unsloth_deepseek/src/train_model_grpo_v0.py

import os
import torch
import torch.distributed as dist
from unsloth import FastLanguageModel
from unsloth import is_bfloat16_supported
from trl import SFTTrainer, GRPOConfig, GRPOTrainer
from datasets import load_dataset
from transformers import TrainingArguments

import re
from datasets import load_dataset, Dataset
from modelscope.msdatasets import MsDataset

# Load and prep dataset
SYSTEM_PROMPT = """
Respond in the following format:
<reasoning>
...
</reasoning>
<answer>
...
</answer>
"""

XML_COT_FORMAT = """\
<reasoning>
{reasoning}
</reasoning>
<answer>
{answer}
</answer>
"""

def extract_xml_answer(text: str) -> str:
    answer = text.split("<answer>")[-1]
    answer = answer.split("</answer>")[0]
    return answer.strip()

def extract_hash_answer(text: str) -> str | None:
    if "####" not in text:
        return None
    return text.split("####")[1].strip()

# uncomment middle messages for 1-shot prompting
def get_gsm8k_questions(split = "train") -> Dataset:
    # data = load_dataset('https://huggingface.co/datasets/openai/gsm8k', 'main')[split] # type: ignore
    
    data =  MsDataset.load('openai-mirror/gsm8k', subset_name='main', split=split)
    data = data.map(lambda x: { # type: ignore
        'prompt': [
            {'role': 'system', 'content': SYSTEM_PROMPT},
            {'role': 'user', 'content': x['question']}
        ],
        'answer': extract_hash_answer(x['answer'])
    }) # type: ignore
    return data # type: ignore

dataset = get_gsm8k_questions()

# 方法 1：使用 datasets 库的 to_json 方法
dataset.to_json(os.path.join("..","data","backup", "gsm8k_dataset_for_train.jsonl"), orient="records", lines=True, force_ascii=False)

# Reward functions
def correctness_reward_func(prompts, completions, answer, **kwargs) -> list[float]:
    responses = [completion[0]['content'] for completion in completions]
    q = prompts[0][-1]['content']
    extracted_responses = [extract_xml_answer(r) for r in responses]
    print('-'*20, f"Question:\n{q}", f"\nAnswer:\n{answer[0]}", f"\nResponse:\n{responses[0]}", f"\nExtracted:\n{extracted_responses[0]}")
    return [2.0 if r == a else 0.0 for r, a in zip(extracted_responses, answer)]

def int_reward_func(completions, **kwargs) -> list[float]:
    responses = [completion[0]['content'] for completion in completions]
    extracted_responses = [extract_xml_answer(r) for r in responses]
    return [0.5 if r.isdigit() else 0.0 for r in extracted_responses]

def strict_format_reward_func(completions, **kwargs) -> list[float]:
    """Reward function that checks if the completion has a specific format."""
    pattern = r"^<reasoning>\n.*?\n</reasoning>\n<answer>\n.*?\n</answer>\n$"
    responses = [completion[0]["content"] for completion in completions]
    matches = [re.match(pattern, r) for r in responses]
    return [0.5 if match else 0.0 for match in matches]

def soft_format_reward_func(completions, **kwargs) -> list[float]:
    """Reward function that checks if the completion has a specific format."""
    pattern = r"<reasoning>.*?</reasoning>\s*<answer>.*?</answer>"
    responses = [completion[0]["content"] for completion in completions]
    matches = [re.match(pattern, r) for r in responses]
    return [0.5 if match else 0.0 for match in matches]

def count_xml(text) -> float:
    count = 0.0
    if text.count("<reasoning>\n") == 1:
        count += 0.125
    if text.count("\n</reasoning>\n") == 1:
        count += 0.125
    if text.count("\n<answer>\n") == 1:
        count += 0.125
        count -= len(text.split("\n</answer>\n")[-1])*0.001
    if text.count("\n</answer>") == 1:
        count += 0.125
        count -= (len(text.split("\n</answer>")[-1]) - 1)*0.001
    return count

def xmlcount_reward_func(completions, **kwargs) -> list[float]:
    contents = [completion[0]["content"] for completion in completions]
    return [count_xml(c) for c in contents]

class ModelTrainer:
    def __init__(self, model_name, max_seq_length, dtype, load_in_4bit,lora_rank=32):
        # 初始化 ModelTrainer 类，设置模型名称、最大序列长度、数据类型和是否以4位加载
        self.model_name = model_name
        self.max_seq_length = max_seq_length
        self.dtype = dtype         # dtype: 数据类型，如 torch.float16 或 torch.bfloat16
        self.load_in_4bit = load_in_4bit         # load_in_4bit: 是否以4位精度加载模型，用于节省显存
        self.lora_rank=lora_rank  #Larger rank = smarter, but slower

    def load_model(self,lora_rank=64):
        # 加载预训练模型和分词器
        model, tokenizer = FastLanguageModel.from_pretrained(
            model_name=self.model_name,
            max_seq_length=self.max_seq_length,
            load_in_4bit=self.load_in_4bit, # 值为True 以 4 bit量化进行微调，为False LoRA 16bit。这将内存使用量减少了 4 倍，使我们能够在免费的 16GB 内存 GPU 中实际进行微调。4 位量化本质上将权重转换为一组有限的数字以减少内存使用量。这样做的缺点是准确度会下降 1-2%。如果您想要这种微小的额外准确度，请在较大的 GPU（如 H100）上将其设置为 False。
            dtype=self.dtype,
            fast_inference = True, # Enable vLLM fast inference
            max_lora_rank = lora_rank,
            gpu_memory_utilization=0.005, # 0.6 # Reduce if out of memory
        )

        # 将模型移动到设备上
        model = model.to_empty(device='cuda')  # 使用 to_empty 而不是 to

        # 初始化模型的权重
        for param in model.parameters():
            if param.is_meta:
                param.data = torch.randn_like(param)  # 随机初始化

        # 添加 LoRA 适配器
        model = FastLanguageModel.get_peft_model(
            model,
            max_seq_length=self.max_seq_length,  # 最大上下文（序列）长度
            r=lora_rank,   # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
            target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                          "gate_proj", "up_proj", "down_proj"],  # 应用 LoRA 的目标模块
            lora_alpha=16,  # LoRA 的 alpha 参数，控制适配器的缩放
            lora_dropout=0,  # LoRA 的 dropout 率，设置为0以优化性能
            bias="none",     # 是否在 LoRA 中添加偏置，设置为 "none" 以优化性能
            use_gradient_checkpointing="unsloth",  # 使用梯度检查点以节省显存,对于非常长的上下文，使用 True 或 "unsloth"
            random_state=3407,  # 随机种子，确保实验可复现
            use_rslora=False,  # 是否使用 rank stabilized LoRA，当前不支持
            loftq_config=None,  # LoftQ 配置，当前不支持
        )

        return model, tokenizer

    def load_data(self, train_data_path):
        # 加载训练集和测试集
        with open(train_data_path, 'r') as f:
            train_dataset = load_dataset("json", data_files={"train": train_data_path}, split="train")

        # train_data_path: 训练数据路径，格式为 JSONL
        return train_dataset

    def train(self, model, tokenizer, train_dataset):
        print("is_bfloat16_supported()=", is_bfloat16_supported())
        
        # 监控显存使用情况
        print(f"Reserved memory: {torch.cuda.memory_reserved()}")
        print(f"Allocated memory: {torch.cuda.memory_allocated()}")

        # 启用 pin_memory  2025年3月10日未能验证通过
        train_loader = torch.utils.data.DataLoader(
            train_dataset, batch_size=1, shuffle=True, pin_memory=True  
        )
        
        # 释放未使用的显存
        torch.cuda.empty_cache()

        training_args = GRPOConfig(
            use_vllm = True, # use vLLM for fast inference!
            learning_rate = 5e-6,
            adam_beta1 = 0.9,
            adam_beta2 = 0.99,
            weight_decay = 0.1,
            warmup_ratio = 0.1,
            lr_scheduler_type = "cosine",
            optim ="adamw_8bit", # "adamw_8bit" if device == "cuda" else "adamw_torch",  # CPU 使用 adamw_torch
            logging_steps = 1,
            bf16 = is_bfloat16_supported(),
            fp16 = not is_bfloat16_supported(),
            per_device_train_batch_size = 1,
            gradient_accumulation_steps = 1, # Increase to 4 for smoother training
            num_generations = 4, # 8 # 每次生成 输出 个数
            max_prompt_length = 256, # 256 # 输入提示的最大长度
            max_completion_length = 200,# 200 # 生成内容的最大长度
            num_train_epochs = 1, # Set to 1 for a full training run
            max_steps = 250,  # 250
            save_steps = 250, # 250
            max_grad_norm = 0.1,
            report_to = "none", # Can use Weights & Biases
            output_dir = os.path.join('..', 'models',"outputs"),
        )

        # 初始化 SFTTrainer
        trainer = GRPOTrainer(
            model = model,
            processing_class = tokenizer,
            reward_funcs = [
                xmlcount_reward_func,
                soft_format_reward_func,
                strict_format_reward_func,
                int_reward_func,
                correctness_reward_func,
            ],
            args = training_args,
            train_dataset = train_dataset,
        )
        
        # 训练模型
        trainer.train()
        
        return model

    def save_model(self, model, tokenizer, save_path):
        # 保存模型和分词器
        model.save_pretrained(save_path)
        tokenizer.save_pretrained(save_path)
        print(f"Model saved to {save_path}")

if __name__ == "__main__":
    # 配置参数
    model_name = os.path.join('..', 'models', 'pretrained', 'DeepSeek-R1-Distill-Qwen-1.5B')
    # model_name: 预训练模型的路径
    max_seq_length = 512  # 单次会话（single session） 的最大 token 长度，一个token大约3-4 字节（Byte）
    dtype = torch.float16  # 数据类型
    load_in_4bit = True  # 是否以4位精度加载模型
    lora_rank=16 

    # 定义训练集和测试集路径
    train_data_path = os.path.join('..', 'data', 'processed', 'train.jsonl')
    # train_data_path: 训练数据路径

    try:
        # 设置环境变量
        # 单机多卡
        os.environ['RANK'] = '0' # 第一张卡的 rank
        os.environ['WORLD_SIZE'] = '1'  # 总共有 1 张卡
        os.environ['MASTER_ADDR'] = 'localhost'
        os.environ['MASTER_PORT'] = '12345'
        # 多机多卡
        # export RANK=0  # 第一台机器的 rank
        # export WORLD_SIZE=4  # 总共有 4 台机器
        # export MASTER_ADDR=<主节点 IP>
        # export MASTER_PORT=12345

        # 初始化进程组
        # dist.init_process_group(backend='nccl', init_method='env://')
        # 初始化 ModelTrainer
        trainer = ModelTrainer(model_name, max_seq_length, dtype, load_in_4bit,lora_rank)
        
        # 加载模型和分词器
        model, tokenizer = trainer.load_model(lora_rank)

        # 加载数据集
        train_dataset = trainer.load_data(train_data_path)

        # 训练模型
        model = trainer.train(model, tokenizer, train_dataset)

        # 保存模型
        save_path = os.path.join('..', 'models', 'trained', 'DeepSeek-R1-Distill-Qwen-1.5B-GRPO')
        trainer.save_model(model, tokenizer, save_path)
    finally:
        # 确保进程组被销毁
        if dist.is_initialized():
            dist.destroy_process_group()
        print("train finally")