Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ 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
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,203 @@
----
-license: mit
----
+---
+license: apache-2.0
+base_model: Qwen/Qwen3-0.6B
+tags:
+  - sat
+  - satisfiability
+  - cube-and-conquer
+  - variable-selection
+  - combinatorial-optimization
+datasets:
+  - Yale-ROSE/SAT-VarSelector-Distilled
+language:
+  - en
+pipeline_tag: text-classification
+---
+
+# Qwen3-0.6B-SAT-VarSelector-Distilled
+
+A **Qwen3-0.6B** model fine-tuned for **SAT variable selection** in the Cube-and-Conquer (CnC) framework. Given a CNF formula state, the model predicts which variable to branch/cube on next.
+
+## Model Description
+
+This model implements a **masked classification head** on top of Qwen3-0.6B to select branching variables for SAT solving. Unlike traditional heuristics (e.g., VSIDS), it learns from expert solver traces to make informed variable selection decisions.
+
+### Key Features
+
+- **Task**: Variable selection for SAT Cube-and-Conquer
+- **Architecture**: Qwen3-0.6B backbone + classification head (601 classes for variables 0-600)
+- **Training**: Supervised fine-tuning on distilled expert data
+- **Output**: Integer variable ID to branch on
+
+## Training Details
+
+| Attribute | Value |
+|-----------|-------|
+| Base Model | `Qwen/Qwen3-0.6B` |
+| Training Dataset | Distilled from GPT expert traces |
+| Best Checkpoint | Step 410 (Epoch ~6.7) |
+| **Eval Accuracy** | **14.75%** |
+| Eval Loss | 3.789 |
+| Training Time | ~53 minutes (8×H100 GPUs) |
+
+### Performance Context
+
+- **Random Baseline**: ~1-2% accuracy (depends on number of valid variables)
+- **This Model**: 14.75% accuracy = **~7-15× better than random**
+
+### Hyperparameters
+
+```yaml
+learning_rate: 5e-6
+warmup_ratio: 0.1
+num_train_epochs: 8
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+max_length: 8192
+max_vars: 600
+optimizer: AdamW
+scheduler: cosine
+deepspeed: ZeRO-3
+```
+
+## Usage
+
+### Loading the Model
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+import torch.nn as nn
+import re
+
+class QwenVarClassifier(nn.Module):
+    def __init__(self, base_model, max_vars=600):
+        super().__init__()
+        self.base = base_model
+        hidden_size = base_model.config.hidden_size
+        self.norm = nn.LayerNorm(hidden_size)
+        self.head = nn.Linear(hidden_size, max_vars + 1)
+    
+    def forward(self, input_ids, attention_mask=None):
+        outputs = self.base(input_ids, attention_mask=attention_mask, output_hidden_states=True)
+        hidden = outputs.hidden_states[-1]  # [B, seq, hidden]
+        
+        # Pool at last non-pad token
+        if attention_mask is not None:
+            lengths = attention_mask.sum(dim=1) - 1
+            pooled = hidden[torch.arange(hidden.size(0)), lengths]
+        else:
+            pooled = hidden[:, -1, :]
+        
+        pooled = self.norm(pooled)
+        logits = self.head(pooled)
+        return logits
+
+# Load
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
+base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")
+model = QwenVarClassifier(base_model, max_vars=600)
+
+# Load fine-tuned weights
+state_dict = torch.load("pytorch_model.bin", map_location="cpu")
+model.load_state_dict(state_dict)
+model.eval()
+```
+
+### Inference
+
+```python
+def get_valid_vars(cnf_text, max_vars=600):
+    """Extract valid variable IDs from CNF text."""
+    valid = set()
+    for line in cnf_text.strip().split('\n'):
+        if line.startswith('c') or line.startswith('p'):
+            continue
+        for tok in line.split():
+            try:
+                lit = int(tok)
+                if lit != 0:
+                    valid.add(abs(lit))
+            except ValueError:
+                pass
+    return valid
+
+def predict_variable(cnf_text, model, tokenizer, max_vars=600):
+    """Predict the next variable to branch on."""
+    inputs = tokenizer(cnf_text, return_tensors="pt", truncation=True, max_length=8192)
+    
+    with torch.no_grad():
+        logits = model(inputs["input_ids"], inputs["attention_mask"])
+    
+    # Mask invalid variables
+    valid_vars = get_valid_vars(cnf_text, max_vars)
+    mask = torch.zeros(max_vars + 1, dtype=torch.bool)
+    for v in valid_vars:
+        if 1 <= v <= max_vars:
+            mask[v] = True
+    
+    logits[0, ~mask] = -1e4
+    predicted_var = logits.argmax(dim=-1).item()
+    
+    return predicted_var
+
+# Example
+cnf_text = """p cnf 100 200
+1 -2 3 0
+-1 4 -5 0
+2 5 6 0
+"""
+
+var = predict_variable(cnf_text, model, tokenizer)
+print(f"Predicted variable: {var}")
+```
+
+## Architecture Details
+
+### Why Masked Classification?
+
+The valid action set is **state-dependent**: not all variables are valid at every step.
+
+- Some variables may be eliminated during simplification
+- Some may be out of range for the specific instance
+
+We use **masked softmax**:
+1. Model outputs logits for all 601 classes (0-600)
+2. Invalid variables get logits set to `-1e4`
+3. Softmax only assigns probability to valid variables
+4. Training uses masked cross-entropy loss
+
+### Why Pool the Last Token?
+
+The last non-pad token has attended to the entire CNF sequence through causal attention, making it a natural summary representation.
+
+### Why LayerNorm Before the Head?
+
+Qwen's hidden states can have large magnitudes. LayerNorm stabilizes the input to the classification head.
+
+## Limitations
+
+- Maximum 600 variables (configurable during training)
+- Maximum sequence length 8192 tokens
+- Trained on specific CNF distribution; may not generalize to all SAT instances
+- Accuracy metric is strict exact-match; the model may predict "good" variables even when not matching the expert label exactly
+
+## Citation
+
+```bibtex
+@misc{qwen-sat-varselector,
+  title={Qwen3-0.6B-SAT-VarSelector-Distilled},
+  author={Yale-ROSE},
+  year={2026},
+  publisher={Hugging Face},
+  url={https://huggingface.co/Yale-ROSE/Qwen3-0.6B-SAT-VarSelector-Distilled}
+}
+```
+
+## Related Models
+
+- [Yale-ROSE/Qwen3-4B-SAT-VarSelector](https://huggingface.co/Yale-ROSE/Qwen3-4B-SAT-VarSelector) - Larger 4B parameter version
+
+## License
+
+Apache 2.0 (following the base Qwen3 license)

added_tokens.json

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

chat_template.jinja

@@ -0,0 +1,89 @@
+{%- 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 %}

global_step410/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ccbbadc33b4d1d3d4347c188c93f82b20051127f3045a5fc6747c5df6214ffb
+size 895007429

global_step410/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27fadac16f054f5a90074ed66b8c7c8c2e3316186d706d3e7afe101ca1437f86
+size 895007429

global_step410/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3d9982e9075caf82c0e3c7ba4dd7545403fdae8b36ef5f27fbeaf9751f1a2e4
+size 895007429

global_step410/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65a08717c58671cf8a36a45d1c02bb95457e049ddb2e01ace410feef7b85b24b
+size 895007429

global_step410/bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:630040bd55b65dca6af42d4c2ba6d6ae343b56d56e6bddeb6196143fc0a68a46
+size 895007429

global_step410/bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c15add7b826bf73a0451326a02b02e542d984d3a6437871d6fbdbda76ae4ae65
+size 895007429

global_step410/bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49aacc04bcb08710c05355788c061bce1c3daa5eb700efebf6f25c4378786ebd
+size 895007429

global_step410/bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77aca66738052ab70e6c446b1932bd7890e61b6162a0a3e5ff8895a0b56a995e
+size 895007429

global_step410/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec781da82ecd6f7882b2e619a9eec4186c8f05b836819871cf7002f731809584
+size 168971

global_step410/zero_pp_rank_1_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45bab9476dcd282273efc27c99b29966b015ca3f0e677e2aac49ebcd1f98d05a
+size 168907

global_step410/zero_pp_rank_2_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28183b923bbe3866413add3a2990c627bf492b9f8079a30657aa167169c02d9a
+size 168907

global_step410/zero_pp_rank_3_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d15c29f8ee05b89067e74db99873b76221035db486799f5c2fd7bb3bf047d8a1
+size 168907

global_step410/zero_pp_rank_4_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:963b1e4e28d6e9e3e51a1572b868cee7c46f09c8f18d5b1479a586049266b350
+size 168907

global_step410/zero_pp_rank_5_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e6ede8583c213cd919570c058669a65481053b91772598b05aad658528464c3
+size 168907

global_step410/zero_pp_rank_6_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e66b39cd5c3ee977b30ef84516bb522a4e2a7909853f8bd6ea2c3a1eebcf2bac
+size 168907

global_step410/zero_pp_rank_7_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:040eae58e0add040cd9f21a859527833cd4aa1e269aa0d7d6f797d6b613db2d7
+size 168907

latest

@@ -0,0 +1 @@
+global_step410

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fdc9339663edaba63e99a0868a29bc05a5642f7875b29957c36212233d88be1
+size 1193440459

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a717e796866a56d9780e272a25aebb426950bce5aa9b5d2497f05b461ff76729
+size 16325

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:979715f4cd80f53fd94508599aca8831163b40d525858bcb11e6454ddb8251c0
+size 16389

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2297b99224f632707d9574fcf1aaa010691c398f9f0f7339d357ff558d79d21
+size 16389

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8977ecd4d73fa6ee50b4f9d14ee8ff5d28167f5540b3271d7b4d4843f152656b
+size 16389

rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aba176670c628753c0cc24485ddacff4538fcc08ea4f35777a7eb84b54dd2922
+size 16389

rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:992b6b333d0e5866749e05e760309ea00aed67c9810d11da4eac4381cc34e022
+size 16389

rng_state_6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac5f361548be76f7934e995622866df9cb9b72bd19b7e323c942a9026e21565f
+size 16389

rng_state_7.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1868f50bff5601996407296d2060a8bb4b443c68d73d4b2442b76fa40389c757
+size 16389

scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7d204098f3bd572f36fccc74c22ca679a5de55f9aee6ea69ea3718f1765f1d7
+size 1465

special_tokens_map.json

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

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7a6a993d40b42d517297bb247ff66679e5bc9dd7a5143be0620faf210b42861
+size 11422753

tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

trainer_state.json

@@ -0,0 +1,1059 @@
+{
+  "best_global_step": 410,
+  "best_metric": 0.14746543778801843,
+  "best_model_checkpoint": "out_qwen_0.6b_sft_distilled/checkpoint-410",
+  "epoch": 6.721311475409836,
+  "eval_steps": 10,
+  "global_step": 410,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.08196721311475409,
+      "grad_norm": 131.30086444913783,
+      "learning_rate": 4.0816326530612243e-07,
+      "loss": 40.5707,
+      "step": 5,
+      "true_loss": 4.9328
+    },
+    {
+      "epoch": 0.16393442622950818,
+      "grad_norm": 126.20257823342612,
+      "learning_rate": 9.183673469387756e-07,
+      "loss": 40.4496,
+      "step": 10,
+      "true_loss": 4.9664
+    },
+    {
+      "epoch": 0.16393442622950818,
+      "eval_accuracy": 0.004608294930875576,
+      "eval_loss": 5.019261360168457,
+      "eval_runtime": 8.3229,
+      "eval_samples_per_second": 52.145,
+      "eval_steps_per_second": 6.608,
+      "step": 10
+    },
+    {
+      "epoch": 0.2459016393442623,
+      "grad_norm": 136.9898759427637,
+      "learning_rate": 1.4285714285714286e-06,
+      "loss": 40.502,
+      "step": 15,
+      "true_loss": 4.9781
+    },
+    {
+      "epoch": 0.32786885245901637,
+      "grad_norm": 134.72385927894973,
+      "learning_rate": 1.938775510204082e-06,
+      "loss": 39.6527,
+      "step": 20,
+      "true_loss": 4.9395
+    },
+    {
+      "epoch": 0.32786885245901637,
+      "eval_accuracy": 0.004608294930875576,
+      "eval_loss": 4.9670939445495605,
+      "eval_runtime": 7.9853,
+      "eval_samples_per_second": 54.35,
+      "eval_steps_per_second": 6.888,
+      "step": 20
+    },
+    {
+      "epoch": 0.4098360655737705,
+      "grad_norm": 127.00331303837481,
+      "learning_rate": 2.4489795918367347e-06,
+      "loss": 38.9625,
+      "step": 25,
+      "true_loss": 4.9176
+    },
+    {
+      "epoch": 0.4918032786885246,
+      "grad_norm": 111.05596081816127,
+      "learning_rate": 2.959183673469388e-06,
+      "loss": 38.8523,
+      "step": 30,
+      "true_loss": 4.9895
+    },
+    {
+      "epoch": 0.4918032786885246,
+      "eval_accuracy": 0.009216589861751152,
+      "eval_loss": 4.8343892097473145,
+      "eval_runtime": 8.399,
+      "eval_samples_per_second": 51.673,
+      "eval_steps_per_second": 6.548,
+      "step": 30
+    },
+    {
+      "epoch": 0.5737704918032787,
+      "grad_norm": 127.72432425565634,
+      "learning_rate": 3.469387755102041e-06,
+      "loss": 38.366,
+      "step": 35,
+      "true_loss": 4.7629
+    },
+    {
+      "epoch": 0.6557377049180327,
+      "grad_norm": 142.80923150687772,
+      "learning_rate": 3.979591836734694e-06,
+      "loss": 38.0703,
+      "step": 40,
+      "true_loss": 4.8902
+    },
+    {
+      "epoch": 0.6557377049180327,
+      "eval_accuracy": 0.055299539170506916,
+      "eval_loss": 4.662118434906006,
+      "eval_runtime": 8.4731,
+      "eval_samples_per_second": 51.221,
+      "eval_steps_per_second": 6.491,
+      "step": 40
+    },
+    {
+      "epoch": 0.7377049180327869,
+      "grad_norm": 179.79048183030648,
+      "learning_rate": 4.489795918367348e-06,
+      "loss": 36.2355,
+      "step": 45,
+      "true_loss": 4.6008
+    },
+    {
+      "epoch": 0.819672131147541,
+      "grad_norm": 161.88532993207818,
+      "learning_rate": 5e-06,
+      "loss": 37.6594,
+      "step": 50,
+      "true_loss": 5.0719
+    },
+    {
+      "epoch": 0.819672131147541,
+      "eval_accuracy": 0.08294930875576037,
+      "eval_loss": 4.533751964569092,
+      "eval_runtime": 8.6568,
+      "eval_samples_per_second": 50.134,
+      "eval_steps_per_second": 6.353,
+      "step": 50
+    },
+    {
+      "epoch": 0.9016393442622951,
+      "grad_norm": 151.13368850588557,
+      "learning_rate": 4.943052391799545e-06,
+      "loss": 36.9232,
+      "step": 55,
+      "true_loss": 4.5199
+    },
+    {
+      "epoch": 0.9836065573770492,
+      "grad_norm": 140.76185960403066,
+      "learning_rate": 4.886104783599089e-06,
+      "loss": 35.987,
+      "step": 60,
+      "true_loss": 4.3872
+    },
+    {
+      "epoch": 0.9836065573770492,
+      "eval_accuracy": 0.08064516129032258,
+      "eval_loss": 4.438742160797119,
+      "eval_runtime": 8.6755,
+      "eval_samples_per_second": 50.026,
+      "eval_steps_per_second": 6.34,
+      "step": 60
+    },
+    {
+      "epoch": 1.0655737704918034,
+      "grad_norm": 141.24088109006144,
+      "learning_rate": 4.829157175398634e-06,
+      "loss": 35.8602,
+      "step": 65,
+      "true_loss": 4.6277
+    },
+    {
+      "epoch": 1.1475409836065573,
+      "grad_norm": 147.2087879935333,
+      "learning_rate": 4.772209567198178e-06,
+      "loss": 35.6707,
+      "step": 70,
+      "true_loss": 4.1293
+    },
+    {
+      "epoch": 1.1475409836065573,
+      "eval_accuracy": 0.08294930875576037,
+      "eval_loss": 4.400156497955322,
+      "eval_runtime": 8.477,
+      "eval_samples_per_second": 51.198,
+      "eval_steps_per_second": 6.488,
+      "step": 70
+    },
+    {
+      "epoch": 1.2295081967213115,
+      "grad_norm": 153.44325788545655,
+      "learning_rate": 4.7152619589977225e-06,
+      "loss": 34.8141,
+      "step": 75,
+      "true_loss": 4.2855
+    },
+    {
+      "epoch": 1.3114754098360657,
+      "grad_norm": 167.21816733466747,
+      "learning_rate": 4.658314350797267e-06,
+      "loss": 35.2213,
+      "step": 80,
+      "true_loss": 4.4329
+    },
+    {
+      "epoch": 1.3114754098360657,
+      "eval_accuracy": 0.08755760368663594,
+      "eval_loss": 4.335811614990234,
+      "eval_runtime": 8.385,
+      "eval_samples_per_second": 51.759,
+      "eval_steps_per_second": 6.559,
+      "step": 80
+    },
+    {
+      "epoch": 1.3934426229508197,
+      "grad_norm": 170.8547829819815,
+      "learning_rate": 4.601366742596811e-06,
+      "loss": 34.6405,
+      "step": 85,
+      "true_loss": 4.2484
+    },
+    {
+      "epoch": 1.4754098360655736,
+      "grad_norm": 169.3659623205823,
+      "learning_rate": 4.544419134396356e-06,
+      "loss": 34.6621,
+      "step": 90,
+      "true_loss": 4.5617
+    },
+    {
+      "epoch": 1.4754098360655736,
+      "eval_accuracy": 0.10829493087557604,
+      "eval_loss": 4.2690629959106445,
+      "eval_runtime": 8.5736,
+      "eval_samples_per_second": 50.62,
+      "eval_steps_per_second": 6.415,
+      "step": 90
+    },
+    {
+      "epoch": 1.5573770491803278,
+      "grad_norm": 181.5354290153052,
+      "learning_rate": 4.4874715261959e-06,
+      "loss": 33.5805,
+      "step": 95,
+      "true_loss": 4.2242
+    },
+    {
+      "epoch": 1.639344262295082,
+      "grad_norm": 183.60203299469077,
+      "learning_rate": 4.4305239179954446e-06,
+      "loss": 34.323,
+      "step": 100,
+      "true_loss": 4.3387
+    },
+    {
+      "epoch": 1.639344262295082,
+      "eval_accuracy": 0.10368663594470046,
+      "eval_loss": 4.227273464202881,
+      "eval_runtime": 8.642,
+      "eval_samples_per_second": 50.22,
+      "eval_steps_per_second": 6.364,
+      "step": 100
+    },
+    {
+      "epoch": 1.721311475409836,
+      "grad_norm": 170.66504985673134,
+      "learning_rate": 4.373576309794989e-06,
+      "loss": 32.0365,
+      "step": 105,
+      "true_loss": 3.9484
+    },
+    {
+      "epoch": 1.8032786885245902,
+      "grad_norm": 185.72351413214275,
+      "learning_rate": 4.316628701594533e-06,
+      "loss": 33.875,
+      "step": 110,
+      "true_loss": 4.1586
+    },
+    {
+      "epoch": 1.8032786885245902,
+      "eval_accuracy": 0.11290322580645161,
+      "eval_loss": 4.191784381866455,
+      "eval_runtime": 8.026,
+      "eval_samples_per_second": 54.074,
+      "eval_steps_per_second": 6.853,
+      "step": 110
+    },
+    {
+      "epoch": 1.8852459016393444,
+      "grad_norm": 187.82807734370155,
+      "learning_rate": 4.259681093394078e-06,
+      "loss": 33.4,
+      "step": 115,
+      "true_loss": 4.2566
+    },
+    {
+      "epoch": 1.9672131147540983,
+      "grad_norm": 177.04120464657095,
+      "learning_rate": 4.202733485193622e-06,
+      "loss": 32.768,
+      "step": 120,
+      "true_loss": 4.1711
+    },
+    {
+      "epoch": 1.9672131147540983,
+      "eval_accuracy": 0.10138248847926268,
+      "eval_loss": 4.166411876678467,
+      "eval_runtime": 8.3731,
+      "eval_samples_per_second": 51.833,
+      "eval_steps_per_second": 6.569,
+      "step": 120
+    },
+    {
+      "epoch": 2.0491803278688523,
+      "grad_norm": 175.70043001685525,
+      "learning_rate": 4.145785876993167e-06,
+      "loss": 32.9279,
+      "step": 125,
+      "true_loss": 4.2443
+    },
+    {
+      "epoch": 2.1311475409836067,
+      "grad_norm": 211.66606785751807,
+      "learning_rate": 4.088838268792711e-06,
+      "loss": 31.4897,
+      "step": 130,
+      "true_loss": 3.9807
+    },
+    {
+      "epoch": 2.1311475409836067,
+      "eval_accuracy": 0.09907834101382489,
+      "eval_loss": 4.117466449737549,
+      "eval_runtime": 9.4127,
+      "eval_samples_per_second": 46.108,
+      "eval_steps_per_second": 5.843,
+      "step": 130
+    },
+    {
+      "epoch": 2.2131147540983607,
+      "grad_norm": 243.03840327182303,
+      "learning_rate": 4.0318906605922555e-06,
+      "loss": 31.5314,
+      "step": 135,
+      "true_loss": 3.9421
+    },
+    {
+      "epoch": 2.2950819672131146,
+      "grad_norm": 220.78266878880433,
+      "learning_rate": 3.9749430523918e-06,
+      "loss": 31.4336,
+      "step": 140,
+      "true_loss": 3.7703
+    },
+    {
+      "epoch": 2.2950819672131146,
+      "eval_accuracy": 0.1336405529953917,
+      "eval_loss": 4.065812110900879,
+      "eval_runtime": 8.6071,
+      "eval_samples_per_second": 50.423,
+      "eval_steps_per_second": 6.39,
+      "step": 140
+    },
+    {
+      "epoch": 2.3770491803278686,
+      "grad_norm": 196.6859886983302,
+      "learning_rate": 3.917995444191344e-06,
+      "loss": 31.3979,
+      "step": 145,
+      "true_loss": 3.8875
+    },
+    {
+      "epoch": 2.459016393442623,
+      "grad_norm": 209.14552747888038,
+      "learning_rate": 3.861047835990889e-06,
+      "loss": 31.5513,
+      "step": 150,
+      "true_loss": 3.65
+    },
+    {
+      "epoch": 2.459016393442623,
+      "eval_accuracy": 0.12903225806451613,
+      "eval_loss": 4.045776844024658,
+      "eval_runtime": 8.5013,
+      "eval_samples_per_second": 51.051,
+      "eval_steps_per_second": 6.47,
+      "step": 150
+    },
+    {
+      "epoch": 2.540983606557377,
+      "grad_norm": 205.92078421639843,
+      "learning_rate": 3.804100227790433e-06,
+      "loss": 31.4361,
+      "step": 155,
+      "true_loss": 3.7869
+    },
+    {
+      "epoch": 2.6229508196721314,
+      "grad_norm": 219.604580334424,
+      "learning_rate": 3.7471526195899776e-06,
+      "loss": 31.6139,
+      "step": 160,
+      "true_loss": 3.9903
+    },
+    {
+      "epoch": 2.6229508196721314,
+      "eval_accuracy": 0.1175115207373272,
+      "eval_loss": 4.006894588470459,
+      "eval_runtime": 8.1353,
+      "eval_samples_per_second": 53.348,
+      "eval_steps_per_second": 6.761,
+      "step": 160
+    },
+    {
+      "epoch": 2.7049180327868854,
+      "grad_norm": 221.00907729918933,
+      "learning_rate": 3.690205011389522e-06,
+      "loss": 31.0029,
+      "step": 165,
+      "true_loss": 3.9396
+    },
+    {
+      "epoch": 2.7868852459016393,
+      "grad_norm": 200.63567221965303,
+      "learning_rate": 3.6332574031890664e-06,
+      "loss": 30.4827,
+      "step": 170,
+      "true_loss": 4.0088
+    },
+    {
+      "epoch": 2.7868852459016393,
+      "eval_accuracy": 0.12442396313364056,
+      "eval_loss": 3.991098403930664,
+      "eval_runtime": 8.4689,
+      "eval_samples_per_second": 51.246,
+      "eval_steps_per_second": 6.494,
+      "step": 170
+    },
+    {
+      "epoch": 2.8688524590163933,
+      "grad_norm": 208.08258154815047,
+      "learning_rate": 3.5763097949886104e-06,
+      "loss": 30.7274,
+      "step": 175,
+      "true_loss": 3.8941
+    },
+    {
+      "epoch": 2.9508196721311473,
+      "grad_norm": 209.30561480651423,
+      "learning_rate": 3.519362186788155e-06,
+      "loss": 31.2113,
+      "step": 180,
+      "true_loss": 3.5918
+    },
+    {
+      "epoch": 2.9508196721311473,
+      "eval_accuracy": 0.1175115207373272,
+      "eval_loss": 3.9841365814208984,
+      "eval_runtime": 8.6637,
+      "eval_samples_per_second": 50.094,
+      "eval_steps_per_second": 6.348,
+      "step": 180
+    },
+    {
+      "epoch": 3.0327868852459017,
+      "grad_norm": 214.16917085440517,
+      "learning_rate": 3.4624145785876997e-06,
+      "loss": 31.0295,
+      "step": 185,
+      "true_loss": 3.5908
+    },
+    {
+      "epoch": 3.1147540983606556,
+      "grad_norm": 235.71482357282463,
+      "learning_rate": 3.405466970387244e-06,
+      "loss": 28.5428,
+      "step": 190,
+      "true_loss": 3.1442
+    },
+    {
+      "epoch": 3.1147540983606556,
+      "eval_accuracy": 0.12442396313364056,
+      "eval_loss": 3.974285364151001,
+      "eval_runtime": 8.0976,
+      "eval_samples_per_second": 53.596,
+      "eval_steps_per_second": 6.792,
+      "step": 190
+    },
+    {
+      "epoch": 3.19672131147541,
+      "grad_norm": 226.00461681805723,
+      "learning_rate": 3.3485193621867885e-06,
+      "loss": 29.5642,
+      "step": 195,
+      "true_loss": 3.5915
+    },
+    {
+      "epoch": 3.278688524590164,
+      "grad_norm": 233.95245184744422,
+      "learning_rate": 3.291571753986333e-06,
+      "loss": 29.4133,
+      "step": 200,
+      "true_loss": 3.518
+    },
+    {
+      "epoch": 3.278688524590164,
+      "eval_accuracy": 0.1313364055299539,
+      "eval_loss": 3.9394891262054443,
+      "eval_runtime": 8.5951,
+      "eval_samples_per_second": 50.494,
+      "eval_steps_per_second": 6.399,
+      "step": 200
+    },
+    {
+      "epoch": 3.360655737704918,
+      "grad_norm": 239.65764806983628,
+      "learning_rate": 3.2346241457858773e-06,
+      "loss": 29.0733,
+      "step": 205,
+      "true_loss": 3.4392
+    },
+    {
+      "epoch": 3.442622950819672,
+      "grad_norm": 251.9041666596011,
+      "learning_rate": 3.1776765375854217e-06,
+      "loss": 28.5826,
+      "step": 210,
+      "true_loss": 3.4814
+    },
+    {
+      "epoch": 3.442622950819672,
+      "eval_accuracy": 0.14055299539170507,
+      "eval_loss": 3.9186437129974365,
+      "eval_runtime": 8.3264,
+      "eval_samples_per_second": 52.123,
+      "eval_steps_per_second": 6.605,
+      "step": 210
+    },
+    {
+      "epoch": 3.5245901639344264,
+      "grad_norm": 236.49516114079387,
+      "learning_rate": 3.120728929384966e-06,
+      "loss": 29.3459,
+      "step": 215,
+      "true_loss": 3.6947
+    },
+    {
+      "epoch": 3.6065573770491803,
+      "grad_norm": 232.10135728893482,
+      "learning_rate": 3.0637813211845106e-06,
+      "loss": 29.1922,
+      "step": 220,
+      "true_loss": 3.7867
+    },
+    {
+      "epoch": 3.6065573770491803,
+      "eval_accuracy": 0.11981566820276497,
+      "eval_loss": 3.924264669418335,
+      "eval_runtime": 8.114,
+      "eval_samples_per_second": 53.488,
+      "eval_steps_per_second": 6.778,
+      "step": 220
+    },
+    {
+      "epoch": 3.6885245901639343,
+      "grad_norm": 248.9470759438453,
+      "learning_rate": 3.0068337129840546e-06,
+      "loss": 29.3719,
+      "step": 225,
+      "true_loss": 3.6912
+    },
+    {
+      "epoch": 3.7704918032786887,
+      "grad_norm": 242.5668319946587,
+      "learning_rate": 2.949886104783599e-06,
+      "loss": 29.0788,
+      "step": 230,
+      "true_loss": 3.7821
+    },
+    {
+      "epoch": 3.7704918032786887,
+      "eval_accuracy": 0.12672811059907835,
+      "eval_loss": 3.9106783866882324,
+      "eval_runtime": 8.4282,
+      "eval_samples_per_second": 51.494,
+      "eval_steps_per_second": 6.526,
+      "step": 230
+    },
+    {
+      "epoch": 3.8524590163934427,
+      "grad_norm": 258.17481354757325,
+      "learning_rate": 2.892938496583144e-06,
+      "loss": 28.8174,
+      "step": 235,
+      "true_loss": 3.3456
+    },
+    {
+      "epoch": 3.9344262295081966,
+      "grad_norm": 242.242892316644,
+      "learning_rate": 2.8359908883826882e-06,
+      "loss": 28.4613,
+      "step": 240,
+      "true_loss": 3.3777
+    },
+    {
+      "epoch": 3.9344262295081966,
+      "eval_accuracy": 0.12211981566820276,
+      "eval_loss": 3.8972718715667725,
+      "eval_runtime": 8.5279,
+      "eval_samples_per_second": 50.892,
+      "eval_steps_per_second": 6.449,
+      "step": 240
+    },
+    {
+      "epoch": 4.016393442622951,
+      "grad_norm": 249.19080164598319,
+      "learning_rate": 2.7790432801822326e-06,
+      "loss": 27.6907,
+      "step": 245,
+      "true_loss": 3.153
+    },
+    {
+      "epoch": 4.098360655737705,
+      "grad_norm": 290.1903243585321,
+      "learning_rate": 2.722095671981777e-06,
+      "loss": 26.9221,
+      "step": 250,
+      "true_loss": 3.2854
+    },
+    {
+      "epoch": 4.098360655737705,
+      "eval_accuracy": 0.1175115207373272,
+      "eval_loss": 3.8916871547698975,
+      "eval_runtime": 8.1923,
+      "eval_samples_per_second": 52.977,
+      "eval_steps_per_second": 6.714,
+      "step": 250
+    },
+    {
+      "epoch": 4.180327868852459,
+      "grad_norm": 284.2190675087926,
+      "learning_rate": 2.6651480637813215e-06,
+      "loss": 26.3206,
+      "step": 255,
+      "true_loss": 3.2506
+    },
+    {
+      "epoch": 4.262295081967213,
+      "grad_norm": 279.44061029669723,
+      "learning_rate": 2.608200455580866e-06,
+      "loss": 26.9143,
+      "step": 260,
+      "true_loss": 3.4412
+    },
+    {
+      "epoch": 4.262295081967213,
+      "eval_accuracy": 0.1152073732718894,
+      "eval_loss": 3.8716788291931152,
+      "eval_runtime": 8.5568,
+      "eval_samples_per_second": 50.72,
+      "eval_steps_per_second": 6.428,
+      "step": 260
+    },
+    {
+      "epoch": 4.344262295081967,
+      "grad_norm": 266.7499728501794,
+      "learning_rate": 2.5512528473804103e-06,
+      "loss": 26.4962,
+      "step": 265,
+      "true_loss": 3.5133
+    },
+    {
+      "epoch": 4.426229508196721,
+      "grad_norm": 260.78822590196535,
+      "learning_rate": 2.4943052391799547e-06,
+      "loss": 26.4451,
+      "step": 270,
+      "true_loss": 3.4446
+    },
+    {
+      "epoch": 4.426229508196721,
+      "eval_accuracy": 0.12211981566820276,
+      "eval_loss": 3.8494653701782227,
+      "eval_runtime": 8.4627,
+      "eval_samples_per_second": 51.284,
+      "eval_steps_per_second": 6.499,
+      "step": 270
+    },
+    {
+      "epoch": 4.508196721311475,
+      "grad_norm": 274.64077279341876,
+      "learning_rate": 2.437357630979499e-06,
+      "loss": 27.3761,
+      "step": 275,
+      "true_loss": 3.4453
+    },
+    {
+      "epoch": 4.590163934426229,
+      "grad_norm": 273.19374429695523,
+      "learning_rate": 2.3804100227790436e-06,
+      "loss": 27.4194,
+      "step": 280,
+      "true_loss": 3.1874
+    },
+    {
+      "epoch": 4.590163934426229,
+      "eval_accuracy": 0.11059907834101383,
+      "eval_loss": 3.846774101257324,
+      "eval_runtime": 8.5426,
+      "eval_samples_per_second": 50.804,
+      "eval_steps_per_second": 6.438,
+      "step": 280
+    },
+    {
+      "epoch": 4.672131147540983,
+      "grad_norm": 275.7740703438606,
+      "learning_rate": 2.323462414578588e-06,
+      "loss": 26.827,
+      "step": 285,
+      "true_loss": 3.3258
+    },
+    {
+      "epoch": 4.754098360655737,
+      "grad_norm": 274.35156164348126,
+      "learning_rate": 2.2665148063781324e-06,
+      "loss": 27.2383,
+      "step": 290,
+      "true_loss": 3.9754
+    },
+    {
+      "epoch": 4.754098360655737,
+      "eval_accuracy": 0.11981566820276497,
+      "eval_loss": 3.8400866985321045,
+      "eval_runtime": 8.4749,
+      "eval_samples_per_second": 51.21,
+      "eval_steps_per_second": 6.49,
+      "step": 290
+    },
+    {
+      "epoch": 4.836065573770492,
+      "grad_norm": 279.2476311679269,
+      "learning_rate": 2.209567198177677e-06,
+      "loss": 26.5837,
+      "step": 295,
+      "true_loss": 3.2732
+    },
+    {
+      "epoch": 4.918032786885246,
+      "grad_norm": 318.84649232107,
+      "learning_rate": 2.1526195899772212e-06,
+      "loss": 28.2213,
+      "step": 300,
+      "true_loss": 3.8258
+    },
+    {
+      "epoch": 4.918032786885246,
+      "eval_accuracy": 0.1336405529953917,
+      "eval_loss": 3.828943967819214,
+      "eval_runtime": 8.3482,
+      "eval_samples_per_second": 51.987,
+      "eval_steps_per_second": 6.588,
+      "step": 300
+    },
+    {
+      "epoch": 5.0,
+      "grad_norm": 281.39270714708135,
+      "learning_rate": 2.0956719817767656e-06,
+      "loss": 26.5872,
+      "step": 305,
+      "true_loss": 3.2237
+    },
+    {
+      "epoch": 5.081967213114754,
+      "grad_norm": 297.3835718534045,
+      "learning_rate": 2.03872437357631e-06,
+      "loss": 24.571,
+      "step": 310,
+      "true_loss": 3.346
+    },
+    {
+      "epoch": 5.081967213114754,
+      "eval_accuracy": 0.14516129032258066,
+      "eval_loss": 3.8257129192352295,
+      "eval_runtime": 8.4071,
+      "eval_samples_per_second": 51.623,
+      "eval_steps_per_second": 6.542,
+      "step": 310
+    },
+    {
+      "epoch": 5.163934426229508,
+      "grad_norm": 333.18631426892176,
+      "learning_rate": 1.9817767653758545e-06,
+      "loss": 25.2512,
+      "step": 315,
+      "true_loss": 3.185
+    },
+    {
+      "epoch": 5.245901639344262,
+      "grad_norm": 307.93977906152685,
+      "learning_rate": 1.924829157175399e-06,
+      "loss": 24.5613,
+      "step": 320,
+      "true_loss": 3.1895
+    },
+    {
+      "epoch": 5.245901639344262,
+      "eval_accuracy": 0.14285714285714285,
+      "eval_loss": 3.8230037689208984,
+      "eval_runtime": 8.2033,
+      "eval_samples_per_second": 52.906,
+      "eval_steps_per_second": 6.705,
+      "step": 320
+    },
+    {
+      "epoch": 5.327868852459017,
+      "grad_norm": 312.5942332522537,
+      "learning_rate": 1.8678815489749433e-06,
+      "loss": 25.0569,
+      "step": 325,
+      "true_loss": 3.3557
+    },
+    {
+      "epoch": 5.409836065573771,
+      "grad_norm": 300.21567029382476,
+      "learning_rate": 1.8109339407744877e-06,
+      "loss": 25.5947,
+      "step": 330,
+      "true_loss": 3.3142
+    },
+    {
+      "epoch": 5.409836065573771,
+      "eval_accuracy": 0.14285714285714285,
+      "eval_loss": 3.8151462078094482,
+      "eval_runtime": 8.197,
+      "eval_samples_per_second": 52.946,
+      "eval_steps_per_second": 6.71,
+      "step": 330
+    },
+    {
+      "epoch": 5.491803278688525,
+      "grad_norm": 355.0409546655278,
+      "learning_rate": 1.753986332574032e-06,
+      "loss": 25.8608,
+      "step": 335,
+      "true_loss": 3.0482
+    },
+    {
+      "epoch": 5.573770491803279,
+      "grad_norm": 328.72773533463106,
+      "learning_rate": 1.6970387243735763e-06,
+      "loss": 25.1268,
+      "step": 340,
+      "true_loss": 3.103
+    },
+    {
+      "epoch": 5.573770491803279,
+      "eval_accuracy": 0.1336405529953917,
+      "eval_loss": 3.811136484146118,
+      "eval_runtime": 8.3155,
+      "eval_samples_per_second": 52.192,
+      "eval_steps_per_second": 6.614,
+      "step": 340
+    },
+    {
+      "epoch": 5.655737704918033,
+      "grad_norm": 335.47496045434843,
+      "learning_rate": 1.640091116173121e-06,
+      "loss": 25.9661,
+      "step": 345,
+      "true_loss": 3.5002
+    },
+    {
+      "epoch": 5.737704918032787,
+      "grad_norm": 318.3200333649924,
+      "learning_rate": 1.5831435079726654e-06,
+      "loss": 24.7426,
+      "step": 350,
+      "true_loss": 3.4293
+    },
+    {
+      "epoch": 5.737704918032787,
+      "eval_accuracy": 0.1336405529953917,
+      "eval_loss": 3.7999937534332275,
+      "eval_runtime": 8.9261,
+      "eval_samples_per_second": 48.621,
+      "eval_steps_per_second": 6.162,
+      "step": 350
+    },
+    {
+      "epoch": 5.8196721311475414,
+      "grad_norm": 315.7081677451309,
+      "learning_rate": 1.5261958997722096e-06,
+      "loss": 25.5119,
+      "step": 355,
+      "true_loss": 3.0279
+    },
+    {
+      "epoch": 5.901639344262295,
+      "grad_norm": 313.72316359405403,
+      "learning_rate": 1.469248291571754e-06,
+      "loss": 24.6192,
+      "step": 360,
+      "true_loss": 3.0785
+    },
+    {
+      "epoch": 5.901639344262295,
+      "eval_accuracy": 0.14055299539170507,
+      "eval_loss": 3.7833380699157715,
+      "eval_runtime": 8.9156,
+      "eval_samples_per_second": 48.679,
+      "eval_steps_per_second": 6.169,
+      "step": 360
+    },
+    {
+      "epoch": 5.983606557377049,
+      "grad_norm": 337.09696857571805,
+      "learning_rate": 1.4123006833712984e-06,
+      "loss": 24.576,
+      "step": 365,
+      "true_loss": 3.0792
+    },
+    {
+      "epoch": 6.065573770491803,
+      "grad_norm": 312.44171671989244,
+      "learning_rate": 1.355353075170843e-06,
+      "loss": 23.0454,
+      "step": 370,
+      "true_loss": 2.8838
+    },
+    {
+      "epoch": 6.065573770491803,
+      "eval_accuracy": 0.14285714285714285,
+      "eval_loss": 3.78341007232666,
+      "eval_runtime": 8.7024,
+      "eval_samples_per_second": 49.871,
+      "eval_steps_per_second": 6.32,
+      "step": 370
+    },
+    {
+      "epoch": 6.147540983606557,
+      "grad_norm": 312.0953900031812,
+      "learning_rate": 1.2984054669703875e-06,
+      "loss": 22.3877,
+      "step": 375,
+      "true_loss": 2.4284
+    },
+    {
+      "epoch": 6.229508196721311,
+      "grad_norm": 335.94846858957845,
+      "learning_rate": 1.2414578587699317e-06,
+      "loss": 23.5471,
+      "step": 380,
+      "true_loss": 3.0217
+    },
+    {
+      "epoch": 6.229508196721311,
+      "eval_accuracy": 0.1382488479262673,
+      "eval_loss": 3.793076753616333,
+      "eval_runtime": 8.1781,
+      "eval_samples_per_second": 53.068,
+      "eval_steps_per_second": 6.725,
+      "step": 380
+    },
+    {
+      "epoch": 6.311475409836065,
+      "grad_norm": 341.87803857824105,
+      "learning_rate": 1.1845102505694763e-06,
+      "loss": 23.2389,
+      "step": 385,
+      "true_loss": 2.6483
+    },
+    {
+      "epoch": 6.39344262295082,
+      "grad_norm": 369.08934119445,
+      "learning_rate": 1.1275626423690205e-06,
+      "loss": 23.6875,
+      "step": 390,
+      "true_loss": 2.932
+    },
+    {
+      "epoch": 6.39344262295082,
+      "eval_accuracy": 0.14055299539170507,
+      "eval_loss": 3.795403480529785,
+      "eval_runtime": 8.1757,
+      "eval_samples_per_second": 53.084,
+      "eval_steps_per_second": 6.727,
+      "step": 390
+    },
+    {
+      "epoch": 6.475409836065574,
+      "grad_norm": 369.5208687625068,
+      "learning_rate": 1.0706150341685651e-06,
+      "loss": 23.193,
+      "step": 395,
+      "true_loss": 2.8789
+    },
+    {
+      "epoch": 6.557377049180328,
+      "grad_norm": 376.6391661100538,
+      "learning_rate": 1.0136674259681093e-06,
+      "loss": 23.5847,
+      "step": 400,
+      "true_loss": 2.9588
+    },
+    {
+      "epoch": 6.557377049180328,
+      "eval_accuracy": 0.14516129032258066,
+      "eval_loss": 3.8012988567352295,
+      "eval_runtime": 8.302,
+      "eval_samples_per_second": 52.276,
+      "eval_steps_per_second": 6.625,
+      "step": 400
+    },
+    {
+      "epoch": 6.639344262295082,
+      "grad_norm": 335.2314208261569,
+      "learning_rate": 9.567198177676538e-07,
+      "loss": 22.4865,
+      "step": 405,
+      "true_loss": 3.1291
+    },
+    {
+      "epoch": 6.721311475409836,
+      "grad_norm": 383.2496458341536,
+      "learning_rate": 8.997722095671982e-07,
+      "loss": 24.1513,
+      "step": 410,
+      "true_loss": 2.858
+    },
+    {
+      "epoch": 6.721311475409836,
+      "eval_accuracy": 0.14746543778801843,
+      "eval_loss": 3.801600217819214,
+      "eval_runtime": 8.4227,
+      "eval_samples_per_second": 51.528,
+      "eval_steps_per_second": 6.53,
+      "step": 410
+    }
+  ],
+  "logging_steps": 5,
+  "max_steps": 488,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 8,
+  "save_steps": 10,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 0.0,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:639ecd139094c2a0c834dd7bd8e1bcc25336952f6cfc37d956b51b3443e52c51
+size 7249

zero_to_fp32.py

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