Converted QwQ notebooks to .py files

tools/train/train/qwq.py

@@ -6,7 +6,6 @@ Fine-tuning for QwQ model
 # Unsloth should be imported before transformers to ensure all optimizations are applied.
 from unsloth import FastLanguageModel, is_bfloat16_supported
 
-from .dataset import Dataset
 from dataclasses import dataclass
 from pathlib import Path
 from transformers import TrainingArguments
@@ -15,6 +14,8 @@ from typing import Optional, List
 import argparse
 import os
 
+from .dataset import Dataset
+
 @dataclass
 class Args:
     def __init__(self, args: Optional[List[str]]):
@@ -78,7 +79,7 @@ def main():
         load_in_4bit = True, # False for LoRA 16bit
         fast_inference = True, # Enable vLLM fast inference
         max_lora_rank = lora_rank,
-        gpu_memory_utilization = 0.85, # Reduce if out of memory
+        gpu_memory_utilization = 0.5, # Reduce if out of memory
     )
 
     model = FastLanguageModel.get_peft_model(
@@ -97,12 +98,6 @@ def main():
         loftq_config = None, # And LoftQ
     )
 
-    response_template = tokenizer.apply_chat_template(
-        [{"role": "assistant", "content": ""}],
-        tokenize=False,
-        add_generation_prompt=True
-    )
-
     training_args = TrainingArguments(
         output_dir=str(args.output_dir),
         num_train_epochs=3,
@@ -129,10 +124,6 @@ def main():
         train_dataset=dataset.to_transformers_dataset(tokenizer),
         dataset_text_field="messages",
         max_seq_length=max_seq_length,
-        data_collator=DataCollatorForCompletionOnlyLM(
-            response_template=response_template, 
-            tokenizer=tokenizer
-        ),
     )
 
     trainer.train()
@@ -140,7 +131,6 @@ def main():
     model.save_pretrained_merged(
         str(args.output_dir), 
         tokenizer=tokenizer,
-        save_method="merged_16bit"
     )
 
 if __name__ == "__main__":

tools/train/train/qwq_train.ipynb

@@ -22,7 +22,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [