train/test_ood_python_lora_rope.py

import lighteval
from lighteval.logging.evaluation_tracker import EvaluationTracker
from lighteval.models.vllm.vllm_model import VLLMModelConfig
from lighteval.pipeline import ParallelismManager, Pipeline, PipelineParameters
from lighteval.utils.imports import is_package_available
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
import os
import torch
import json

if is_package_available("accelerate"):
    from datetime import timedelta
    from accelerate import Accelerator, InitProcessGroupKwargs
    accelerator = Accelerator(kwargs_handlers=[InitProcessGroupKwargs(timeout=timedelta(seconds=3000))])
else:
    accelerator = None

def merge_lora_if_needed():
    """Merge LoRA model and preserve RoPE scaling configuration"""
    merged_path = "/public/home/lshi/yoAI/projects/Online_CL/train/model_sft_save/Qwen2.5-Math-1.5B-Lora-Merged"
    
    if os.path.exists(os.path.join(merged_path, "config.json")):
        print(f"Merged model already exists at {merged_path}")
        
        # Verify RoPE scaling in existing merged model
        config_path = os.path.join(merged_path, "config.json")
        with open(config_path, 'r') as f:
            config = json.load(f)
            if 'rope_scaling' in config:
                print(f"✓ Existing merged model has RoPE scaling: {config['rope_scaling']}")
                print(f"✓ Max position embeddings: {config.get('max_position_embeddings', 'N/A')}")
            else:
                print("⚠ Warning: Existing merged model does NOT have RoPE scaling config!")
                print("  Deleting and re-creating with RoPE scaling...")
                import shutil
                shutil.rmtree(merged_path)
                return merge_lora_if_needed()  # Recursive call to re-create
        
        return merged_path

    print("="*100)
    print("Merged model not found. Starting merge process...")
    print("="*100)

    lora_path = "/public/home/lshi/yoAI/projects/Online_CL/train/model_sft_save/Qwen2.5-Math-1.5B-Lora"
    
    # Step 1: Load base model
    print("\n[1/5] Loading base model...")
    base_model = AutoModelForCausalLM.from_pretrained(
        "Qwen/Qwen2.5-Math-1.5B",
        torch_dtype=torch.bfloat16,
        trust_remote_code=True,
        device_map="auto"
    )
    
    # Step 2: Load LoRA adapter
    print("\n[2/5] Loading LoRA adapter...")
    model = PeftModel.from_pretrained(base_model, lora_path)
    
    # Step 3: Merge and unload
    print("\n[3/5] Merging LoRA weights with base model...")
    merged_model = model.merge_and_unload()
    
    # Step 4: Save merged model
    print(f"\n[4/5] Saving merged model to {merged_path}...")
    os.makedirs(merged_path, exist_ok=True)
    merged_model.save_pretrained(merged_path, safe_serialization=True)
    
    # Step 5: Add RoPE scaling configuration
    print("\n[5/5] Adding RoPE scaling configuration...")
    merged_config_path = os.path.join(merged_path, "config.json")
    with open(merged_config_path, 'r') as f:
        merged_config = json.load(f)
    
    # ========== RoPE scaling: 4096 -> 8192, factor = 2.0 ==========
    merged_config['rope_scaling'] = {
        "type": "linear",
        "factor": 2.0
    }
    
    print(f"✓ Added RoPE scaling: {merged_config['rope_scaling']}")
    
    # 更新 max_position_embeddings 从 4096 到 8192
    original_max_pos = merged_config.get('max_position_embeddings', 4096)
    scaling_factor = merged_config['rope_scaling']['factor']
    new_max_pos = int(original_max_pos * scaling_factor)
    merged_config['max_position_embeddings'] = new_max_pos
    print(f"✓ Updated max_position_embeddings: {original_max_pos} -> {new_max_pos}")
    
    # Save updated config
    with open(merged_config_path, 'w') as f:
        json.dump(merged_config, f, indent=2, ensure_ascii=False)
    
    # Save tokenizer
    print("Saving tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Math-1.5B", trust_remote_code=True)
    tokenizer.save_pretrained(merged_path)
    
    # Clean up memory
    del base_model
    del model
    del merged_model
    torch.cuda.empty_cache()
    
    print("\n" + "="*100)
    print("✓ Merge completed successfully!")
    print(f"✓ Merged model saved to: {merged_path}")
    print(f"✓ RoPE scaling config: {merged_config['rope_scaling']}")
    print(f"✓ Max position embeddings: {merged_config['max_position_embeddings']}")
    print("="*100 + "\n")
    
    return merged_path

def main():
    # Set CUDA device FIRST before any CUDA operations
    os.environ["CUDA_VISIBLE_DEVICES"] = "2"
    os.environ["VLLM_ALLOW_LONG_MAX_MODEL_LEN"] = "1"
    
    print("Checking for merged model...")
    merged_model_path = merge_lora_if_needed()
    
    # Detect number of GPUs
    num_gpus = torch.cuda.device_count()
    print(f"\n{'='*100}")
    print(f"Detected {num_gpus} GPU(s)")
    if num_gpus > 0:
        for i in range(num_gpus):
            print(f"  GPU {i}: {torch.cuda.get_device_name(i)}")
    print(f"{'='*100}\n")
    
    # Read the merged model config to get max_model_length
    config_path = os.path.join(merged_model_path, "config.json")
    with open(config_path, 'r') as f:
        model_config_dict = json.load(f)
        max_position_embeddings = model_config_dict.get('max_position_embeddings', 4096)
        rope_scaling = model_config_dict.get('rope_scaling', None)
    
    print(f"Model max_position_embeddings: {max_position_embeddings}")
    print(f"Model RoPE scaling config: {rope_scaling}")
    
    # 使用 8192 作为 max_model_length（你训练时扩展后的长度）
    max_model_length = 8192
    print(f"Using max_model_length: {max_model_length}\n")
    
    print("Setting up evaluation pipeline...")
    
    evaluation_tracker = EvaluationTracker(
        output_dir="./results",
        save_details=True,
        push_to_hub=False,
    )

    pipeline_params = PipelineParameters(
        launcher_type=ParallelismManager.ACCELERATE,
        custom_tasks_directory=None,
        max_samples=500
    )

    model_config = VLLMModelConfig(
        model_name=merged_model_path,
        dtype="bfloat16",
        max_model_length=max_model_length,  # 使用 8192
        trust_remote_code=True,
        tensor_parallel_size=num_gpus,
    )

    task = "lighteval|math_500|0"

    print(f"Using {num_gpus} GPU(s) with tensor parallelism")
    print(f"Task: {task}")
    print(f"Max model length: {max_model_length}\n")

    print("Creating pipeline...")
    pipeline = Pipeline(
        tasks=task,
        pipeline_parameters=pipeline_params,
        evaluation_tracker=evaluation_tracker,
        model_config=model_config,
    )

    # Fix generation_size
    print("Configuring generation parameters...")
    for task_name, task_obj in pipeline.tasks_dict.items():
        for doc in task_obj._docs:
            doc.generation_size = 2048

    print("\nStarting evaluation...")
    print("="*100)
    pipeline.evaluate()
    
    print("\nSaving results...")
    pipeline.save_and_push_results()
    
    print("\nShowing results...")
    pipeline.show_results()
    
    print("\n" + "="*100)
    print("✓ Evaluation completed!")
    print("="*100)

if __name__ == "__main__":
    main()