.0041          | 1712.0         | 1712.0          | 98.6565           | 96.6706               | 99.9491          | 99.8942          | 99.7857          |
+
+
+### Framework versions
+
+- PEFT 0.18.1
+- Transformers 5.8.0
+- Pytorch 2.9.1+cu128
+- Datasets 4.0.0
+- Tokenizers 0.22.2
\ No newline at end of file
diff --git a/qwen3/csd/adapter_config.json b/qwen3/csd/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9959003652c9aaae4cece005b46aaf87e6d0f224
--- /dev/null
+++ b/qwen3/csd/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "up_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "o_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/csd/adapter_model.safetensors b/qwen3/csd/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..57ad3d3cb4fe71a4d7f36999029d2b5e7ba925c6
--- /dev/null
+++ b/qwen3/csd/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4406b594fb337274c887bfad3ad274f42413413cfb01fb17f61a907e23a5aca
+size 40422560
diff --git a/qwen3/csd/all_results.json b/qwen3/csd/all_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..908a5e1203e65961ea2e79bf98da56f598aafe74
--- /dev/null
+++ b/qwen3/csd/all_results.json
@@ -0,0 +1,19 @@
+{
+    "epoch": 5.0,
+    "eval_generator_count": 1712.0,
+    "eval_generator_exact_match": 96.6121,
+    "eval_generator_rouge1": 99.9371,
+    "eval_generator_rouge2": 99.8891,
+    "eval_generator_rougeL": 99.7571,
+    "eval_loss": 0.00408697547391057,
+    "eval_runtime": 585.4618,
+    "eval_samples_per_second": 5.831,
+    "eval_selector_accuracy": 98.5981,
+    "eval_selector_count": 1712.0,
+    "eval_steps_per_second": 0.183,
+    "total_flos": 5.871444840310374e+16,
+    "train_loss": -0.013058029026516779,
+    "train_runtime": 7965.6329,
+    "train_samples_per_second": 6.294,
+    "train_steps_per_second": 0.197
+}
\ No newline at end of file
diff --git a/qwen3/csd/chat_template.jinja b/qwen3/csd/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/csd/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1256/README.md b/qwen3/csd/checkpoint-1256/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1256/adapter_config.json b/qwen3/csd/checkpoint-1256/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9959003652c9aaae4cece005b46aaf87e6d0f224
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "up_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "o_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1256/adapter_model.safetensors b/qwen3/csd/checkpoint-1256/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..b85b79f15631bffb3a36568946231c30f5ae1ed7
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a383f12c8d4db1e02f37692723d39177089a0b1b09d598ae61ae9426efbd636d
+size 40422560
diff --git a/qwen3/csd/checkpoint-1256/chat_template.jinja b/qwen3/csd/checkpoint-1256/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..6b188dc8c0d02140d1df35c04a75c2d9216379da
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:227e3b62ef4c65d8e242a7eeb6ddd96c65729363487d6a9c33822628441b6536
+size 121135557
diff --git a/qwen3/csd/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..637ca16f174ab2ad45ac986d6fd8ae4e28883b86
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d301a9eb5a44c64fb9dea101f07e48b82c5fa08985e9d57e1e2bfd2c25c951a4
+size 121135749
diff --git a/qwen3/csd/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt b/qwen3/csd/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..e8ee839ba70ea1beff0a62ccab838366bd7a35c2
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b75ae1ee7c799972ebedcb20fb08609db04762d45a3030f24f465c8605eeb8a9
+size 351841765
diff --git a/qwen3/csd/checkpoint-1256/latest b/qwen3/csd/checkpoint-1256/latest
new file mode 100644
index 0000000000000000000000000000000000000000..74abb3d23ea83ebdac59918752db5d1782834ca7
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/latest
@@ -0,0 +1 @@
+global_step1256
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1256/rng_state_0.pth b/qwen3/csd/checkpoint-1256/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..f71a7dc0719c09b17fc6c626e62684f6ce78877f
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53905562edd591f558a2d5e21b3ce7bcf87bff6ea468e990980a48731ed07327
+size 14917
diff --git a/qwen3/csd/checkpoint-1256/rng_state_1.pth b/qwen3/csd/checkpoint-1256/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..2f8498e702b57f175b94b9411aacb88c870385bf
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f73798ef6ab5a59b84975e3c8385db7b9dde1508715924c2c18cba2d5c8ac9c2
+size 14917
diff --git a/qwen3/csd/checkpoint-1256/scheduler.pt b/qwen3/csd/checkpoint-1256/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..40e863d07cfed22cd912014eb3ba3fd2b16b173f
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1775297f14e75f7463ae51983a66296f81f4b2e2e5d7234f1a6d295abfa4cf23
+size 1465
diff --git a/qwen3/csd/checkpoint-1256/tokenizer.json b/qwen3/csd/checkpoint-1256/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/csd/checkpoint-1256/tokenizer_config.json b/qwen3/csd/checkpoint-1256/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/csd/checkpoint-1256/trainer_state.json b/qwen3/csd/checkpoint-1256/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..9f637fb733126d725c0abe0e8f2bd2ac8b8cc7ab
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/trainer_state.json
@@ -0,0 +1,528 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 4.0,
+  "eval_steps": 500,
+  "global_step": 1256,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 6.0161848068237305,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.11036398410797119,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 4.099686622619629,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.01695394217967987,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 3.3755345344543457,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.03482693731784821,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 1.194334864616394,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.027955988049507143,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 6.6159281730651855,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.005286623537540436,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 2.851339817047119,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.01179061084985733,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.989425539970398,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.006131124496459961,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 1.3493149280548096,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.007052437216043472,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.9337753057479858,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": -0.002346178889274597,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 1.5435981750488281,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.013309524953365326,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.7995954155921936,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.008521870523691178,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 1.6571325063705444,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.008653326332569123,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.519936203956604,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.004485934227705002,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.31598174571990967,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.003059915825724602,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.48084524273872375,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.005025627836585045,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 86.2734,
+      "eval_generator_rouge1": 99.1343,
+      "eval_generator_rouge2": 98.8297,
+      "eval_generator_rougeL": 98.6074,
+      "eval_loss": 0.019235465675592422,
+      "eval_runtime": 603.0951,
+      "eval_samples_per_second": 5.661,
+      "eval_selector_accuracy": 93.3995,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.177,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.34810933470726013,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.008805014193058014,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.19888198375701904,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.0013909661211073399,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 1.0506941080093384,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.0005901294760406017,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 2.363455057144165,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": -0.0003092627972364426,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.0766368880867958,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": -0.002663038857281208,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.1551397442817688,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0026730677112936974,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.7773767113685608,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": -0.0011709232814610005,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.42017874121665955,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": -0.0019059792160987855,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 1.0544211864471436,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": -0.004772515594959259,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.4706478416919708,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": -0.0002438830444589257,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.22088219225406647,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": -0.004391691088676453,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.5972660183906555,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": -0.007541544735431671,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.19400978088378906,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": -0.004146666824817657,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 1.2469046115875244,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": -0.0019529350101947784,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.14316241443157196,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": -0.008291246742010117,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.4657886326313019,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": -0.006107022240757942,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 94.6846,
+      "eval_generator_rouge1": 99.7652,
+      "eval_generator_rouge2": 99.667,
+      "eval_generator_rougeL": 99.5471,
+      "eval_loss": 0.006834100466221571,
+      "eval_runtime": 601.0294,
+      "eval_samples_per_second": 5.68,
+      "eval_selector_accuracy": 97.8388,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.22094473242759705,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": -0.005953450500965118,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.5406356453895569,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": -0.0025423141196370125,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.09373176842927933,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": -0.0044939789921045305,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.1080167219042778,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": -0.00820673182606697,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.05399267002940178,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": -0.009635555744171142,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.2048628032207489,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": -0.010086850821971893,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.08765007555484772,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": -0.008542184531688691,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.2229253351688385,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": -0.011658547818660736,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.18513049185276031,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": -0.00930779054760933,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.3498702347278595,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": -0.014870831370353698,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.16160713136196136,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": -0.020250825583934783,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.16046860814094543,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": -0.018617722392082214,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.17135752737522125,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": -0.020963044464588167,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.24867695569992065,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": -0.020418259501457214,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.12315966188907623,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": -0.018548570573329926,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.14333273470401764,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": -0.018815448880195616,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.736,
+      "eval_generator_rouge1": 99.9465,
+      "eval_generator_rouge2": 99.8866,
+      "eval_generator_rougeL": 99.7132,
+      "eval_loss": 0.005336109548807144,
+      "eval_runtime": 608.1178,
+      "eval_samples_per_second": 5.614,
+      "eval_selector_accuracy": 97.722,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.176,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.1492357701063156,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": -0.02238200604915619,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.0778684988617897,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": -0.02527567744255066,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.07070758193731308,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": -0.02356054335832596,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.10038785636425018,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": -0.019069214165210725,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.33050113916397095,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": -0.022282478213310242,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.12148360908031464,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": -0.02368807643651962,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.16367903351783752,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": -0.02528309226036072,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.0877315029501915,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": -0.026697894930839537,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.06676916033029556,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": -0.028009063005447386,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.09503277391195297,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": -0.0323164701461792,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.08010950684547424,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": -0.03129267990589142,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.15882308781147003,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": -0.03188144564628601,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.07064881920814514,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": -0.03279578983783722,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.24478258192539215,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": -0.03418469429016113,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.09533516317605972,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": -0.036141425371170044,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.2617,
+      "eval_generator_rouge1": 99.8812,
+      "eval_generator_rouge2": 99.8259,
+      "eval_generator_rougeL": 99.6849,
+      "eval_loss": 0.0041860658675432205,
+      "eval_runtime": 606.5994,
+      "eval_samples_per_second": 5.628,
+      "eval_selector_accuracy": 98.3645,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.176,
+      "step": 1256
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 4.697643190059008e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/csd/checkpoint-1256/training_args.bin b/qwen3/csd/checkpoint-1256/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4b4d426a661bf64c99e9ff7730557b24ebf0e60c
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e0687327fbf0a1cbac0cde783b20e396ed970bb1eed13dcf33e672089c5f44
+size 7185
diff --git a/qwen3/csd/checkpoint-1256/zero_to_fp32.py b/qwen3/csd/checkpoint-1256/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/csd/checkpoint-1256/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/csd/checkpoint-1570/README.md b/qwen3/csd/checkpoint-1570/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1570/adapter_config.json b/qwen3/csd/checkpoint-1570/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9959003652c9aaae4cece005b46aaf87e6d0f224
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "up_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "o_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1570/adapter_model.safetensors b/qwen3/csd/checkpoint-1570/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..57ad3d3cb4fe71a4d7f36999029d2b5e7ba925c6
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4406b594fb337274c887bfad3ad274f42413413cfb01fb17f61a907e23a5aca
+size 40422560
diff --git a/qwen3/csd/checkpoint-1570/chat_template.jinja b/qwen3/csd/checkpoint-1570/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..fd9527739658e30908693b44d17300d100d06c78
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f526c3c23ccaa39490051bd670107dfeaff6c30303d8ef087b0d3c4445db877b
+size 121135557
diff --git a/qwen3/csd/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..e5988f596beaba9f11b34f411b9203f01b45d79a
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f06c88d8898c46f359073bab03b26000345de6d57b6df8cc7253c0fc3ce00c1
+size 121135749
diff --git a/qwen3/csd/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt b/qwen3/csd/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..2edd69cc73af7ebbb3495460dfbd834aa27970b3
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27dc4b35b498bb01ea52e73c324084875d55b19be49beb4afe8c9de2329148bf
+size 351841765
diff --git a/qwen3/csd/checkpoint-1570/latest b/qwen3/csd/checkpoint-1570/latest
new file mode 100644
index 0000000000000000000000000000000000000000..18c2fcff58fa6daa0a2e7c3a015552a69085778b
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/latest
@@ -0,0 +1 @@
+global_step1570
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-1570/rng_state_0.pth b/qwen3/csd/checkpoint-1570/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..082dc13ddd57b832c2b8b2ef156ed202eb5ff879
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e473b79bd0bf657ca035b0fcf6dccc1812c2cd6097dfcd0e7d6b7c3eca4e823
+size 14917
diff --git a/qwen3/csd/checkpoint-1570/rng_state_1.pth b/qwen3/csd/checkpoint-1570/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..8e2f5a7cefae455902b8f7d01023194cd292e775
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97d16f50612975ec353b47b7227fe03226cb18ce7ed72b85448e93b7059d5f45
+size 14917
diff --git a/qwen3/csd/checkpoint-1570/scheduler.pt b/qwen3/csd/checkpoint-1570/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..e22954463492d8a070ae427c95418df61893c54c
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6332fac17ae99a07597512c705658da2fd8429d1c39bb776b30e107c2cb7aae
+size 1465
diff --git a/qwen3/csd/checkpoint-1570/tokenizer.json b/qwen3/csd/checkpoint-1570/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/csd/checkpoint-1570/tokenizer_config.json b/qwen3/csd/checkpoint-1570/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/csd/checkpoint-1570/trainer_state.json b/qwen3/csd/checkpoint-1570/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..b233a2ea933f8c557d54743d3eee5cb00ec7191d
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/trainer_state.json
@@ -0,0 +1,655 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 5.0,
+  "eval_steps": 500,
+  "global_step": 1570,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 6.0161848068237305,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.11036398410797119,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 4.099686622619629,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.01695394217967987,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 3.3755345344543457,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.03482693731784821,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 1.194334864616394,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.027955988049507143,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 6.6159281730651855,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.005286623537540436,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 2.851339817047119,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.01179061084985733,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.989425539970398,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.006131124496459961,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 1.3493149280548096,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.007052437216043472,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.9337753057479858,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": -0.002346178889274597,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 1.5435981750488281,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.013309524953365326,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.7995954155921936,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.008521870523691178,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 1.6571325063705444,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.008653326332569123,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.519936203956604,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.004485934227705002,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.31598174571990967,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.003059915825724602,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.48084524273872375,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.005025627836585045,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 86.2734,
+      "eval_generator_rouge1": 99.1343,
+      "eval_generator_rouge2": 98.8297,
+      "eval_generator_rougeL": 98.6074,
+      "eval_loss": 0.019235465675592422,
+      "eval_runtime": 603.0951,
+      "eval_samples_per_second": 5.661,
+      "eval_selector_accuracy": 93.3995,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.177,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.34810933470726013,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.008805014193058014,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.19888198375701904,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.0013909661211073399,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 1.0506941080093384,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.0005901294760406017,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 2.363455057144165,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": -0.0003092627972364426,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.0766368880867958,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": -0.002663038857281208,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.1551397442817688,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0026730677112936974,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.7773767113685608,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": -0.0011709232814610005,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.42017874121665955,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": -0.0019059792160987855,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 1.0544211864471436,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": -0.004772515594959259,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.4706478416919708,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": -0.0002438830444589257,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.22088219225406647,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": -0.004391691088676453,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.5972660183906555,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": -0.007541544735431671,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.19400978088378906,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": -0.004146666824817657,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 1.2469046115875244,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": -0.0019529350101947784,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.14316241443157196,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": -0.008291246742010117,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.4657886326313019,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": -0.006107022240757942,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 94.6846,
+      "eval_generator_rouge1": 99.7652,
+      "eval_generator_rouge2": 99.667,
+      "eval_generator_rougeL": 99.5471,
+      "eval_loss": 0.006834100466221571,
+      "eval_runtime": 601.0294,
+      "eval_samples_per_second": 5.68,
+      "eval_selector_accuracy": 97.8388,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.22094473242759705,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": -0.005953450500965118,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.5406356453895569,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": -0.0025423141196370125,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.09373176842927933,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": -0.0044939789921045305,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.1080167219042778,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": -0.00820673182606697,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.05399267002940178,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": -0.009635555744171142,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.2048628032207489,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": -0.010086850821971893,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.08765007555484772,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": -0.008542184531688691,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.2229253351688385,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": -0.011658547818660736,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.18513049185276031,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": -0.00930779054760933,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.3498702347278595,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": -0.014870831370353698,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.16160713136196136,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": -0.020250825583934783,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.16046860814094543,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": -0.018617722392082214,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.17135752737522125,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": -0.020963044464588167,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.24867695569992065,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": -0.020418259501457214,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.12315966188907623,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": -0.018548570573329926,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.14333273470401764,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": -0.018815448880195616,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.736,
+      "eval_generator_rouge1": 99.9465,
+      "eval_generator_rouge2": 99.8866,
+      "eval_generator_rougeL": 99.7132,
+      "eval_loss": 0.005336109548807144,
+      "eval_runtime": 608.1178,
+      "eval_samples_per_second": 5.614,
+      "eval_selector_accuracy": 97.722,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.176,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.1492357701063156,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": -0.02238200604915619,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.0778684988617897,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": -0.02527567744255066,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.07070758193731308,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": -0.02356054335832596,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.10038785636425018,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": -0.019069214165210725,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.33050113916397095,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": -0.022282478213310242,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.12148360908031464,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": -0.02368807643651962,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.16367903351783752,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": -0.02528309226036072,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.0877315029501915,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": -0.026697894930839537,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.06676916033029556,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": -0.028009063005447386,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.09503277391195297,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": -0.0323164701461792,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.08010950684547424,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": -0.03129267990589142,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.15882308781147003,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": -0.03188144564628601,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.07064881920814514,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": -0.03279578983783722,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.24478258192539215,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": -0.03418469429016113,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.09533516317605972,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": -0.036141425371170044,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.2617,
+      "eval_generator_rouge1": 99.8812,
+      "eval_generator_rouge2": 99.8259,
+      "eval_generator_rougeL": 99.6849,
+      "eval_loss": 0.0041860658675432205,
+      "eval_runtime": 606.5994,
+      "eval_samples_per_second": 5.628,
+      "eval_selector_accuracy": 98.3645,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.176,
+      "step": 1256
+    },
+    {
+      "epoch": 4.0127642600717985,
+      "grad_norm": 0.527934730052948,
+      "learning_rate": 9.464081721960592e-06,
+      "loss": -0.034009939432144164,
+      "step": 1260
+    },
+    {
+      "epoch": 4.076585560430794,
+      "grad_norm": 0.09058220684528351,
+      "learning_rate": 8.331628936475818e-06,
+      "loss": -0.037159329652786253,
+      "step": 1280
+    },
+    {
+      "epoch": 4.140406860789788,
+      "grad_norm": 0.05895598232746124,
+      "learning_rate": 7.2659843568748385e-06,
+      "loss": -0.0379990428686142,
+      "step": 1300
+    },
+    {
+      "epoch": 4.204228161148784,
+      "grad_norm": 0.0801444798707962,
+      "learning_rate": 6.268854516770008e-06,
+      "loss": -0.03669540882110596,
+      "step": 1320
+    },
+    {
+      "epoch": 4.268049461507778,
+      "grad_norm": 0.08857632428407669,
+      "learning_rate": 5.341836229602331e-06,
+      "loss": -0.03779337704181671,
+      "step": 1340
+    },
+    {
+      "epoch": 4.331870761866773,
+      "grad_norm": 0.07376661151647568,
+      "learning_rate": 4.4864140314888216e-06,
+      "loss": -0.03757229745388031,
+      "step": 1360
+    },
+    {
+      "epoch": 4.395692062225768,
+      "grad_norm": 0.08888167142868042,
+      "learning_rate": 3.703957803871932e-06,
+      "loss": -0.03948601186275482,
+      "step": 1380
+    },
+    {
+      "epoch": 4.4595133625847625,
+      "grad_norm": 0.12080858647823334,
+      "learning_rate": 2.9957205797780495e-06,
+      "loss": -0.039844200015068054,
+      "step": 1400
+    },
+    {
+      "epoch": 4.523334662943758,
+      "grad_norm": 0.16051147878170013,
+      "learning_rate": 2.362836537198199e-06,
+      "loss": -0.0381706565618515,
+      "step": 1420
+    },
+    {
+      "epoch": 4.587155963302752,
+      "grad_norm": 0.06491488963365555,
+      "learning_rate": 1.8063191828043703e-06,
+      "loss": -0.04070441722869873,
+      "step": 1440
+    },
+    {
+      "epoch": 4.650977263661747,
+      "grad_norm": 0.10599425435066223,
+      "learning_rate": 1.3270597289100437e-06,
+      "loss": -0.03966156244277954,
+      "step": 1460
+    },
+    {
+      "epoch": 4.714798564020742,
+      "grad_norm": 0.07826361805200577,
+      "learning_rate": 9.258256662741205e-07,
+      "loss": -0.040459868311882016,
+      "step": 1480
+    },
+    {
+      "epoch": 4.778619864379737,
+      "grad_norm": 0.1433020979166031,
+      "learning_rate": 6.032595350337341e-07,
+      "loss": -0.04145862460136414,
+      "step": 1500
+    },
+    {
+      "epoch": 4.842441164738732,
+      "grad_norm": 0.1149950921535492,
+      "learning_rate": 3.5987789573421356e-07,
+      "loss": -0.03695116639137268,
+      "step": 1520
+    },
+    {
+      "epoch": 4.9062624650977265,
+      "grad_norm": 0.15655829012393951,
+      "learning_rate": 1.9607050210401192e-07,
+      "loss": -0.03868899643421173,
+      "step": 1540
+    },
+    {
+      "epoch": 4.970083765456721,
+      "grad_norm": 0.1153496727347374,
+      "learning_rate": 1.1209967689925832e-07,
+      "loss": -0.04158148169517517,
+      "step": 1560
+    },
+    {
+      "epoch": 5.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.6706,
+      "eval_generator_rouge1": 99.9491,
+      "eval_generator_rouge2": 99.8942,
+      "eval_generator_rougeL": 99.7857,
+      "eval_loss": 0.00408697547391057,
+      "eval_runtime": 602.4809,
+      "eval_samples_per_second": 5.667,
+      "eval_selector_accuracy": 98.6565,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 1570
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.871444840310374e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/csd/checkpoint-1570/training_args.bin b/qwen3/csd/checkpoint-1570/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4b4d426a661bf64c99e9ff7730557b24ebf0e60c
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e0687327fbf0a1cbac0cde783b20e396ed970bb1eed13dcf33e672089c5f44
+size 7185
diff --git a/qwen3/csd/checkpoint-1570/zero_to_fp32.py b/qwen3/csd/checkpoint-1570/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/csd/checkpoint-1570/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/csd/checkpoint-314/README.md b/qwen3/csd/checkpoint-314/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-314/adapter_config.json b/qwen3/csd/checkpoint-314/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9959003652c9aaae4cece005b46aaf87e6d0f224
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "up_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "o_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-314/adapter_model.safetensors b/qwen3/csd/checkpoint-314/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9f36257165d7e784154c2d7a4d688b2ba006f2f4
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebd48a4d1bd54ec404bfba54b062fb4fa983e40c96d41af40783c5cb4aad657e
+size 40422560
diff --git a/qwen3/csd/checkpoint-314/chat_template.jinja b/qwen3/csd/checkpoint-314/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..05cbd8920430bc10062224b8016a59aae1751a81
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ce8d1702f1514699432b046f0ef58b2f04ea1b82291ba4f0b1b546a6d1b35c3
+size 121135557
diff --git a/qwen3/csd/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..78cb9c5f7bc802f3b31fffe2ad92054f9835fc71
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa07716581ebb388af7d9c9203b7394be2b7185ded218f2634e8fe00a84cd88a
+size 121135749
diff --git a/qwen3/csd/checkpoint-314/global_step314/mp_rank_00_model_states.pt b/qwen3/csd/checkpoint-314/global_step314/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..6eb375f7716d95229e922481bc9082283af06320
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/global_step314/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59c7fa30945e5c87c2563ea4376bc682c8d6a00241fefb4e343232c29e718819
+size 351841765
diff --git a/qwen3/csd/checkpoint-314/latest b/qwen3/csd/checkpoint-314/latest
new file mode 100644
index 0000000000000000000000000000000000000000..a9786513b1a514db2c3a9da7408844635663d55d
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/latest
@@ -0,0 +1 @@
+global_step314
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-314/rng_state_0.pth b/qwen3/csd/checkpoint-314/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..3b6097fde6a02b320536c7bdde68856bb00040e0
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d03899d91988ccdf5b43b8e389e1d59f52e564bf6a9bf1856760aa48aeb981e7
+size 14917
diff --git a/qwen3/csd/checkpoint-314/rng_state_1.pth b/qwen3/csd/checkpoint-314/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..17250b5b460827c4940a09b0d33317c023323af3
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8934ac8fd8d0950e8dbb853693c27862f9c724e4b15f7094c963a73102efe9b1
+size 14917
diff --git a/qwen3/csd/checkpoint-314/scheduler.pt b/qwen3/csd/checkpoint-314/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..a3cb460802cda03bf0833a98a6f1428a3fd4f585
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17e254b5642bb75983fd5105bec79b35376cfbc99d29702287668b8767d46149
+size 1465
diff --git a/qwen3/csd/checkpoint-314/tokenizer.json b/qwen3/csd/checkpoint-314/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/csd/checkpoint-314/tokenizer_config.json b/qwen3/csd/checkpoint-314/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/csd/checkpoint-314/trainer_state.json b/qwen3/csd/checkpoint-314/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..8ccbed4ae3a3f19b35d60739e7985899cd7d50b1
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/trainer_state.json
@@ -0,0 +1,154 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.0,
+  "eval_steps": 500,
+  "global_step": 314,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 6.0161848068237305,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.11036398410797119,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 4.099686622619629,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.01695394217967987,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 3.3755345344543457,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.03482693731784821,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 1.194334864616394,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.027955988049507143,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 6.6159281730651855,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.005286623537540436,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 2.851339817047119,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.01179061084985733,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.989425539970398,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.006131124496459961,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 1.3493149280548096,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.007052437216043472,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.9337753057479858,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": -0.002346178889274597,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 1.5435981750488281,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.013309524953365326,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.7995954155921936,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.008521870523691178,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 1.6571325063705444,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.008653326332569123,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.519936203956604,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.004485934227705002,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.31598174571990967,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.003059915825724602,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.48084524273872375,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.005025627836585045,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 86.2734,
+      "eval_generator_rouge1": 99.1343,
+      "eval_generator_rouge2": 98.8297,
+      "eval_generator_rougeL": 98.6074,
+      "eval_loss": 0.019235465675592422,
+      "eval_runtime": 603.0951,
+      "eval_samples_per_second": 5.661,
+      "eval_selector_accuracy": 93.3995,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.177,
+      "step": 314
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.172842003562496e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/csd/checkpoint-314/training_args.bin b/qwen3/csd/checkpoint-314/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4b4d426a661bf64c99e9ff7730557b24ebf0e60c
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e0687327fbf0a1cbac0cde783b20e396ed970bb1eed13dcf33e672089c5f44
+size 7185
diff --git a/qwen3/csd/checkpoint-314/zero_to_fp32.py b/qwen3/csd/checkpoint-314/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/csd/checkpoint-314/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/csd/checkpoint-628/README.md b/qwen3/csd/checkpoint-628/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-628/adapter_config.json b/qwen3/csd/checkpoint-628/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9959003652c9aaae4cece005b46aaf87e6d0f224
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "up_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "o_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-628/adapter_model.safetensors b/qwen3/csd/checkpoint-628/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..ec3c97587cc829db63bb240f54c9a0cf52ec20c2
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9609ac073c9317ba19f5183dd90b8f0715e88ecc3e43641751d94edab30627a7
+size 40422560
diff --git a/qwen3/csd/checkpoint-628/chat_template.jinja b/qwen3/csd/checkpoint-628/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..af65e2889606f52ccb404bc4fdcb7944d36b5e9f
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08c79db1fbeb51b289aec741be1778f94f7c6acf51342178e850f20230dc8014
+size 121135557
diff --git a/qwen3/csd/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..93f632bc9599d692e2b69443b86b021134493db3
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e4ad0a0f5046f675230dda0c95b0b6188dad500ad126197b27a3114fbd4ed9c
+size 121135749
diff --git a/qwen3/csd/checkpoint-628/global_step628/mp_rank_00_model_states.pt b/qwen3/csd/checkpoint-628/global_step628/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..9f6de6beaddecd21b1ec5671ce2c981729cb8527
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/global_step628/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4fde01b7832da220a471ca8dee9501fde630f156f350360fd013c021b5a9413
+size 351841765
diff --git a/qwen3/csd/checkpoint-628/latest b/qwen3/csd/checkpoint-628/latest
new file mode 100644
index 0000000000000000000000000000000000000000..af055c3db05dcb2af541005f22308dbd946aba65
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/latest
@@ -0,0 +1 @@
+global_step628
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-628/rng_state_0.pth b/qwen3/csd/checkpoint-628/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..049cafc473aaf75140ed892b0788e082252fd13b
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4891a7ab4bb37eee26320259e8915c51843a2b2e8bdf0f9b6411c3aab0364aaf
+size 14917
diff --git a/qwen3/csd/checkpoint-628/rng_state_1.pth b/qwen3/csd/checkpoint-628/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..32dc1962680376cbde07b14cc66a5c038d431239
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bbf6452fd7043dc78650fd01fb93174fa7facae52d6322b9f296b20a0c8daa3
+size 14917
diff --git a/qwen3/csd/checkpoint-628/scheduler.pt b/qwen3/csd/checkpoint-628/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..96c0cabcd99be4945ffbd7897ecc5367ac4beb28
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb823a2c51d7fb42dfe3b693a7159784285dbf38c5a797e6e73292096123ab7a
+size 1465
diff --git a/qwen3/csd/checkpoint-628/tokenizer.json b/qwen3/csd/checkpoint-628/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/csd/checkpoint-628/tokenizer_config.json b/qwen3/csd/checkpoint-628/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/csd/checkpoint-628/trainer_state.json b/qwen3/csd/checkpoint-628/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..44eac57dfb5a590e892fdcc04b48c04ccaf36616
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/trainer_state.json
@@ -0,0 +1,281 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.0,
+  "eval_steps": 500,
+  "global_step": 628,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 6.0161848068237305,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.11036398410797119,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 4.099686622619629,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.01695394217967987,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 3.3755345344543457,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.03482693731784821,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 1.194334864616394,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.027955988049507143,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 6.6159281730651855,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.005286623537540436,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 2.851339817047119,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.01179061084985733,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.989425539970398,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.006131124496459961,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 1.3493149280548096,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.007052437216043472,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.9337753057479858,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": -0.002346178889274597,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 1.5435981750488281,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.013309524953365326,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.7995954155921936,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.008521870523691178,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 1.6571325063705444,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.008653326332569123,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.519936203956604,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.004485934227705002,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.31598174571990967,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.003059915825724602,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.48084524273872375,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.005025627836585045,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 86.2734,
+      "eval_generator_rouge1": 99.1343,
+      "eval_generator_rouge2": 98.8297,
+      "eval_generator_rougeL": 98.6074,
+      "eval_loss": 0.019235465675592422,
+      "eval_runtime": 603.0951,
+      "eval_samples_per_second": 5.661,
+      "eval_selector_accuracy": 93.3995,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.177,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.34810933470726013,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.008805014193058014,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.19888198375701904,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.0013909661211073399,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 1.0506941080093384,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.0005901294760406017,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 2.363455057144165,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": -0.0003092627972364426,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.0766368880867958,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": -0.002663038857281208,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.1551397442817688,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0026730677112936974,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.7773767113685608,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": -0.0011709232814610005,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.42017874121665955,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": -0.0019059792160987855,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 1.0544211864471436,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": -0.004772515594959259,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.4706478416919708,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": -0.0002438830444589257,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.22088219225406647,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": -0.004391691088676453,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.5972660183906555,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": -0.007541544735431671,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.19400978088378906,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": -0.004146666824817657,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 1.2469046115875244,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": -0.0019529350101947784,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.14316241443157196,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": -0.008291246742010117,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.4657886326313019,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": -0.006107022240757942,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 94.6846,
+      "eval_generator_rouge1": 99.7652,
+      "eval_generator_rouge2": 99.667,
+      "eval_generator_rougeL": 99.5471,
+      "eval_loss": 0.006834100466221571,
+      "eval_runtime": 601.0294,
+      "eval_samples_per_second": 5.68,
+      "eval_selector_accuracy": 97.8388,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 628
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.346351770520781e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/csd/checkpoint-628/training_args.bin b/qwen3/csd/checkpoint-628/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4b4d426a661bf64c99e9ff7730557b24ebf0e60c
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e0687327fbf0a1cbac0cde783b20e396ed970bb1eed13dcf33e672089c5f44
+size 7185
diff --git a/qwen3/csd/checkpoint-628/zero_to_fp32.py b/qwen3/csd/checkpoint-628/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/csd/checkpoint-628/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/csd/checkpoint-942/README.md b/qwen3/csd/checkpoint-942/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-942/adapter_config.json b/qwen3/csd/checkpoint-942/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9959003652c9aaae4cece005b46aaf87e6d0f224
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "up_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "o_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-942/adapter_model.safetensors b/qwen3/csd/checkpoint-942/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..274d1cb91f7532fc63d04c058e05383bfdcb87bc
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41307e764f199471398c357bef49275eed521835fb93c390c9aa923b54b8fbac
+size 40422560
diff --git a/qwen3/csd/checkpoint-942/chat_template.jinja b/qwen3/csd/checkpoint-942/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-942/global_step942/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-942/global_step942/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..eb9a594bd30d3d3b8f8153a30d966cc4f1279bd8
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/global_step942/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3759ef693f54f35b13069160c5d8a07559e35684ed5bf38a0d9ecb784ece88d4
+size 121135557
diff --git a/qwen3/csd/checkpoint-942/global_step942/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/csd/checkpoint-942/global_step942/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..e5590bfbee8e31c7d9f0366003b00786718d0a93
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/global_step942/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b0d384d4366696a0b6e1d09d5af28e0405cc8fdcc9391aea83352acf2c9df51
+size 121135749
diff --git a/qwen3/csd/checkpoint-942/global_step942/mp_rank_00_model_states.pt b/qwen3/csd/checkpoint-942/global_step942/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..9682e6908906693685dcc06a8c9fe01363a46000
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/global_step942/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e8a2874917593811eb141fe4e6b049753292d104edb3763755e7ee0ea61d64a
+size 351841765
diff --git a/qwen3/csd/checkpoint-942/latest b/qwen3/csd/checkpoint-942/latest
new file mode 100644
index 0000000000000000000000000000000000000000..4e9a68e4460aca593c480611a486d2a4cd73ef42
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/latest
@@ -0,0 +1 @@
+global_step942
\ No newline at end of file
diff --git a/qwen3/csd/checkpoint-942/rng_state_0.pth b/qwen3/csd/checkpoint-942/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..8f8b77d4cdc54516505d36ac7fb5715b58436c78
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18d771c9a513dbae7970979494c25db1b118adc2e913f25175b0c80845d96441
+size 14917
diff --git a/qwen3/csd/checkpoint-942/rng_state_1.pth b/qwen3/csd/checkpoint-942/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..7f8a81fb86858f1045478c8dcec2eb804ff91120
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b4d52eca5a91b412c929ff7e86bc2b8e4d0c0eacf6a27c8afc10c8a1d585a8b
+size 14917
diff --git a/qwen3/csd/checkpoint-942/scheduler.pt b/qwen3/csd/checkpoint-942/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..6b1f8af0480bf83ab728abfeff11bc3288214349
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8076da6f897951c5bdf8649e4fba596e8c49819927746017442ffd85f40b18a3
+size 1465
diff --git a/qwen3/csd/checkpoint-942/tokenizer.json b/qwen3/csd/checkpoint-942/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/csd/checkpoint-942/tokenizer_config.json b/qwen3/csd/checkpoint-942/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/csd/checkpoint-942/trainer_state.json b/qwen3/csd/checkpoint-942/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..b0bd542aa0d9b80226abb6e1c26820d3bdb6b96e
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/trainer_state.json
@@ -0,0 +1,408 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 3.0,
+  "eval_steps": 500,
+  "global_step": 942,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 6.0161848068237305,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.11036398410797119,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 4.099686622619629,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.01695394217967987,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 3.3755345344543457,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.03482693731784821,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 1.194334864616394,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.027955988049507143,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 6.6159281730651855,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.005286623537540436,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 2.851339817047119,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.01179061084985733,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.989425539970398,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.006131124496459961,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 1.3493149280548096,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.007052437216043472,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.9337753057479858,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": -0.002346178889274597,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 1.5435981750488281,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.013309524953365326,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.7995954155921936,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.008521870523691178,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 1.6571325063705444,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.008653326332569123,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.519936203956604,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.004485934227705002,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.31598174571990967,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.003059915825724602,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.48084524273872375,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.005025627836585045,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 86.2734,
+      "eval_generator_rouge1": 99.1343,
+      "eval_generator_rouge2": 98.8297,
+      "eval_generator_rougeL": 98.6074,
+      "eval_loss": 0.019235465675592422,
+      "eval_runtime": 603.0951,
+      "eval_samples_per_second": 5.661,
+      "eval_selector_accuracy": 93.3995,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.177,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.34810933470726013,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.008805014193058014,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.19888198375701904,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.0013909661211073399,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 1.0506941080093384,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.0005901294760406017,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 2.363455057144165,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": -0.0003092627972364426,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.0766368880867958,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": -0.002663038857281208,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.1551397442817688,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0026730677112936974,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.7773767113685608,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": -0.0011709232814610005,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.42017874121665955,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": -0.0019059792160987855,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 1.0544211864471436,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": -0.004772515594959259,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.4706478416919708,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": -0.0002438830444589257,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.22088219225406647,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": -0.004391691088676453,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.5972660183906555,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": -0.007541544735431671,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.19400978088378906,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": -0.004146666824817657,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 1.2469046115875244,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": -0.0019529350101947784,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.14316241443157196,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": -0.008291246742010117,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.4657886326313019,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": -0.006107022240757942,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 94.6846,
+      "eval_generator_rouge1": 99.7652,
+      "eval_generator_rouge2": 99.667,
+      "eval_generator_rougeL": 99.5471,
+      "eval_loss": 0.006834100466221571,
+      "eval_runtime": 601.0294,
+      "eval_samples_per_second": 5.68,
+      "eval_selector_accuracy": 97.8388,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.22094473242759705,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": -0.005953450500965118,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.5406356453895569,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": -0.0025423141196370125,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.09373176842927933,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": -0.0044939789921045305,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.1080167219042778,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": -0.00820673182606697,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.05399267002940178,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": -0.009635555744171142,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.2048628032207489,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": -0.010086850821971893,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.08765007555484772,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": -0.008542184531688691,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.2229253351688385,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": -0.011658547818660736,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.18513049185276031,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": -0.00930779054760933,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.3498702347278595,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": -0.014870831370353698,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.16160713136196136,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": -0.020250825583934783,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.16046860814094543,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": -0.018617722392082214,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.17135752737522125,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": -0.020963044464588167,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.24867695569992065,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": -0.020418259501457214,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.12315966188907623,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": -0.018548570573329926,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.14333273470401764,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": -0.018815448880195616,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.736,
+      "eval_generator_rouge1": 99.9465,
+      "eval_generator_rouge2": 99.8866,
+      "eval_generator_rougeL": 99.7132,
+      "eval_loss": 0.005336109548807144,
+      "eval_runtime": 608.1178,
+      "eval_samples_per_second": 5.614,
+      "eval_selector_accuracy": 97.722,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.176,
+      "step": 942
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 3.524157743353037e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/csd/checkpoint-942/training_args.bin b/qwen3/csd/checkpoint-942/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4b4d426a661bf64c99e9ff7730557b24ebf0e60c
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e0687327fbf0a1cbac0cde783b20e396ed970bb1eed13dcf33e672089c5f44
+size 7185
diff --git a/qwen3/csd/checkpoint-942/zero_to_fp32.py b/qwen3/csd/checkpoint-942/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/csd/checkpoint-942/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/csd/eval_results.json b/qwen3/csd/eval_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..112d25829a87ac1815d231290e097df398cdcd0d
--- /dev/null
+++ b/qwen3/csd/eval_results.json
@@ -0,0 +1,14 @@
+{
+    "epoch": 5.0,
+    "eval_generator_count": 1712.0,
+    "eval_generator_exact_match": 96.6121,
+    "eval_generator_rouge1": 99.9371,
+    "eval_generator_rouge2": 99.8891,
+    "eval_generator_rougeL": 99.7571,
+    "eval_loss": 0.00408697547391057,
+    "eval_runtime": 585.4618,
+    "eval_samples_per_second": 5.831,
+    "eval_selector_accuracy": 98.5981,
+    "eval_selector_count": 1712.0,
+    "eval_steps_per_second": 0.183
+}
\ No newline at end of file
diff --git a/qwen3/csd/tokenizer.json b/qwen3/csd/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/csd/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/csd/tokenizer_config.json b/qwen3/csd/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/csd/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/csd/train_results.json b/qwen3/csd/train_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..c3a68d3096d434cf6cb3f07aa92a7d532179dc95
--- /dev/null
+++ b/qwen3/csd/train_results.json
@@ -0,0 +1,8 @@
+{
+    "epoch": 5.0,
+    "total_flos": 5.871444840310374e+16,
+    "train_loss": -0.013058029026516779,
+    "train_runtime": 7965.6329,
+    "train_samples_per_second": 6.294,
+    "train_steps_per_second": 0.197
+}
\ No newline at end of file
diff --git a/qwen3/csd/trainer_state.json b/qwen3/csd/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..955f43064efc44d4f8660d3bba16953130447406
--- /dev/null
+++ b/qwen3/csd/trainer_state.json
@@ -0,0 +1,664 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 5.0,
+  "eval_steps": 500,
+  "global_step": 1570,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 6.0161848068237305,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.11036398410797119,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 4.099686622619629,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.01695394217967987,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 3.3755345344543457,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.03482693731784821,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 1.194334864616394,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.027955988049507143,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 6.6159281730651855,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.005286623537540436,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 2.851339817047119,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.01179061084985733,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.989425539970398,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.006131124496459961,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 1.3493149280548096,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.007052437216043472,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 0.9337753057479858,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": -0.002346178889274597,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 1.5435981750488281,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.013309524953365326,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.7995954155921936,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.008521870523691178,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 1.6571325063705444,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.008653326332569123,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.519936203956604,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.004485934227705002,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.31598174571990967,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.003059915825724602,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.48084524273872375,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.005025627836585045,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 86.2734,
+      "eval_generator_rouge1": 99.1343,
+      "eval_generator_rouge2": 98.8297,
+      "eval_generator_rougeL": 98.6074,
+      "eval_loss": 0.019235465675592422,
+      "eval_runtime": 603.0951,
+      "eval_samples_per_second": 5.661,
+      "eval_selector_accuracy": 93.3995,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.177,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 0.34810933470726013,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.008805014193058014,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.19888198375701904,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.0013909661211073399,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 1.0506941080093384,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.0005901294760406017,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 2.363455057144165,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": -0.0003092627972364426,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.0766368880867958,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": -0.002663038857281208,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.1551397442817688,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.0026730677112936974,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.7773767113685608,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": -0.0011709232814610005,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.42017874121665955,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": -0.0019059792160987855,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 1.0544211864471436,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": -0.004772515594959259,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.4706478416919708,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": -0.0002438830444589257,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.22088219225406647,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": -0.004391691088676453,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.5972660183906555,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": -0.007541544735431671,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.19400978088378906,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": -0.004146666824817657,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 1.2469046115875244,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": -0.0019529350101947784,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 0.14316241443157196,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": -0.008291246742010117,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.4657886326313019,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": -0.006107022240757942,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 94.6846,
+      "eval_generator_rouge1": 99.7652,
+      "eval_generator_rouge2": 99.667,
+      "eval_generator_rougeL": 99.5471,
+      "eval_loss": 0.006834100466221571,
+      "eval_runtime": 601.0294,
+      "eval_samples_per_second": 5.68,
+      "eval_selector_accuracy": 97.8388,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.22094473242759705,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": -0.005953450500965118,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.5406356453895569,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": -0.0025423141196370125,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.09373176842927933,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": -0.0044939789921045305,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.1080167219042778,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": -0.00820673182606697,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.05399267002940178,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": -0.009635555744171142,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.2048628032207489,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": -0.010086850821971893,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.08765007555484772,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": -0.008542184531688691,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.2229253351688385,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": -0.011658547818660736,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.18513049185276031,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": -0.00930779054760933,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.3498702347278595,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": -0.014870831370353698,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.16160713136196136,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": -0.020250825583934783,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.16046860814094543,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": -0.018617722392082214,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.17135752737522125,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": -0.020963044464588167,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.24867695569992065,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": -0.020418259501457214,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.12315966188907623,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": -0.018548570573329926,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.14333273470401764,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": -0.018815448880195616,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.736,
+      "eval_generator_rouge1": 99.9465,
+      "eval_generator_rouge2": 99.8866,
+      "eval_generator_rougeL": 99.7132,
+      "eval_loss": 0.005336109548807144,
+      "eval_runtime": 608.1178,
+      "eval_samples_per_second": 5.614,
+      "eval_selector_accuracy": 97.722,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.176,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.1492357701063156,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": -0.02238200604915619,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.0778684988617897,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": -0.02527567744255066,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.07070758193731308,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": -0.02356054335832596,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.10038785636425018,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": -0.019069214165210725,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.33050113916397095,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": -0.022282478213310242,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.12148360908031464,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": -0.02368807643651962,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.16367903351783752,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": -0.02528309226036072,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.0877315029501915,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": -0.026697894930839537,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.06676916033029556,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": -0.028009063005447386,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.09503277391195297,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": -0.0323164701461792,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.08010950684547424,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": -0.03129267990589142,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.15882308781147003,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": -0.03188144564628601,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.07064881920814514,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": -0.03279578983783722,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.24478258192539215,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": -0.03418469429016113,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.09533516317605972,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": -0.036141425371170044,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.2617,
+      "eval_generator_rouge1": 99.8812,
+      "eval_generator_rouge2": 99.8259,
+      "eval_generator_rougeL": 99.6849,
+      "eval_loss": 0.0041860658675432205,
+      "eval_runtime": 606.5994,
+      "eval_samples_per_second": 5.628,
+      "eval_selector_accuracy": 98.3645,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.176,
+      "step": 1256
+    },
+    {
+      "epoch": 4.0127642600717985,
+      "grad_norm": 0.527934730052948,
+      "learning_rate": 9.464081721960592e-06,
+      "loss": -0.034009939432144164,
+      "step": 1260
+    },
+    {
+      "epoch": 4.076585560430794,
+      "grad_norm": 0.09058220684528351,
+      "learning_rate": 8.331628936475818e-06,
+      "loss": -0.037159329652786253,
+      "step": 1280
+    },
+    {
+      "epoch": 4.140406860789788,
+      "grad_norm": 0.05895598232746124,
+      "learning_rate": 7.2659843568748385e-06,
+      "loss": -0.0379990428686142,
+      "step": 1300
+    },
+    {
+      "epoch": 4.204228161148784,
+      "grad_norm": 0.0801444798707962,
+      "learning_rate": 6.268854516770008e-06,
+      "loss": -0.03669540882110596,
+      "step": 1320
+    },
+    {
+      "epoch": 4.268049461507778,
+      "grad_norm": 0.08857632428407669,
+      "learning_rate": 5.341836229602331e-06,
+      "loss": -0.03779337704181671,
+      "step": 1340
+    },
+    {
+      "epoch": 4.331870761866773,
+      "grad_norm": 0.07376661151647568,
+      "learning_rate": 4.4864140314888216e-06,
+      "loss": -0.03757229745388031,
+      "step": 1360
+    },
+    {
+      "epoch": 4.395692062225768,
+      "grad_norm": 0.08888167142868042,
+      "learning_rate": 3.703957803871932e-06,
+      "loss": -0.03948601186275482,
+      "step": 1380
+    },
+    {
+      "epoch": 4.4595133625847625,
+      "grad_norm": 0.12080858647823334,
+      "learning_rate": 2.9957205797780495e-06,
+      "loss": -0.039844200015068054,
+      "step": 1400
+    },
+    {
+      "epoch": 4.523334662943758,
+      "grad_norm": 0.16051147878170013,
+      "learning_rate": 2.362836537198199e-06,
+      "loss": -0.0381706565618515,
+      "step": 1420
+    },
+    {
+      "epoch": 4.587155963302752,
+      "grad_norm": 0.06491488963365555,
+      "learning_rate": 1.8063191828043703e-06,
+      "loss": -0.04070441722869873,
+      "step": 1440
+    },
+    {
+      "epoch": 4.650977263661747,
+      "grad_norm": 0.10599425435066223,
+      "learning_rate": 1.3270597289100437e-06,
+      "loss": -0.03966156244277954,
+      "step": 1460
+    },
+    {
+      "epoch": 4.714798564020742,
+      "grad_norm": 0.07826361805200577,
+      "learning_rate": 9.258256662741205e-07,
+      "loss": -0.040459868311882016,
+      "step": 1480
+    },
+    {
+      "epoch": 4.778619864379737,
+      "grad_norm": 0.1433020979166031,
+      "learning_rate": 6.032595350337341e-07,
+      "loss": -0.04145862460136414,
+      "step": 1500
+    },
+    {
+      "epoch": 4.842441164738732,
+      "grad_norm": 0.1149950921535492,
+      "learning_rate": 3.5987789573421356e-07,
+      "loss": -0.03695116639137268,
+      "step": 1520
+    },
+    {
+      "epoch": 4.9062624650977265,
+      "grad_norm": 0.15655829012393951,
+      "learning_rate": 1.9607050210401192e-07,
+      "loss": -0.03868899643421173,
+      "step": 1540
+    },
+    {
+      "epoch": 4.970083765456721,
+      "grad_norm": 0.1153496727347374,
+      "learning_rate": 1.1209967689925832e-07,
+      "loss": -0.04158148169517517,
+      "step": 1560
+    },
+    {
+      "epoch": 5.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.6706,
+      "eval_generator_rouge1": 99.9491,
+      "eval_generator_rouge2": 99.8942,
+      "eval_generator_rougeL": 99.7857,
+      "eval_loss": 0.00408697547391057,
+      "eval_runtime": 602.4809,
+      "eval_samples_per_second": 5.667,
+      "eval_selector_accuracy": 98.6565,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 1570
+    },
+    {
+      "epoch": 5.0,
+      "step": 1570,
+      "total_flos": 5.871444840310374e+16,
+      "train_loss": -0.013058029026516779,
+      "train_runtime": 7965.6329,
+      "train_samples_per_second": 6.294,
+      "train_steps_per_second": 0.197
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.871444840310374e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/csd/training_args.bin b/qwen3/csd/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4b4d426a661bf64c99e9ff7730557b24ebf0e60c
--- /dev/null
+++ b/qwen3/csd/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e0687327fbf0a1cbac0cde783b20e396ed970bb1eed13dcf33e672089c5f44
+size 7185
diff --git a/qwen3/distillm_adaptive_sfkl/README.md b/qwen3/distillm_adaptive_sfkl/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..112a32962ddbbc571ea4b0e02cc7e2527c21da68
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/README.md
@@ -0,0 +1,79 @@
+---
+library_name: peft
+license: other
+base_model: Qwen/Qwen3-0.6B
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+pipeline_tag: text-generation
+model-index:
+- name: distillm_adaptive_sfkl
+  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. -->
+
+# distillm_adaptive_sfkl
+
+This model is a fine-tuned version of [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) on the cypher_prepared_train dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.0032
+- Selector Count: 1712.0
+- Generator Count: 1712.0
+- Selector Accuracy: 98.715
+- Generator Exact Match: 96.3785
+- Generator Rouge1: 99.9305
+- Generator Rouge2: 99.8682
+- Generator Rougel: 99.7529
+
+## 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: 0.0001
+- train_batch_size: 2
+- eval_batch_size: 16
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 2
+- gradient_accumulation_steps: 8
+- total_train_batch_size: 32
+- total_eval_batch_size: 32
+- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine_with_min_lr
+- num_epochs: 5
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss | Selector Count | Generator Count | Selector Accuracy | Generator Exact Match | Generator Rouge1 | Generator Rouge2 | Generator Rougel |
+|:-------------:|:-----:|:----:|:---------------:|:--------------:|:---------------:|:-----------------:|:---------------------:|:----------------:|:----------------:|:----------------:|
+| 0.0097        | 1.0   | 314  | 0.0078          | 1712.0         | 1712.0          | 95.5607           | 90.0701               | 99.4297          | 99.1758          | 99.0503          |
+| 0.0035        | 2.0   | 628  | 0.0057          | 1712.0         | 1712.0          | 97.722            | 92.9907               | 99.8109          | 99.6137          | 99.5606          |
+| 0.0014        | 3.0   | 942  | 0.0040          | 1712.0         | 1712.0          | 98.0724           | 95.3271               | 99.9061          | 99.7991          | 99.6925          |
+| 0.0009        | 4.0   | 1256 | 0.0032          | 1712.0         | 1712.0          | 98.5981           | 96.6706               | 99.9312          | 99.8709          | 99.7682          |
+| 0.0011        | 5.0   | 1570 | 0.0032          | 1712.0         | 1712.0          | 98.7734           | 96.4953               | 99.9278          | 99.8599          | 99.777           |
+
+
+### Framework versions
+
+- PEFT 0.18.1
+- Transformers 5.8.0
+- Pytorch 2.9.1+cu128
+- Datasets 4.0.0
+- Tokenizers 0.22.2
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/adapter_config.json b/qwen3/distillm_adaptive_sfkl/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d88d3ddd2a1fe81f4cf32eb7f116cfcaecff483e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "down_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/adapter_model.safetensors b/qwen3/distillm_adaptive_sfkl/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..906be98e4f889370ba814eecd51b8209eb2a3815
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a99e7a0decfbd11795996653da7c5bc4e9262944d052503644f577fff3ab8626
+size 40422560
diff --git a/qwen3/distillm_adaptive_sfkl/all_results.json b/qwen3/distillm_adaptive_sfkl/all_results.json
new file mode 100644
index 0000000000000000000000000000000000000000..b6c1608e9d66e99a3172e8a9b93ecd591af7801b
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/all_results.json
@@ -0,0 +1,24 @@
+{
+    "epoch": 5.0,
+    "eval_distillm_threshold": 0.0,
+    "eval_distillm_threshold_changed": 0.0,
+    "eval_generator_count": 1712.0,
+    "eval_generator_exact_match": 96.3785,
+    "eval_generator_rouge1": 99.9305,
+    "eval_generator_rouge2": 99.8682,
+    "eval_generator_rougeL": 99.7529,
+    "eval_loss": 0.003217515302821994,
+    "eval_rollout_dataset_steps": 12535,
+    "eval_rollout_fresh_steps": 0,
+    "eval_rollout_replay_steps": 0,
+    "eval_runtime": 579.9368,
+    "eval_samples_per_second": 5.887,
+    "eval_selector_accuracy": 98.715,
+    "eval_selector_count": 1712.0,
+    "eval_steps_per_second": 0.185,
+    "total_flos": 5.871444840310374e+16,
+    "train_loss": 0.011652989937051847,
+    "train_runtime": 11256.8941,
+    "train_samples_per_second": 4.454,
+    "train_steps_per_second": 0.139
+}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/chat_template.jinja b/qwen3/distillm_adaptive_sfkl/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/README.md b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/adapter_config.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d88d3ddd2a1fe81f4cf32eb7f116cfcaecff483e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "down_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/adapter_model.safetensors b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..372afd30ebfa129646c46970f09e3bf30c8a13a1
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f58e3b2fa10c1d01313865a7357234cc6333c4bd4046e8d6688a71a03aa4d13
+size 40422560
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/chat_template.jinja b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/distillm_state_rank0.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/distillm_state_rank0.pt
new file mode 100644
index 0000000000000000000000000000000000000000..ba55d0738648d06c6b1157efaafadb891c3ede66
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/distillm_state_rank0.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a9b5a84ccbbe30ec49486231d2b6d24ec7caa2a645eaeb1ef1ac1046bacb30f
+size 9403
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/distillm_state_rank1.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/distillm_state_rank1.pt
new file mode 100644
index 0000000000000000000000000000000000000000..7d6f6e0743fd8e8410d9dfaa520a36a19db24f2f
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/distillm_state_rank1.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55b029e6dfb2341a0bd92c4bf670fd9d1504f4e3c5bfe7a81b0984f5e033f257
+size 9339
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..87579889f5e5584141ab4a19dd302d972ada8d32
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58b78c70f5888623805998daab9f14c7e2814bd923742b7f2f0b621ee3f04007
+size 121135557
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..9483b621110fc9dda6dc979a47f28d2e46c22851
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59424cfcbfd035bb92538ce0e6f10147c53b280b7714a52322c8a59e77787f3a
+size 121135749
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..d91391fa7c229c8eec9b1b6d05fd03b3b03d559b
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/global_step1256/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26c5dd1b685e3e6a02aa27262d7215ba3e7570008ea738399c9cf0fceb699c6d
+size 351841765
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/latest b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/latest
new file mode 100644
index 0000000000000000000000000000000000000000..74abb3d23ea83ebdac59918752db5d1782834ca7
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/latest
@@ -0,0 +1 @@
+global_step1256
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/rng_state_0.pth b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..fff1c25f9b0daed6d86b07a8a74fb5a26e71a4cc
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3938975c7fa1c658632f9261d8c2105e3cb55a634ad1d7f0de51ab9cba34601b
+size 14917
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/rng_state_1.pth b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..4ffa55d98ace7ca67ac9fe09cc26debcaf9358b9
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08b65105f0441eb434f927c31aeb8917c9a0b1ce46d94c0a321be03146ac2bdd
+size 14917
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/scheduler.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..40e863d07cfed22cd912014eb3ba3fd2b16b173f
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1775297f14e75f7463ae51983a66296f81f4b2e2e5d7234f1a6d295abfa4cf23
+size 1465
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/tokenizer.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/tokenizer_config.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/trainer_state.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..7026bc2435f14e47e197ec7304b6752cea12e1bb
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/trainer_state.json
@@ -0,0 +1,543 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 4.0,
+  "eval_steps": 500,
+  "global_step": 1256,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "distillm_initial_generator_count": 1712.0,
+      "distillm_initial_generator_exact_match": 0.0,
+      "distillm_initial_generator_rouge1": 16.2149,
+      "distillm_initial_generator_rouge2": 3.2276,
+      "distillm_initial_generator_rougeL": 11.6725,
+      "distillm_initial_loss": 2.494271755218506,
+      "distillm_initial_runtime": 2988.1286,
+      "distillm_initial_samples_per_second": 1.143,
+      "distillm_initial_selector_accuracy": 52.278,
+      "distillm_initial_selector_count": 1712.0,
+      "distillm_initial_steps_per_second": 0.036,
+      "epoch": 0,
+      "step": 0
+    },
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 1.4179933071136475,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.360275936126709,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.6145927309989929,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.0921450138092041,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.949298083782196,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.06155582666397095,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.32781288027763367,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.03148526251316071,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.42831170558929443,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.0304107666015625,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.37467315793037415,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.018133948743343353,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.26231899857521057,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.023359841108322142,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.23304614424705505,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.015377765893936158,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 4.825753688812256,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.0134853795170784,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.29226061701774597,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.022103385627269746,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.1759956330060959,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.011787685751914977,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.18942032754421234,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.009058185666799546,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.2657412886619568,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.009212753176689148,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.28189197182655334,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.009374459087848664,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.22191843390464783,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.009671492874622345,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 90.0701,
+      "eval_generator_rouge1": 99.4297,
+      "eval_generator_rouge2": 99.1758,
+      "eval_generator_rougeL": 99.0503,
+      "eval_loss": 0.007805492263287306,
+      "eval_runtime": 601.5417,
+      "eval_samples_per_second": 5.675,
+      "eval_selector_accuracy": 95.5607,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 1.0133247375488281,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.013477623462677002,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.07633373141288757,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.0067503504455089566,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 0.10817505419254303,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.005428457632660866,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 0.13295228779315948,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": 0.004656650871038437,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.09486064314842224,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": 0.008230840414762497,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.3019433617591858,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.008267045021057129,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.06807491928339005,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": 0.006827282160520554,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.31115707755088806,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": 0.005447100847959518,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 0.09037721902132034,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": 0.0037935201078653337,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.050306856632232666,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": 0.004351192712783813,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.06550924479961395,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": 0.004179144278168678,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.037223804742097855,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": 0.006012631207704544,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.07440082728862762,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": 0.0067665338516235355,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 0.4098421037197113,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": 0.007050857692956924,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 1.077409029006958,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": 0.0025460906326770782,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.14294511079788208,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": 0.0034835845232009887,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 92.9907,
+      "eval_generator_rouge1": 99.8109,
+      "eval_generator_rouge2": 99.6137,
+      "eval_generator_rougeL": 99.5606,
+      "eval_loss": 0.005652441177517176,
+      "eval_runtime": 602.3066,
+      "eval_samples_per_second": 5.668,
+      "eval_selector_accuracy": 97.722,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.15983566641807556,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": 0.0025436000898480414,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.24636076390743256,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": 0.005166642367839813,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.05653794854879379,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": 0.0020019154995679854,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.020209209993481636,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": 0.003964858502149582,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.015605684369802475,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": 0.005196084827184677,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.05083553120493889,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": 0.002757585048675537,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.023810580372810364,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": 0.0015414590016007423,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.022960327565670013,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": 0.0012374447658658027,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.08331727981567383,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": 0.0037563137710094454,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.3172169327735901,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": 0.0025980696082115173,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.0642293244600296,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": 0.0016859130933880805,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.07357119768857956,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": 0.002408899739384651,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.12400355935096741,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": 0.0020469389855861664,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.21444188058376312,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": 0.0023752737790346144,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.07703635096549988,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": 0.004145454987883568,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.11405396461486816,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": 0.0013555588200688362,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.3271,
+      "eval_generator_rouge1": 99.9061,
+      "eval_generator_rouge2": 99.7991,
+      "eval_generator_rougeL": 99.6925,
+      "eval_loss": 0.003950633108615875,
+      "eval_runtime": 595.0919,
+      "eval_samples_per_second": 5.737,
+      "eval_selector_accuracy": 98.0724,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.18,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.01742524839937687,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": 0.0017176205292344093,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.003188655711710453,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": 0.0026764504611492155,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.01925867609679699,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": 0.0012796200811862946,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.02856561355292797,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": 0.003274943307042122,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.28238120675086975,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": 0.0020828992128372192,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.0034029490780085325,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": 0.0026421770453453064,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.08193282783031464,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": 0.002248910814523697,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.006192841101437807,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": 0.0024001479148864748,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.023488029837608337,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": 0.001568210683763027,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.10149848461151123,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": 0.0015563317574560642,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.016474615782499313,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": 0.003182568401098251,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.015502875670790672,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": 0.002123878337442875,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.013245862908661366,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": 0.0021719962358474733,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.176155224442482,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": 0.003167513757944107,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.00724257156252861,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": 0.0008703418076038361,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.6706,
+      "eval_generator_rouge1": 99.9312,
+      "eval_generator_rouge2": 99.8709,
+      "eval_generator_rougeL": 99.7682,
+      "eval_loss": 0.0032213269732892513,
+      "eval_runtime": 594.8325,
+      "eval_samples_per_second": 5.739,
+      "eval_selector_accuracy": 98.5981,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.18,
+      "step": 1256
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 4.697643190059008e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/training_args.bin b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..9965b3704c5694d942c823f03f32ef4aa365d5a2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4080ce6c122d894dce2a6242fb58ae879822809092a9484cf3c306a91aacf376
+size 7185
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1256/zero_to_fp32.py b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1256/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/README.md b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/adapter_config.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d88d3ddd2a1fe81f4cf32eb7f116cfcaecff483e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "down_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/adapter_model.safetensors b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..906be98e4f889370ba814eecd51b8209eb2a3815
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a99e7a0decfbd11795996653da7c5bc4e9262944d052503644f577fff3ab8626
+size 40422560
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/chat_template.jinja b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/distillm_state_rank0.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/distillm_state_rank0.pt
new file mode 100644
index 0000000000000000000000000000000000000000..becca175688219a9b4a422f423edefb47590217d
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/distillm_state_rank0.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c45fea5636ba58d552a49ba5a651afe08fe7e86b64bd29d82359e1ffb3df111f
+size 9403
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/distillm_state_rank1.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/distillm_state_rank1.pt
new file mode 100644
index 0000000000000000000000000000000000000000..2d7c319e37a117a78ed636f3109afc6ef3c8fd98
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/distillm_state_rank1.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34b6ffca878fd86b6133b9eb70d94fee805ff00ef1d33571b54dc0de3a307df3
+size 9339
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..3b24b94f1149c43a697a92512f3edd89444400ec
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dea9f965e183cf6e75cc938d2101c21aebc2cc249d288e312672c5b1c17a5de6
+size 121135557
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..9c18f90b893aa54624a6cf9a7820d62d24dc8a50
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:712782b89f3f2a08edd6b14367d0d1c1f316231e121f4e9f2ad44141927379c7
+size 121135749
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..39a22f7bb9c8a25d3c0bf9adaa55a6adbe3cfe2f
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/global_step1570/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4eb5d89461998455656d00f5f8b8b13c336625ba0180c96c30ae3fcddb30b535
+size 351841765
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/latest b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/latest
new file mode 100644
index 0000000000000000000000000000000000000000..18c2fcff58fa6daa0a2e7c3a015552a69085778b
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/latest
@@ -0,0 +1 @@
+global_step1570
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/rng_state_0.pth b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..d4b6079ff4d744e907933ebe2f9df086a16f163f
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4b7c956a417bcc6c356c97ca412ef5c5e2b5fc70fba8223db55cb48ac282753
+size 14917
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/rng_state_1.pth b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..d90069af091242b4bae104ce1480faa3b6b594c2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10569e1274020cceec29016fb35178a5dffe88d926ec542d32d89b0016d59ccf
+size 14917
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/scheduler.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..e22954463492d8a070ae427c95418df61893c54c
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6332fac17ae99a07597512c705658da2fd8429d1c39bb776b30e107c2cb7aae
+size 1465
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/tokenizer.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/tokenizer_config.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/trainer_state.json b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..aeeaa7242620077d49ec3151c11f99879000d0a3
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/trainer_state.json
@@ -0,0 +1,670 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 5.0,
+  "eval_steps": 500,
+  "global_step": 1570,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "distillm_initial_generator_count": 1712.0,
+      "distillm_initial_generator_exact_match": 0.0,
+      "distillm_initial_generator_rouge1": 16.2149,
+      "distillm_initial_generator_rouge2": 3.2276,
+      "distillm_initial_generator_rougeL": 11.6725,
+      "distillm_initial_loss": 2.494271755218506,
+      "distillm_initial_runtime": 2988.1286,
+      "distillm_initial_samples_per_second": 1.143,
+      "distillm_initial_selector_accuracy": 52.278,
+      "distillm_initial_selector_count": 1712.0,
+      "distillm_initial_steps_per_second": 0.036,
+      "epoch": 0,
+      "step": 0
+    },
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 1.4179933071136475,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.360275936126709,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.6145927309989929,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.0921450138092041,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.949298083782196,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.06155582666397095,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.32781288027763367,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.03148526251316071,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.42831170558929443,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.0304107666015625,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.37467315793037415,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.018133948743343353,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.26231899857521057,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.023359841108322142,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.23304614424705505,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.015377765893936158,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 4.825753688812256,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.0134853795170784,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.29226061701774597,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.022103385627269746,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.1759956330060959,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.011787685751914977,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.18942032754421234,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.009058185666799546,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.2657412886619568,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.009212753176689148,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.28189197182655334,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.009374459087848664,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.22191843390464783,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.009671492874622345,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 90.0701,
+      "eval_generator_rouge1": 99.4297,
+      "eval_generator_rouge2": 99.1758,
+      "eval_generator_rougeL": 99.0503,
+      "eval_loss": 0.007805492263287306,
+      "eval_runtime": 601.5417,
+      "eval_samples_per_second": 5.675,
+      "eval_selector_accuracy": 95.5607,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 314
+    },
+    {
+      "epoch": 1.0191463901076985,
+      "grad_norm": 1.0133247375488281,
+      "learning_rate": 9.01646335654399e-05,
+      "loss": 0.013477623462677002,
+      "step": 320
+    },
+    {
+      "epoch": 1.0829676904666932,
+      "grad_norm": 0.07633373141288757,
+      "learning_rate": 8.894171942290817e-05,
+      "loss": 0.0067503504455089566,
+      "step": 340
+    },
+    {
+      "epoch": 1.146788990825688,
+      "grad_norm": 0.10817505419254303,
+      "learning_rate": 8.765652370197003e-05,
+      "loss": 0.005428457632660866,
+      "step": 360
+    },
+    {
+      "epoch": 1.210610291184683,
+      "grad_norm": 0.13295228779315948,
+      "learning_rate": 8.631110452757467e-05,
+      "loss": 0.004656650871038437,
+      "step": 380
+    },
+    {
+      "epoch": 1.2744315915436777,
+      "grad_norm": 0.09486064314842224,
+      "learning_rate": 8.490761646709642e-05,
+      "loss": 0.008230840414762497,
+      "step": 400
+    },
+    {
+      "epoch": 1.3382528919026724,
+      "grad_norm": 0.3019433617591858,
+      "learning_rate": 8.344830707998953e-05,
+      "loss": 0.008267045021057129,
+      "step": 420
+    },
+    {
+      "epoch": 1.4020741922616673,
+      "grad_norm": 0.06807491928339005,
+      "learning_rate": 8.193551331852458e-05,
+      "loss": 0.006827282160520554,
+      "step": 440
+    },
+    {
+      "epoch": 1.4658954926206622,
+      "grad_norm": 0.31115707755088806,
+      "learning_rate": 8.037165778537033e-05,
+      "loss": 0.005447100847959518,
+      "step": 460
+    },
+    {
+      "epoch": 1.5297167929796571,
+      "grad_norm": 0.09037721902132034,
+      "learning_rate": 7.875924485401427e-05,
+      "loss": 0.0037935201078653337,
+      "step": 480
+    },
+    {
+      "epoch": 1.5935380933386518,
+      "grad_norm": 0.050306856632232666,
+      "learning_rate": 7.710085665823444e-05,
+      "loss": 0.004351192712783813,
+      "step": 500
+    },
+    {
+      "epoch": 1.6573593936976465,
+      "grad_norm": 0.06550924479961395,
+      "learning_rate": 7.539914895704521e-05,
+      "loss": 0.004179144278168678,
+      "step": 520
+    },
+    {
+      "epoch": 1.7211806940566414,
+      "grad_norm": 0.037223804742097855,
+      "learning_rate": 7.365684688173908e-05,
+      "loss": 0.006012631207704544,
+      "step": 540
+    },
+    {
+      "epoch": 1.7850019944156363,
+      "grad_norm": 0.07440082728862762,
+      "learning_rate": 7.18767405718347e-05,
+      "loss": 0.0067665338516235355,
+      "step": 560
+    },
+    {
+      "epoch": 1.848823294774631,
+      "grad_norm": 0.4098421037197113,
+      "learning_rate": 7.006168070692036e-05,
+      "loss": 0.007050857692956924,
+      "step": 580
+    },
+    {
+      "epoch": 1.9126445951336257,
+      "grad_norm": 1.077409029006958,
+      "learning_rate": 6.821457394154786e-05,
+      "loss": 0.0025460906326770782,
+      "step": 600
+    },
+    {
+      "epoch": 1.9764658954926206,
+      "grad_norm": 0.14294511079788208,
+      "learning_rate": 6.633837825048768e-05,
+      "loss": 0.0034835845232009887,
+      "step": 620
+    },
+    {
+      "epoch": 2.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 92.9907,
+      "eval_generator_rouge1": 99.8109,
+      "eval_generator_rouge2": 99.6137,
+      "eval_generator_rougeL": 99.5606,
+      "eval_loss": 0.005652441177517176,
+      "eval_runtime": 602.3066,
+      "eval_samples_per_second": 5.668,
+      "eval_selector_accuracy": 97.722,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 628
+    },
+    {
+      "epoch": 2.038292780215397,
+      "grad_norm": 0.15983566641807556,
+      "learning_rate": 6.443609819179926e-05,
+      "loss": 0.0025436000898480414,
+      "step": 640
+    },
+    {
+      "epoch": 2.102114080574392,
+      "grad_norm": 0.24636076390743256,
+      "learning_rate": 6.251078009530256e-05,
+      "loss": 0.005166642367839813,
+      "step": 660
+    },
+    {
+      "epoch": 2.1659353809333863,
+      "grad_norm": 0.05653794854879379,
+      "learning_rate": 6.0565507184155677e-05,
+      "loss": 0.0020019154995679854,
+      "step": 680
+    },
+    {
+      "epoch": 2.2297566812923812,
+      "grad_norm": 0.020209209993481636,
+      "learning_rate": 5.860339463735136e-05,
+      "loss": 0.003964858502149582,
+      "step": 700
+    },
+    {
+      "epoch": 2.293577981651376,
+      "grad_norm": 0.015605684369802475,
+      "learning_rate": 5.662758460103895e-05,
+      "loss": 0.005196084827184677,
+      "step": 720
+    },
+    {
+      "epoch": 2.357399282010371,
+      "grad_norm": 0.05083553120493889,
+      "learning_rate": 5.464124115666095e-05,
+      "loss": 0.002757585048675537,
+      "step": 740
+    },
+    {
+      "epoch": 2.421220582369366,
+      "grad_norm": 0.023810580372810364,
+      "learning_rate": 5.264754525396212e-05,
+      "loss": 0.0015414590016007423,
+      "step": 760
+    },
+    {
+      "epoch": 2.4850418827283605,
+      "grad_norm": 0.022960327565670013,
+      "learning_rate": 5.06496896169857e-05,
+      "loss": 0.0012374447658658027,
+      "step": 780
+    },
+    {
+      "epoch": 2.5488631830873554,
+      "grad_norm": 0.08331727981567383,
+      "learning_rate": 4.865087363121384e-05,
+      "loss": 0.0037563137710094454,
+      "step": 800
+    },
+    {
+      "epoch": 2.6126844834463503,
+      "grad_norm": 0.3172169327735901,
+      "learning_rate": 4.6654298220040673e-05,
+      "loss": 0.0025980696082115173,
+      "step": 820
+    },
+    {
+      "epoch": 2.6765057838053448,
+      "grad_norm": 0.0642293244600296,
+      "learning_rate": 4.466316071878229e-05,
+      "loss": 0.0016859130933880805,
+      "step": 840
+    },
+    {
+      "epoch": 2.74032708416434,
+      "grad_norm": 0.07357119768857956,
+      "learning_rate": 4.2680649754432913e-05,
+      "loss": 0.002408899739384651,
+      "step": 860
+    },
+    {
+      "epoch": 2.8041483845233346,
+      "grad_norm": 0.12400355935096741,
+      "learning_rate": 4.070994013936646e-05,
+      "loss": 0.0020469389855861664,
+      "step": 880
+    },
+    {
+      "epoch": 2.8679696848823295,
+      "grad_norm": 0.21444188058376312,
+      "learning_rate": 3.8754187787161155e-05,
+      "loss": 0.0023752737790346144,
+      "step": 900
+    },
+    {
+      "epoch": 2.9317909852413244,
+      "grad_norm": 0.07703635096549988,
+      "learning_rate": 3.681652465868875e-05,
+      "loss": 0.004145454987883568,
+      "step": 920
+    },
+    {
+      "epoch": 2.995612285600319,
+      "grad_norm": 0.11405396461486816,
+      "learning_rate": 3.490005374656184e-05,
+      "loss": 0.0013555588200688362,
+      "step": 940
+    },
+    {
+      "epoch": 3.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 95.3271,
+      "eval_generator_rouge1": 99.9061,
+      "eval_generator_rouge2": 99.7991,
+      "eval_generator_rougeL": 99.6925,
+      "eval_loss": 0.003950633108615875,
+      "eval_runtime": 595.0919,
+      "eval_samples_per_second": 5.737,
+      "eval_selector_accuracy": 98.0724,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.18,
+      "step": 942
+    },
+    {
+      "epoch": 3.057439170323095,
+      "grad_norm": 0.01742524839937687,
+      "learning_rate": 3.3007844105971336e-05,
+      "loss": 0.0017176205292344093,
+      "step": 960
+    },
+    {
+      "epoch": 3.12126047068209,
+      "grad_norm": 0.003188655711710453,
+      "learning_rate": 3.114292593987149e-05,
+      "loss": 0.0026764504611492155,
+      "step": 980
+    },
+    {
+      "epoch": 3.185081771041085,
+      "grad_norm": 0.01925867609679699,
+      "learning_rate": 2.9308285746383408e-05,
+      "loss": 0.0012796200811862946,
+      "step": 1000
+    },
+    {
+      "epoch": 3.24890307140008,
+      "grad_norm": 0.02856561355292797,
+      "learning_rate": 2.7506861536187867e-05,
+      "loss": 0.003274943307042122,
+      "step": 1020
+    },
+    {
+      "epoch": 3.3127243717590744,
+      "grad_norm": 0.28238120675086975,
+      "learning_rate": 2.5741538127566396e-05,
+      "loss": 0.0020828992128372192,
+      "step": 1040
+    },
+    {
+      "epoch": 3.3765456721180693,
+      "grad_norm": 0.0034029490780085325,
+      "learning_rate": 2.4015142526625115e-05,
+      "loss": 0.0026421770453453064,
+      "step": 1060
+    },
+    {
+      "epoch": 3.4403669724770642,
+      "grad_norm": 0.08193282783031464,
+      "learning_rate": 2.233043940009967e-05,
+      "loss": 0.002248910814523697,
+      "step": 1080
+    },
+    {
+      "epoch": 3.504188272836059,
+      "grad_norm": 0.006192841101437807,
+      "learning_rate": 2.0690126647990974e-05,
+      "loss": 0.0024001479148864748,
+      "step": 1100
+    },
+    {
+      "epoch": 3.568009573195054,
+      "grad_norm": 0.023488029837608337,
+      "learning_rate": 1.9096831083121828e-05,
+      "loss": 0.001568210683763027,
+      "step": 1120
+    },
+    {
+      "epoch": 3.6318308735540485,
+      "grad_norm": 0.10149848461151123,
+      "learning_rate": 1.7553104224533324e-05,
+      "loss": 0.0015563317574560642,
+      "step": 1140
+    },
+    {
+      "epoch": 3.6956521739130435,
+      "grad_norm": 0.016474615782499313,
+      "learning_rate": 1.606141821145735e-05,
+      "loss": 0.003182568401098251,
+      "step": 1160
+    },
+    {
+      "epoch": 3.7594734742720384,
+      "grad_norm": 0.015502875670790672,
+      "learning_rate": 1.4624161844408763e-05,
+      "loss": 0.002123878337442875,
+      "step": 1180
+    },
+    {
+      "epoch": 3.823294774631033,
+      "grad_norm": 0.013245862908661366,
+      "learning_rate": 1.324363675973698e-05,
+      "loss": 0.0021719962358474733,
+      "step": 1200
+    },
+    {
+      "epoch": 3.8871160749900278,
+      "grad_norm": 0.176155224442482,
+      "learning_rate": 1.1922053743763052e-05,
+      "loss": 0.003167513757944107,
+      "step": 1220
+    },
+    {
+      "epoch": 3.9509373753490227,
+      "grad_norm": 0.00724257156252861,
+      "learning_rate": 1.0661529192404973e-05,
+      "loss": 0.0008703418076038361,
+      "step": 1240
+    },
+    {
+      "epoch": 4.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.6706,
+      "eval_generator_rouge1": 99.9312,
+      "eval_generator_rouge2": 99.8709,
+      "eval_generator_rougeL": 99.7682,
+      "eval_loss": 0.0032213269732892513,
+      "eval_runtime": 594.8325,
+      "eval_samples_per_second": 5.739,
+      "eval_selector_accuracy": 98.5981,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.18,
+      "step": 1256
+    },
+    {
+      "epoch": 4.0127642600717985,
+      "grad_norm": 0.2515179216861725,
+      "learning_rate": 9.464081721960592e-06,
+      "loss": 0.0021197797730565073,
+      "step": 1260
+    },
+    {
+      "epoch": 4.076585560430794,
+      "grad_norm": 0.052773285657167435,
+      "learning_rate": 8.331628936475818e-06,
+      "loss": 0.0007191885262727738,
+      "step": 1280
+    },
+    {
+      "epoch": 4.140406860789788,
+      "grad_norm": 0.020960243418812752,
+      "learning_rate": 7.2659843568748385e-06,
+      "loss": 0.0012289942242205143,
+      "step": 1300
+    },
+    {
+      "epoch": 4.204228161148784,
+      "grad_norm": 0.009606960229575634,
+      "learning_rate": 6.268854516770008e-06,
+      "loss": 0.0012783183716237544,
+      "step": 1320
+    },
+    {
+      "epoch": 4.268049461507778,
+      "grad_norm": 0.03356768190860748,
+      "learning_rate": 5.341836229602331e-06,
+      "loss": 0.0015904821455478668,
+      "step": 1340
+    },
+    {
+      "epoch": 4.331870761866773,
+      "grad_norm": 0.43399015069007874,
+      "learning_rate": 4.4864140314888216e-06,
+      "loss": 0.0019017960876226425,
+      "step": 1360
+    },
+    {
+      "epoch": 4.395692062225768,
+      "grad_norm": 0.09213124215602875,
+      "learning_rate": 3.703957803871932e-06,
+      "loss": 0.000992914568632841,
+      "step": 1380
+    },
+    {
+      "epoch": 4.4595133625847625,
+      "grad_norm": 0.030944187194108963,
+      "learning_rate": 2.9957205797780495e-06,
+      "loss": 0.0019235227257013321,
+      "step": 1400
+    },
+    {
+      "epoch": 4.523334662943758,
+      "grad_norm": 0.04038539156317711,
+      "learning_rate": 2.362836537198199e-06,
+      "loss": 0.001333768479526043,
+      "step": 1420
+    },
+    {
+      "epoch": 4.587155963302752,
+      "grad_norm": 0.009491882286965847,
+      "learning_rate": 1.8063191828043703e-06,
+      "loss": 0.0010302967391908168,
+      "step": 1440
+    },
+    {
+      "epoch": 4.650977263661747,
+      "grad_norm": 0.02417566254734993,
+      "learning_rate": 1.3270597289100437e-06,
+      "loss": 0.0008206027559936047,
+      "step": 1460
+    },
+    {
+      "epoch": 4.714798564020742,
+      "grad_norm": 0.0027866039890795946,
+      "learning_rate": 9.258256662741205e-07,
+      "loss": 0.0008394632488489151,
+      "step": 1480
+    },
+    {
+      "epoch": 4.778619864379737,
+      "grad_norm": 0.01856832019984722,
+      "learning_rate": 6.032595350337341e-07,
+      "loss": 0.0011719820089638234,
+      "step": 1500
+    },
+    {
+      "epoch": 4.842441164738732,
+      "grad_norm": 0.013062495738267899,
+      "learning_rate": 3.5987789573421356e-07,
+      "loss": 0.0029238536953926086,
+      "step": 1520
+    },
+    {
+      "epoch": 4.9062624650977265,
+      "grad_norm": 0.094195157289505,
+      "learning_rate": 1.9607050210401192e-07,
+      "loss": 0.0008629086427390575,
+      "step": 1540
+    },
+    {
+      "epoch": 4.970083765456721,
+      "grad_norm": 0.02496415562927723,
+      "learning_rate": 1.1209967689925832e-07,
+      "loss": 0.0011463512666523457,
+      "step": 1560
+    },
+    {
+      "epoch": 5.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 96.4953,
+      "eval_generator_rouge1": 99.9278,
+      "eval_generator_rouge2": 99.8599,
+      "eval_generator_rougeL": 99.777,
+      "eval_loss": 0.003217515302821994,
+      "eval_runtime": 592.4524,
+      "eval_samples_per_second": 5.762,
+      "eval_selector_accuracy": 98.7734,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.181,
+      "step": 1570
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.871444840310374e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/training_args.bin b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..9965b3704c5694d942c823f03f32ef4aa365d5a2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4080ce6c122d894dce2a6242fb58ae879822809092a9484cf3c306a91aacf376
+size 7185
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-1570/zero_to_fp32.py b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-1570/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/README.md b/qwen3/distillm_adaptive_sfkl/checkpoint-314/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/adapter_config.json b/qwen3/distillm_adaptive_sfkl/checkpoint-314/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d88d3ddd2a1fe81f4cf32eb7f116cfcaecff483e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "down_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/adapter_model.safetensors b/qwen3/distillm_adaptive_sfkl/checkpoint-314/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5d32a50fe0b6e1b4144cd92235a437b2c4306cb3
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fd31ad3c06c6ba32fe311104350ef79b3b7f71d978406e20a89363c362ed6b8
+size 40422560
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/chat_template.jinja b/qwen3/distillm_adaptive_sfkl/checkpoint-314/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/distillm_state_rank0.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-314/distillm_state_rank0.pt
new file mode 100644
index 0000000000000000000000000000000000000000..c92ac759103ef6b2a0cb377b030a5ee272ca1dbc
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/distillm_state_rank0.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9208c430a8d43fa0e53ede8a6e5eab75fb565f1ad94e1b66dd5a8fad4f2d279d
+size 9403
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/distillm_state_rank1.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-314/distillm_state_rank1.pt
new file mode 100644
index 0000000000000000000000000000000000000000..16e0bb3962fa3f007370bb097abbec17d52cdc05
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/distillm_state_rank1.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06c0f9ff15c6f6d1dc5269643a7219fdaa4612ecfea5313c5b9359e6d4f4e8a3
+size 9339
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..014b686c47a8d627fb2d342a3bb19f6c4ce5bf13
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bb7a56206ec184dbcfe3370795df6c6b26b16cfe15e4a5c9d96dcf0ae02e829
+size 121135557
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..95f042ba4456ef3e294347567a93e5d6fb71fa21
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40bb47e84ae23c36d478d29a33a002027f11197936156de3c8d0b5aadc1b533d
+size 121135749
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/mp_rank_00_model_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..545a097c0a57ff9365f05b2c6d9485eafd8c0ab0
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/global_step314/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b88bce20f8997184325ad4c5820b1f0ba9771b61b725ec98f3f53ed6da5bc319
+size 351841765
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/latest b/qwen3/distillm_adaptive_sfkl/checkpoint-314/latest
new file mode 100644
index 0000000000000000000000000000000000000000..a9786513b1a514db2c3a9da7408844635663d55d
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/latest
@@ -0,0 +1 @@
+global_step314
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/rng_state_0.pth b/qwen3/distillm_adaptive_sfkl/checkpoint-314/rng_state_0.pth
new file mode 100644
index 0000000000000000000000000000000000000000..4e26866d298b00dd623ba4a091e965dfa4ed3760
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/rng_state_0.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:addc9cd09c422791daca0c22561ff8b62387ed0000b8a8d0870193b03682d5ea
+size 14917
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/rng_state_1.pth b/qwen3/distillm_adaptive_sfkl/checkpoint-314/rng_state_1.pth
new file mode 100644
index 0000000000000000000000000000000000000000..4b37f01bd6e69f7ee3e287a310189b0b1ed52236
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/rng_state_1.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:224ceacd3737b01636f96f069023026c038f76b816e103d596860a265f268d13
+size 14917
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/scheduler.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-314/scheduler.pt
new file mode 100644
index 0000000000000000000000000000000000000000..a3cb460802cda03bf0833a98a6f1428a3fd4f585
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/scheduler.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17e254b5642bb75983fd5105bec79b35376cfbc99d29702287668b8767d46149
+size 1465
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/tokenizer.json b/qwen3/distillm_adaptive_sfkl/checkpoint-314/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/tokenizer.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/tokenizer_config.json b/qwen3/distillm_adaptive_sfkl/checkpoint-314/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..6623da03457d79d2cd4af9557e9f31407cdb835c
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/tokenizer_config.json
@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_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|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/trainer_state.json b/qwen3/distillm_adaptive_sfkl/checkpoint-314/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..b1518143e28e6df62fb5dfacdb228ea2cdc68e29
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/trainer_state.json
@@ -0,0 +1,169 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.0,
+  "eval_steps": 500,
+  "global_step": 314,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "distillm_initial_generator_count": 1712.0,
+      "distillm_initial_generator_exact_match": 0.0,
+      "distillm_initial_generator_rouge1": 16.2149,
+      "distillm_initial_generator_rouge2": 3.2276,
+      "distillm_initial_generator_rougeL": 11.6725,
+      "distillm_initial_loss": 2.494271755218506,
+      "distillm_initial_runtime": 2988.1286,
+      "distillm_initial_samples_per_second": 1.143,
+      "distillm_initial_selector_accuracy": 52.278,
+      "distillm_initial_selector_count": 1712.0,
+      "distillm_initial_steps_per_second": 0.036,
+      "epoch": 0,
+      "step": 0
+    },
+    {
+      "epoch": 0.06382130035899482,
+      "grad_norm": 1.4179933071136475,
+      "learning_rate": 9.996390385474569e-05,
+      "loss": 0.360275936126709,
+      "step": 20
+    },
+    {
+      "epoch": 0.12764260071798963,
+      "grad_norm": 0.6145927309989929,
+      "learning_rate": 9.984797509888031e-05,
+      "loss": 0.0921450138092041,
+      "step": 40
+    },
+    {
+      "epoch": 0.19146390107698444,
+      "grad_norm": 0.949298083782196,
+      "learning_rate": 9.965229938050925e-05,
+      "loss": 0.06155582666397095,
+      "step": 60
+    },
+    {
+      "epoch": 0.25528520143597927,
+      "grad_norm": 0.32781288027763367,
+      "learning_rate": 9.937719005663432e-05,
+      "loss": 0.03148526251316071,
+      "step": 80
+    },
+    {
+      "epoch": 0.3191065017949741,
+      "grad_norm": 0.42831170558929443,
+      "learning_rate": 9.902308769000724e-05,
+      "loss": 0.0304107666015625,
+      "step": 100
+    },
+    {
+      "epoch": 0.3829278021539689,
+      "grad_norm": 0.37467315793037415,
+      "learning_rate": 9.859055934360789e-05,
+      "loss": 0.018133948743343353,
+      "step": 120
+    },
+    {
+      "epoch": 0.4467491025129637,
+      "grad_norm": 0.26231899857521057,
+      "learning_rate": 9.808029767254437e-05,
+      "loss": 0.023359841108322142,
+      "step": 140
+    },
+    {
+      "epoch": 0.5105704028719585,
+      "grad_norm": 0.23304614424705505,
+      "learning_rate": 9.749311981482817e-05,
+      "loss": 0.015377765893936158,
+      "step": 160
+    },
+    {
+      "epoch": 0.5743917032309533,
+      "grad_norm": 4.825753688812256,
+      "learning_rate": 9.682996608280154e-05,
+      "loss": 0.0134853795170784,
+      "step": 180
+    },
+    {
+      "epoch": 0.6382130035899481,
+      "grad_norm": 0.29226061701774597,
+      "learning_rate": 9.609189845731207e-05,
+      "loss": 0.022103385627269746,
+      "step": 200
+    },
+    {
+      "epoch": 0.702034303948943,
+      "grad_norm": 0.1759956330060959,
+      "learning_rate": 9.52800988870462e-05,
+      "loss": 0.011787685751914977,
+      "step": 220
+    },
+    {
+      "epoch": 0.7658556043079378,
+      "grad_norm": 0.18942032754421234,
+      "learning_rate": 9.439586739574513e-05,
+      "loss": 0.009058185666799546,
+      "step": 240
+    },
+    {
+      "epoch": 0.8296769046669326,
+      "grad_norm": 0.2657412886619568,
+      "learning_rate": 9.344062000033414e-05,
+      "loss": 0.009212753176689148,
+      "step": 260
+    },
+    {
+      "epoch": 0.8934982050259274,
+      "grad_norm": 0.28189197182655334,
+      "learning_rate": 9.241588644329919e-05,
+      "loss": 0.009374459087848664,
+      "step": 280
+    },
+    {
+      "epoch": 0.9573195053849222,
+      "grad_norm": 0.22191843390464783,
+      "learning_rate": 9.132330774294252e-05,
+      "loss": 0.009671492874622345,
+      "step": 300
+    },
+    {
+      "epoch": 1.0,
+      "eval_generator_count": 1712.0,
+      "eval_generator_exact_match": 90.0701,
+      "eval_generator_rouge1": 99.4297,
+      "eval_generator_rouge2": 99.1758,
+      "eval_generator_rougeL": 99.0503,
+      "eval_loss": 0.007805492263287306,
+      "eval_runtime": 601.5417,
+      "eval_samples_per_second": 5.675,
+      "eval_selector_accuracy": 95.5607,
+      "eval_selector_count": 1712.0,
+      "eval_steps_per_second": 0.178,
+      "step": 314
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 1570,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.172842003562496e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/training_args.bin b/qwen3/distillm_adaptive_sfkl/checkpoint-314/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..9965b3704c5694d942c823f03f32ef4aa365d5a2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4080ce6c122d894dce2a6242fb58ae879822809092a9484cf3c306a91aacf376
+size 7185
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-314/zero_to_fp32.py b/qwen3/distillm_adaptive_sfkl/checkpoint-314/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..f40f1f0acbbff30e8e8d29906b56aae4e8b8a138
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-314/zero_to_fp32.py
@@ -0,0 +1,790 @@
+#!/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, AUTOEP_LAYERS_KEY,
+                                            AUTOEP_LAYERS_KEY_LEGACY, AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY,
+                                            AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT)
+
+
+@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 _has_autoep_zero3_partitioned_metadata(state_dict):
+    autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY)
+    if autoep_layers is None:
+        autoep_layers = state_dict.get(AUTOEP_LAYERS_KEY_LEGACY)
+    if not isinstance(autoep_layers, list):
+        return False
+    return any(
+        isinstance(entry, dict)
+        and entry.get(AUTOEP_ZERO3_EXPERT_STATE_FORMAT_KEY) == AUTOEP_ZERO3_PARTITIONED_EXPERT_STATE_FORMAT
+        for entry in autoep_layers)
+
+
+def _raise_if_autoep_zero3_partitioned_state(state_dict):
+    if _has_autoep_zero3_partitioned_metadata(state_dict):
+        raise NotImplementedError("zero_to_fp32 does not support AutoEP ZeRO-3 partition-native checkpoints. "
+                                  "AutoEP expert parameters are partitioned over expert replica groups, so "
+                                  "global data-parallel consolidation would produce incomplete expert tensors. "
+                                  "Use ds_to_universal.py for expert-aware conversion.")
+
+
+def _raise_if_autoep_zero3_partitioned_checkpoint(model_files):
+    for file in model_files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+        _raise_if_autoep_zero3_partitioned_state(state_dict)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    # parse_model_states rejects AutoEP ZeRO-3 partition-native checkpoints
+    # before the expensive optimizer-shard load below.
+    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}')
+
+    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}")
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/README.md b/qwen3/distillm_adaptive_sfkl/checkpoint-628/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..776f98d43659dd923b8c3e5649c5ae66d1f99f3e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/README.md
@@ -0,0 +1,208 @@
+---
+base_model: Qwen/Qwen3-0.6B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:Qwen/Qwen3-0.6B
+- llama-factory
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/adapter_config.json b/qwen3/distillm_adaptive_sfkl/checkpoint-628/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d88d3ddd2a1fe81f4cf32eb7f116cfcaecff483e
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/adapter_config.json
@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "down_proj",
+    "v_proj",
+    "q_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/adapter_model.safetensors b/qwen3/distillm_adaptive_sfkl/checkpoint-628/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..3a587664cabe2cd3c49db2e6ef31d40fb9cdb05a
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25185e10cd53d2e1d7ac55987d0e84b185172802124cf27164ddebc2678aaa66
+size 40422560
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/chat_template.jinja b/qwen3/distillm_adaptive_sfkl/checkpoint-628/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/chat_template.jinja
@@ -0,0 +1,89 @@
+{%- 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 message.content is string 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.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.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' }}
+        {{- 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' }}
+    {%- endif %}
+{%- endif %}
\ No newline at end of file
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/distillm_state_rank0.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-628/distillm_state_rank0.pt
new file mode 100644
index 0000000000000000000000000000000000000000..a9bffbbbd85eed6d205abf4ba497b58a9c2ab686
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/distillm_state_rank0.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9083bb0185b673b3026a71a4e1b07808667f534cb608256c2f138888949155e4
+size 9403
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/distillm_state_rank1.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-628/distillm_state_rank1.pt
new file mode 100644
index 0000000000000000000000000000000000000000..cb46275912c78d8615a32230c877ae737b79e87d
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/distillm_state_rank1.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ade08e81abc31e53e5fa43cc5e43d33641c8cb156fb645f56ff2473327d9a0d
+size 9339
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..3c1cf86e1a1a6c8cc5c07b40e8d1d801e5fe50c2
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:becf8dbf056d2918d68224f88e165a67737125ccb57bf085549418cf2348d419
+size 121135557
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..7bfd77c468e563a6e19658f7aaf9ac4651e830a3
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:851ff806acb5365b9157d74e461ed2c106e766cbcfe17f279f195e0ffec76520
+size 121135749
diff --git a/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/mp_rank_00_model_states.pt b/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..f41cbda545d04506eaa0b3ff3498b13a1ca5dbb0
--- /dev/null
+++ b/qwen3/distillm_adaptive_sfkl/checkpoint-628/global_step628/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b7109baa0ea62e4c188d495711855d5589a88cf6ff9e8a19c40f7056aa5dd69
+size 351841765