FineTurnX/examples/quantization_workflow.py

"""
完整的量化工作流程示例

这个脚本演示了从微调到量化的完整流程。

使用方法:
    python examples/quantization_workflow.py
"""

import os
import sys
import json

sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))

from finetunex.quantization import (
    quantize_model,
    get_model_size,
    estimate_quantized_size,
    compare_models,
)


def print_section(title):
    """打印章节标题"""
    print("\n" + "=" * 70)
    print(f"  {title}")
    print("=" * 70 + "\n")


def main():
    print_section("模型量化完整工作流程")
    
    # 配置
    finetuned_model = "./outputs/qwen3.5-0.8b-finetuned"
    quantized_output = "./outputs/qwen3.5-0.8b-quantized"
    
    print("配置信息:")
    print(f"  微调模型路径：{finetuned_model}")
    print(f"  量化输出路径：{quantized_output}")
    
    # 步骤 1: 检查模型
    print_section("步骤 1: 检查微调模型")
    
    if not os.path.exists(finetuned_model):
        print(f"错误：找不到微调模型：{finetuned_model}")
        print("请先运行微调脚本：python examples/qwen3.5_0.8b_local_finetune.py")
        sys.exit(1)
    
    print(f"✓ 找到微调模型：{finetuned_model}")
    
    # 步骤 2: 查看原始模型大小
    print_section("步骤 2: 查看原始模型大小")
    
    original_size = get_model_size(finetuned_model)
    print(f"原始模型大小:")
    print(f"  总大小：{original_size['total_size_formatted']}")
    print(f"  文件数：{original_size['file_count']}")
    
    # 步骤 3: 估算量化后大小
    print_section("步骤 3: 估算量化后大小")
    
    print("不同量化级别的估算:")
    for bits in [4, 8]:
        estimate = estimate_quantized_size(finetuned_model, quantization_bits=bits)
        print(f"\n{bits}bit 量化:")
        print(f"  原始大小：{estimate['original_size']}")
        print(f"  估算大小：{estimate['estimated_size']}")
        print(f"  压缩比：{estimate['compression_ratio']}")
        print(f"  节省空间：{estimate['space_saved']} ({estimate['space_saved_percent']})")
    
    # 步骤 4: 选择量化方法
    print_section("步骤 4: 选择量化方法")
    
    print("可用的量化方法:")
    print("  1. AWQ  - 快速，精度高，适合 GPU 推理")
    print("  2. GPTQ - 精度高，适合 GPU 推理")
    print("  3. GGUF - 支持 CPU 推理")
    
    method = input("\n请选择量化方法 (awq/gptq/gguf)，默认 awq: ").strip().lower()
    if not method:
        method = "awq"
    
    if method not in ["awq", "gptq", "gguf"]:
        print(f"错误：不支持的量化方法：{method}")
        sys.exit(1)
    
    print(f"✓ 选择量化方法：{method}")
    
    # 步骤 5: 执行量化
    print_section("步骤 5: 执行量化")
    
    output_path = os.path.join(quantized_output, method)
    
    print(f"开始量化...")
    print(f"  方法：{method}")
    print(f"  输出：{output_path}")
    
    try:
        # 执行量化
        result = quantize_model(
            model_path=finetuned_model,
            output_path=output_path,
            method=method,
            bits=4,
            group_size=128,
        )
        
        if result["success"]:
            print(f"\n✓ 量化成功！")
            print(f"  输出路径：{output_path}")
        else:
            print(f"\n✗ 量化失败！")
            sys.exit(1)
            
    except Exception as e:
        print(f"\n✗ 量化过程出错：{e}")
        import traceback
        traceback.print_exc()
        sys.exit(1)
    
    # 步骤 6: 比较模型大小
    print_section("步骤 6: 比较模型大小")
    
    comparison = compare_models(
        finetuned_model,
        output_path,
        label_1="原始模型",
        label_2=f"{method.upper()} 量化模型",
    )
    
    print(f"{comparison['原始模型']['size']} -> {comparison[f'{method.upper()} 量化模型']['size']}")
    print(f"减少了：{comparison['difference']} ({comparison['difference_percent']})")
    print(f"更小的模型：{comparison['smaller']}")
    
    # 步骤 7: 使用建议
    print_section("步骤 7: 使用建议")
    
    if method == "awq":
        print("""
AWQ 量化模型使用示例:

from transformers import AutoTokenizer
from awq import AutoAWQForCausalLM

# 加载模型
model = AutoAWQForCausalLM.from_quantized(
    "{output_path}",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("{output_path}")

# 推理
prompt = "你好"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))
        """.format(output_path=output_path))
        
    elif method == "gptq":
        print(f"""
GPTQ 量化模型使用示例:

from auto_gptq import AutoGPTQForCausalLM
from transformers import AutoTokenizer

# 加载模型
model = AutoGPTQForCausalLM.from_quantized(
    "{output_path}",
    device="cuda:0",
)
tokenizer = AutoTokenizer.from_pretrained("{output_path}")

# 推理
prompt = "你好"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))
        """)
        
    elif method == "gguf":
        print(f"""
GGUF 量化模型使用示例:

# 命令行推理
./llama.cpp/main -m {output_path}/*.gguf -p "你好" -n 512

# Python 推理
from llama_cpp import Llama

llm = Llama(model_path="{output_path}/*.gguf")
output = llm("你好", max_tokens=100)
print(output)
        """)
    
    # 完成
    print_section("工作流程完成")
    
    print("✓ 所有步骤完成！")
    print(f"\n量化模型已保存到：{output_path}")
    print("\n下一步:")
    print("  1. 测试量化模型的性能")
    print("  2. 比较量化前后的推理速度")
    print("  3. 评估量化对模型精度的影响")
    print("  4. 部署量化模型到生产环境")
    
    print("\n" + "=" * 70)


if __name__ == "__main__":
    main()