unsloth_deepseek/src/train_model_grpo_v1.1.py

# train_model_grpo_v1.py

import os
import torch
import torch.distributed as dist
from unsloth import FastLanguageModel
from unsloth import is_bfloat16_supported
from trl import GRPOConfig, GRPOTrainer
from datasets import load_dataset
from conf_train import Config, load_config  # 导入配置文件
import re
# from transformers import BertTokenizer, BertModel  # 分词模型最大支持 512 个token
from transformers import LongformerTokenizer, LongformerModel # 分词模型最大支持 4096 个token
import numpy as np

class ModelTrainer:
    def __init__(self, config: Config):
        """
        初始化 ModelTrainer 类，加载配置参数。
        :param config: 配置对象，包含模型训练所需的参数
        """
        self.config: Config = config
        self.model_name = config.model_name
        self.max_seq_length = config.max_seq_length
        self.dtype = torch.float16 if config.dtype == "float16" else torch.bfloat16
        self.load_in_4bit = config.load_in_4bit
        self.fast_inference = config.fast_inference
        self.lora_rank = config.lora_rank
        self.gpu_memory_utilization = config.gpu_memory_utilization
        # 初始化 BERT 模型和分词器
        self.tokenizer = LongformerTokenizer.from_pretrained(f'../models/allenai/longformer-base-4096')
        self.longformer_model = LongformerModel.from_pretrained(f'../models/allenai/longformer-base-4096')

    def load_model(self):
        """
        加载预训练模型和分词器。
        :return: 返回加载的模型和分词器
        """
        model, tokenizer = FastLanguageModel.from_pretrained(
            model_name=self.model_name,
            max_seq_length=self.max_seq_length,
            load_in_4bit=self.load_in_4bit, # False for LoRA 16bit
            dtype=self.dtype,
            fast_inference=self.fast_inference,
            max_lora_rank=self.lora_rank,
            gpu_memory_utilization=self.gpu_memory_utilization,
        )

        model = model.to_empty(device='cuda')

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

        # 添加 LoRA 适配器
        model = FastLanguageModel.get_peft_model(
            model,
            max_seq_length=self.max_seq_length,
            r=self.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"],  # Remove QKVO if out of memory
            lora_alpha=16,
            lora_dropout=0, #Supports any, but = 0 is optimized
            bias="none", # Supports any, but = "none" is optimized
            #[NEW]"unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
            use_gradient_checkpointing="unsloth",  # True or "unsloth" for very long context
            random_state=3407,
            use_rslora=False,  # We support rank stabilized LoRA
            loftq_config=None,  # And LoftQ
        )

        return model, tokenizer

    def load_data(self, train_data_path):
        """
        加载训练数据集。
        :param train_data_path: 训练数据路径
        :return: 返回加载的训练数据集
        """
        with open(train_data_path, 'r') as f:
            train_dataset = load_dataset("json", data_files={"train": train_data_path}, split="train")
            print("train_dataset",train_dataset)
        return train_dataset

    def train(self, model, tokenizer, train_dataset):
        """
        训练模型。
        :param model: 预训练模型
        :param tokenizer: 分词器
        :param train_dataset: 训练数据集
        :return: 返回训练后的模型
        """
        print("is_bfloat16_supported()=", is_bfloat16_supported())
        print(f"Reserved memory: {torch.cuda.memory_reserved()}")
        print(f"Allocated memory: {torch.cuda.memory_allocated()}")

        train_loader = torch.utils.data.DataLoader(
            train_dataset, batch_size=1, shuffle=True, pin_memory=True
        )

        torch.cuda.empty_cache()
        print("self.config.learning_rate=", float(self.config.learning_rate))
        training_args = GRPOConfig(
            use_vllm=self.config.use_vllm,
            learning_rate=float(self.config.learning_rate),
            adam_beta1=self.config.adam_beta1,
            adam_beta2=self.config.adam_beta2,
            weight_decay=self.config.weight_decay,
            warmup_ratio=self.config.warmup_ratio,
            lr_scheduler_type=self.config.lr_scheduler_type,
            optim=self.config.optim,
            logging_steps=self.config.logging_steps,
            bf16=is_bfloat16_supported(),
            fp16=not is_bfloat16_supported(),
            per_device_train_batch_size=self.config.per_device_train_batch_size,
            gradient_accumulation_steps=self.config.gradient_accumulation_steps,
            num_generations=self.config.num_generations,
            max_prompt_length=self.config.max_prompt_length,
            max_completion_length=self.config.max_completion_length,
            num_train_epochs=self.config.num_train_epochs,
            max_steps=self.config.max_steps,
            save_steps=self.config.save_steps,
            max_grad_norm=self.config.max_grad_norm,
            report_to=self.config.report_to,
            output_dir=self.config.output_dir,
        )

        """
        PyTorch 的分布式进程组已初始化，但并行模式不等于 “分布式并行模式（ParallelMode.DISTRIBUTED）”。
        为了使用 PyTorch 的分布式数据并行（DDP），请使用 python -m torch.distributed.launch 来启动你的脚本。
        """

        trainer = GRPOTrainer(
            model=model,
            processing_class=tokenizer, # 用于处理输入文本的分词器(tokenizer)。它将文本转换为模型可以理解的数字格式。
            reward_funcs=[
                self.xmlcount_reward_func, # 某种特定的基于XML计数的奖励函数
                self.soft_format_reward_func, # 基于软格式的奖励函数。
                self.strict_format_reward_func, # 基于严格格式的奖励函数。
                self.int_reward_func,  # 整数奖励函数。
                self.correctness_reward_func, # 基于输出正确性的奖励函数
                self.semantic_correctness_reward_func, # 语义正确性奖励函数
                self.reasoning_quality_reward_func,  # 推理质量奖励函数
                self.combined_reward_func,  # combined_reward_func
            ],   # 这是一个奖励函数的列表，决定了模型输出的好坏。在GRPO训练中，奖励函数通常用来评估模型输出的质量。
            args=training_args, # 定义的训练超参数。
            train_dataset=train_dataset, # 训练数据集,
        )

        trainer.train()
        return model

    def save_model(self, model, tokenizer, save_path):
        """
        保存训练后的模型和分词器。
        :param model: 训练后的模型
        :param tokenizer: 分词器
        :param save_path: 保存路径
        """
        """
        # Save to 8bit Q8_0
        if False: model.save_pretrained_gguf("model", tokenizer,)
        # Remember to go to https://huggingface.co/settings/tokens for a token!
        # And change hf to your username!
        if False: model.push_to_hub_gguf("hf/model", tokenizer, token = "")

        # Save to 16bit GGUF
        if False: model.save_pretrained_gguf("model", tokenizer, quantization_method = "f16")
        if False: model.push_to_hub_gguf("hf/model", tokenizer, quantization_method = "f16", token = "")

        # Save to q4_k_m GGUF
        if False: model.save_pretrained_gguf("model", tokenizer, quantization_method = "q4_k_m")
        if False: model.push_to_hub_gguf("hf/model", tokenizer, quantization_method = "q4_k_m", token = "")

        ###

        model.save_pretrained_merged("model", tokenizer, save_method = "merged_16bit",) # save_method = "merged_4bit" Merge to 4bit ; save_method = "lora"  Just LoRA adapters ;
        """

        model.save_pretrained(save_path)
        tokenizer.save_pretrained(save_path)
        
        print(f"Model saved to {save_path}")
    
    @staticmethod
    def cosine_similarity(vec1, vec2):
        """
        计算两个向量的余弦相似度。
        :param vec1: 第一个向量，形状为 (1, 768)
        :param vec2: 第二个向量，形状为 (1, 768)
        :return: 余弦相似度
        """
        # 将 (1, 768) 的矩阵转换为 (768,) 的一维向量
        vec1 = vec1.squeeze()  # 形状从 (1, 768) 变为 (768,)
        vec2 = vec2.squeeze()  # 形状从 (1, 768) 变为 (768,)
        # print(f"vec1 shape: {vec1.shape}, vec2 shape: {vec2.shape}")
        # 计算余弦相似度
        return np.dot(vec1, vec2) / (np.linalg.norm(vec1) * np.linalg.norm(vec2))

    def semantic_correctness_reward_func(self, prompts, completions, answer, **kwargs):
        """
        使用 Longformer 计算生成答案与标准答案的语义相似度。
        """
        responses = [completion[0]['content'] for completion in completions]
        extracted_responses = [self.extract_xml_answer(r) for r in responses]
        scores = []
        for resp, ans in zip(extracted_responses, answer):
            # 截断文本，确保长度不超过 4096
            resp = self.tokenizer.decode(self.tokenizer.encode(resp, truncation=True, max_length=4096))
            ans = self.tokenizer.decode(self.tokenizer.encode(ans, truncation=True, max_length=4096))
            # 编码生成答案和标准答案
            inputs_resp = self.tokenizer(resp, return_tensors='pt', padding=True, truncation=True, max_length=4096)
            inputs_ans = self.tokenizer(ans, return_tensors='pt', padding=True, truncation=True, max_length=4096)
            with torch.no_grad():
                outputs_resp = self.longformer_model(**inputs_resp).last_hidden_state.mean(dim=1)  # 形状为 (1, 768)
                outputs_ans = self.longformer_model(**inputs_ans).last_hidden_state.mean(dim=1)  # 形状为 (1, 768)
            # 计算余弦相似度
            similarity = self.cosine_similarity(outputs_resp.numpy(), outputs_ans.numpy())
            scores.append(similarity)
        return scores
    
    def combined_reward_func(self, prompts, completions, answer, **kwargs):
        """
        综合多个奖励函数，动态调整权重。
        :param prompts: 输入提示
        :param completions: 模型生成的补全内容
        :param answer: 标准答案
        :return: 综合得分列表
        """
        # 计算各奖励函数的得分
        format_score = self.strict_format_reward_func(completions)
        semantic_score = self.semantic_correctness_reward_func(prompts, completions, answer)
        correctness_score = self.correctness_reward_func(prompts, completions, answer)

        # 动态调整权重
        combined_scores = []
        for fs, ss, cs in zip(format_score, semantic_score, correctness_score):
            if cs == 2.0:  # 答案完全正确
                combined_scores.append(fs * 0.2 + ss * 0.3 + cs * 0.5)
            else:  # 答案不完全正确
                combined_scores.append(fs * 0.4 + ss * 0.4 + cs * 0.2)
        return combined_scores
    
    @staticmethod
    def reasoning_quality_reward_func(completions, **kwargs):
        """
        检查推理过程的质量。
        :param completions: 模型生成的补全内容
        :return: 推理过程质量得分列表
        """
        responses = [completion[0]["content"] for completion in completions]
        scores = []
        for response in responses:
            reasoning_match = re.search(r"<reasoning>\n(.+?)\n</reasoning>", response, re.DOTALL)
            if reasoning_match:
                reasoning_content = reasoning_match.group(1).strip()
                # 简单检查推理内容是否包含关键词
                if "诊断" in reasoning_content and "原因" in reasoning_content:
                    scores.append(1.0)
                else:
                    scores.append(0.5)
            else:
                scores.append(0.0)
        return scores

    @staticmethod
    def extract_xml_answer(text: str) -> str:
        """
        从文本中提取 XML 格式的答案。
        :param text: 包含 XML 格式的文本
        :return: 提取的答案
        """
        answer = text.split("<answer>")[-1]
        answer = answer.split("</answer>")[0]
        return answer.strip()

    @staticmethod
    def count_xml(text) -> float:
        """
        计算 XML 标签的数量和完整性。
        :param text: 包含 XML 格式的文本
        :return: XML 标签的完整性得分
        """
        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

    @staticmethod
    def xmlcount_reward_func(completions, **kwargs):
        """
        计算 XML 标签的完整性得分。
        :param completions: 模型生成的补全内容
        :return: XML 标签的完整性得分列表
        """
        contents = [completion[0]["content"] for completion in completions]
        return [ModelTrainer.count_xml(c) for c in contents]

    @staticmethod
    def soft_format_reward_func(completions, **kwargs):
        """
        检查补全内容是否符合软格式要求。
        :param completions: 模型生成的补全内容
        :return: 符合软格式要求的得分列表
        """
        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]

    @staticmethod
    def strict_format_reward_func(completions, **kwargs):
        """
        检查响应是否符合严格的 XML 格式要求，并确保标签内容非空。
        :param completions: 模型生成的补全内容
        :return: 符合严格格式要求的得分列表
        """
        pattern = r"^<reasoning>\n(.+?)\n</reasoning>\n<answer>\n(.+?)\n</answer>\n$"
        responses = [completion[0]["content"] for completion in completions]
        scores = []
        for response in responses:
            match = re.match(pattern, response, re.DOTALL)
            if match:
                reasoning_content = match.group(1).strip()
                answer_content = match.group(2).strip()
                # 检查内容是否非空
                if reasoning_content and answer_content:
                    scores.append(1.0)  # 格式和内容均符合要求
                else:
                    scores.append(0.5)  # 格式符合但内容为空
            else:
                scores.append(0.0)  # 格式不符合
        return scores

    @staticmethod
    def int_reward_func(completions, **kwargs):
        """
        检查补全内容是否包含整数。
        :param completions: 模型生成的补全内容
        :return: 包含整数的得分列表
        """
        responses = [completion[0]['content'] for completion in completions]
        extracted_responses = [ModelTrainer.extract_xml_answer(r) for r in responses]
        return [0.5 if r.isdigit() else 0.0 for r in extracted_responses]

    @staticmethod
    def correctness_reward_func(prompts, completions, answer, **kwargs):
        """
        检查补全内容是否正确。
        :param prompts: 输入提示
        :param completions: 模型生成的补全内容
        :param answer: 正确答案
        :return: 补全内容正确的得分列表
        """
        print("completions : \n ",completions)
        responses = [completion[0]['content'] for completion in completions]
        q = prompts[0][-1]['content']
        extracted_responses = [ModelTrainer.extract_xml_answer(r) for r in responses]
        # print(f"\n Response:\n {responses}",f"\n Extracted:\n {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)]

if __name__ == "__main__":
    try:
        # 加载配置文件
        config = load_config(f"../conf/conf_train.yaml")
        print("train config:\n",config)

        # 设置环境变量
        """
            # 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

            # 单机多卡
            os.environ['RANK'] = '0'
            os.environ['WORLD_SIZE'] = '1'
            os.environ['MASTER_ADDR'] = 'localhost'
            os.environ['MASTER_PORT'] = '12345'

            # 根据操作系统选择后端
            backend = 'gloo' if os.name == 'nt' else 'nccl'
            # 使用文件初始化方法  2025-3-11 成功验证支持windows
            init_method = f'env://'  # env://  # 文件路径需要所有进程都能访问
            dist.init_process_group(backend=backend, init_method=init_method)
            print(f"Initialized distributed training with backend: {backend}")
        """

        # 初始化 ModelTrainer
        trainer = ModelTrainer(config)

        # 加载模型和分词器
        model, tokenizer = trainer.load_model()

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

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

        # 保存模型
        trainer.save_model(model, tokenizer, config.save_path)
        
        print("Training completed.")
    except Exception as e:
        print("exception \n ",e)
    finally:
        # # 确保进程组被销毁
        # if dist.is_initialized():
        #     dist.destroy_process_group()
        print("end")