ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">一、配置环境ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">1. 安装 unslothpipinstallunsloth-ihttps://pypi.tuna.tsinghua.edu.cn/simple ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">2. 查看 GPU 信息import torch
def print_cuda_info(): try: print("-" * 40) print(" yTorch CUDA Environment Information:") print("-" * 40) if torch.cuda.is_available(): device_count = torch.cuda.device_count() print(f"Number of CUDA devices: {device_count}") if device_count > 0: device_name = torch.cuda.get_device_name(0) print(f"0th CUDA Device Name: {device_name}") total_memory = torch.cuda.get_device_properties(0).total_memory allocated_memory = torch.cuda.memory_allocated(0) free_memory = total_memory - allocated_memory print(f"Total Memory: {total_memory / (1024 ** 3):.2f} GB") print(f"Allocated Memory: {allocated_memory / (1024 ** 3):.2f} GB") print(f"Free Memory: {free_memory / (1024 ** 3):.2f} GB") else: print("No CUDA devices found.") else: print("CUDA is not available.") print("-" * 40) except Exception as e: print("-" * 40) print(f"An error occurred: {e}") print("-" * 40)
if __name__ == "__main__": print_cuda_info()
ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">3. 安装 Ollamacurl-fsSLhttps://ollama.com/install.sh|sh ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">4. 启动 ollama 服务ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">5. 下载 Qwen3-4B 镜像ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">（下载完成后需结束当前运行进程）ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">二、构建数据集ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">1. 加载模型和数据集（验证可用性）from unsloth import FastLanguageModelfrom datasets import load_datasetimport torch
# 配置max_seq_length = 2048load_in_4bit = True # 4bit量化
# 从 Hugging Face Hub 加载模型和分词器model, tokenizer = FastLanguageModel.from_pretrained( "Qwen/Qwen3-4B-Instruct-2507", # 模型名称 max_seq_length=max_seq_length, load_in_4bit=load_in_4bit, trust_remote_code=True # Qwen模型需此参数)
# 配置 LoRA 适配器model = FastLanguageModel.get_peft_model( model, r=16, target_modules=["q_proj","k_proj","v_proj","o_proj", "gate_proj","up_proj","down_proj",], lora_alpha=16, lora_dropout=0, bias="none", use_gradient_checkpointing="unsloth", random_state=3407, use_rslora=False, loftq_config=None,)
EOS_TOKEN = tokenizer.eos_token
# 加载数据集（需修改为你的数据集路径）dataset = load_dataset("json", data_files="/content/noli.json", split="train")
# 查看数据集信息print("数据集样例:", dataset[0])print("数据集大小:", len(dataset))
三、模型微调1. 完整微调代码# 导入必要的库from unsloth import FastLanguageModelfrom datasets import load_dataset, Datasetfrom trl import SFTTrainerfrom transformers import TrainingArgumentsfrom unsloth import is_bfloat16_supportedimport torch
# --- 1. 模型和分词器加载 ---print("正在加载模型和分词器...")model, tokenizer = FastLanguageModel.from_pretrained( model_name="Qwen/Qwen3-4B-Instruct-2507", max_seq_length=2048, load_in_4bit=True, trust_remote_code=True # Qwen 模型需要此参数)print("模型和分词器加载完成。")
# --- 2. LoRA 配置 ---print("正在配置LoRA适配器...")model = FastLanguageModel.get_peft_model( model, r=16, target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], lora_alpha=16, lora_dropout=0, bias="none", use_gradient_checkpointing="unsloth", random_state=3407, use_rslora=False, loftq_config=None,)print("LoRA适配器配置完成。")
# --- 3. 数据集加载 ---print("正在加载数据集...")raw_dataset = load_dataset("json", data_files="/content/NOLI.json", split="train") # 修改为你的数据集路径print(f"原始数据集加载完成。数据集大小: {len(raw_dataset)}")print("原始数据集样例:", raw_dataset[0])
# --- 4. 预处理数据集：添加 'text' 列 ---def create_text_column(example): """将单个样本格式化为模型训练所需的文本格式。""" # 安全地获取字段，确保是字符串 instruction = str(example.get("instruction", "")).strip() input_text = str(example.get("input", "")).strip() output_text = str(example.get("output", "")).strip()
# 构建用户部分 if input_text: user_content = f"{instruction}\n{input_text}" else: user_content = instruction
# 构建完整的提示（符合Qwen3对话格式） full_prompt = ( f"<|im_start|>user\n{user_content}<|im_end|>\n" f"<|im_start|>assistant\n{output_text}<|im_end|>" ) return {"text": full_prompt}
print("正在预处理数据集，添加 'text' 列...")# 使用 map 函数为数据集中的每个样本添加 'text' 列dataset = raw_dataset.map(create_text_column)print("数据集预处理完成。")print("处理后数据集样例:", dataset[0])
# --- 5. 配置并创建 SFTTrainer ---print("正在配置SFTTrainer...")trainer = SFTTrainer( model=model, tokenizer=tokenizer, train_dataset=dataset, # 使用预处理后的数据集 dataset_text_field="text", # 指定使用 'text' 列 max_seq_length=2048, dataset_num_proc=2, packing=False, # 此格式下禁用packing args=TrainingArguments( per_device_train_batch_size=2, gradient_accumulation_steps=4, warmup_steps=5, max_steps=100, # 可根据需求调整训练步数 learning_rate=2e-4, fp16=not is_bfloat16_supported(), bf16=is_bfloat16_supported(), logging_steps=5, optim="adamw_8bit", output_dir="./qwen_finetune_output_v2", overwrite_output_dir=True, report_to="none", # 禁用外部日志记录 seed=3407, ),)print("SFTTrainer配置完成。")
# --- 6. 开始训练 ---print("开始训练...")trainer.train()print("训练完成。")
# --- 7. 保存 LoRA 权重 ---print("正在保存LoRA适配器权重...")model.save_pretrained("./lora_adapters_v2")tokenizer.save_pretrained("./lora_adapters_v2") # 同时保存分词器配置print("LoRA适配器已保存到 './lora_adapters_v2' 目录。")
2. 合并 LoRA 权重（保存为 32 位和 16 位模型）# 导入必要的库from unsloth import FastLanguageModelimport torchfrom peft import PeftModel # 用于加载和合并LoRA权重
# --- 1. 加载基础模型（非4bit量化，用于完整权重合并）---print("正在加载基础模型...")model, tokenizer = FastLanguageModel.from_pretrained( model_name="Qwen/Qwen3-4B-Instruct-2507", max_seq_length=2048, load_in_4bit=False, # 加载完整精度模型 trust_remote_code=True)print("基础模型加载完成。")
# --- 2. 加载 LoRA 适配器 ---print("正在加载LoRA适配器...")model = PeftModel.from_pretrained( model=model, model_id="/content/lora_adapters_v2" # LoRA权重保存路径)print("LoRA适配器加载完成。")
# --- 3. 合并 LoRA 权重到基础模型 ---print("正在合并LoRA权重到基础模型...")model = model.merge_and_unload() # 执行权重合并print("权重合并完成。")
# --- 4. 保存完整模型（32位和16位）---print("正在保存完整模型...")# 保存为32位完整模型（高精度，体积较大）model.save_pretrained("./qwen_merged_full_model")tokenizer.save_pretrained("./qwen_merged_full_model")
# 保存为16位模型（平衡精度与体积）model.save_pretrained("./qwen_merged_full_model_16bit", torch_dtype=torch.float16)tokenizer.save_pretrained("./qwen_merged_full_model_16bit")
print("完整模型保存完成！")print("32位模型保存路径：./qwen_merged_full_model")print("16位模型保存路径：./qwen_merged_full_model_16bit")
四、使用微调后的 16 位模型from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamerimport torchfrom typing import List, Tuplefrom threading import Thread
# --- 1. 加载模型和分词器 ---model_path = "./qwen_merged_full_model_16bit" # 16位模型路径print(f"正在加载模型: {model_path}...")tokenizer = AutoTokenizer.from_pretrained( model_path, trust_remote_code=True, padding_side="left")# 确保pad_token存在（使用eos_token作为pad_token）if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained( model_path, torch_dtype=torch.float16, device_map="auto", # 自动分配设备（优先GPU） trust_remote_code=True)model.eval() # 推理模式print("模型加载完成，可开始对话。")
# --- 2. 自定义流式输出器（仅打印新生成内容）---class CurrentResponseStreamer(TextStreamer): def __init__(self, tokenizer, input_prompt_length: int, skip_prompt: bool = True, **decode_kwargs): super().__init__(tokenizer, skip_prompt=skip_prompt, **decode_kwargs) self.input_prompt_length = input_prompt_length self.first_token = True
def on_finalized_text(self, text: str, stream_end: bool = False): if self.first_token: print("Noli: ", end="", flush=True) self.first_token = False print(text, end="", flush=True) if stream_end: print() # 结束时换行
# --- 3. 流式生成函数 ---def generate_response_streaming(conversation_history: List[Tuple[str, str]]): """根据对话历史生成流式响应，并返回完整响应文本""" # 构建完整对话prompt prompt = "" for role, content in conversation_history: if role == "user": prompt += f"<|im_start|>user\n{content}<|im_end|>\n" else: prompt += f"<|im_start|>assistant\n{content}<|im_end|>\n" prompt += f"<|im_start|>assistant\n" # 启动助手回复生成
# 计算输入prompt的token长度 input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"] input_prompt_length = input_ids.shape[1]
# 移动输入到模型设备 inputs = { "input_ids": input_ids.to(model.device), "attention_mask": tokenizer(prompt, return_tensors="pt")["attention_mask"].to(model.device), }
# 初始化流式输出器 streamer = CurrentResponseStreamer( tokenizer, input_prompt_length=input_prompt_length, skip_special_tokens=True )
# 启动流式生成（独立线程） generation_kwargs = dict( **inputs, max_new_tokens=512, temperature=0.7, top_p=0.9, repetition_penalty=1.1, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id, do_sample=True, streamer=streamer ) thread = Thread(target=model.generate, kwargs=generation_kwargs) thread.start() thread.join() # 等待生成完成
# 生成完整响应文本（用于更新对话历史） with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=512, temperature=0.7, top_p=0.9, repetition_penalty=1.1, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id, do_sample=True ) # 仅解码新生成的部分 generated_ids = outputs[:, inputs['input_ids'].shape[1]:] full_response = tokenizer.decode(generated_ids[0], skip_special_tokens=True) return full_response
# --- 4. 对话主程序 ---if __name__ == "__main__": conversation_history: List[Tuple[str, str]] = [] print("你好！我是诺丽，一个AI助手。输入'退出'即可结束对话。")
while True: try: user_input = input("\n你: ").strip() if not user_input: continue if user_input.lower() in ["退出", "quit", "exit"]: print("Noli: 再见！很高兴与你交谈。") break
# 更新对话历史 conversation_history.append(("user", user_input)) # 流式生成响应 current_response = generate_response_streaming(conversation_history) # 保存完整响应到历史 conversation_history.append(("assistant", current_response))
except KeyboardInterrupt: print("\n\nNoli: 看起来你中断了对话。再见！") break except Exception as e: print(f"\nNoli: 抱歉，处理你的请求时出现了错误: {e}")
五、常见问题解决1. GPU 显存占满、资源无法释放解决办法：重启Colab会话（菜单栏->Runtime->Restartsession） 六、扩展：保存 LoRA 权重到谷歌云盘from google.colab import driveimport osimport shutil
# 1. 挂载 Google 云盘print("正在挂载 Google 云盘...")if not os.path.ismount('/content/drive'): drive.mount('/content/drive') print("Google 云盘已挂载。")else: print("Google 云盘已挂载。")
# 2. 定义路径（可修改目标路径）source_dir_path = '/content/lora_adapters_v2' # LoRA权重源路径destination_folder_path = '/content/drive/MyDrive/lora' # 云盘目标文件夹destination_dir_path = os.path.join(destination_folder_path, os.path.basename(source_dir_path))
# 配置：是否覆盖已存在的目标目录OVERWRITE_EXISTING = True
# 3. 创建目标文件夹（若不存在）os.makedirs(destination_folder_path, exist_ok=True)print(f"目标备份文件夹: {destination_folder_path}")
# 4. 检查源目录有效性if not os.path.exists(source_dir_path) or not os.path.isdir(source_dir_path): print(f"❌ 错误：源目录不存在或不是目录 - {source_dir_path}")else: try: # 计算源目录大小 source_size_mb = sum( os.path.getsize(os.path.join(dirpath, filename)) for dirpath, dirnames, filenames in os.walk(source_dir_path) for filename in filenames ) / (1024 * 1024) print(f"✅ 找到源目录: {source_dir_path} (估算大小: {source_size_mb:.2f} MB)")
# 处理目标目录已存在的情况 if os.path.exists(destination_dir_path): dest_size_mb = sum( os.path.getsize(os.path.join(dirpath, filename)) for dirpath, dirnames, filenames in os.walk(destination_dir_path) for filename in filenames ) / (1024 * 1024) print(f"⚠️ 警告：目标目录已存在 - {destination_dir_path} (估算大小: {dest_size_mb:.2f} MB)") if OVERWRITE_EXISTING: print("正在删除旧目录以覆盖...") shutil.rmtree(destination_dir_path) print("旧目录已删除。") else: print("OVERWRITE_EXISTING=False，跳过复制。") print("如需覆盖，请将 OVERWRITE_EXISTING 设为 True。")
# 复制目录到云盘 if OVERWRITE_EXISTING or not os.path.exists(destination_dir_path): print(f"正在复制目录到云盘...") print(f" 源: {source_dir_path}") print(f" 目标: {destination_dir_path}") shutil.copytree(source_dir_path, destination_dir_path) print(f"✅ 目录已备份到云盘: {destination_dir_path}") else: print("操作已取消或跳过。")
except Exception as e: print(f"❌ 复制错误: {e}")
|
