公司网站模版 dedecms,广西和住房城乡建设厅网站,推广步骤,网站伪静态全站伪静态前言
大语言模型(LLM)的微调是当前AI领域的热门话题#xff0c;而参数高效微调方法(如LoRA)因其低成本和高效率备受关注。本文将手把手教你如何使用Qwen2.5-3B-Instruct模型进行LoRA微调#xff0c;并构建完整的推理流程。
一、环境准备
1.1 硬件要求
• GPU: 至少16GB显存(如…前言
大语言模型(LLM)的微调是当前AI领域的热门话题而参数高效微调方法(如LoRA)因其低成本和高效率备受关注。本文将手把手教你如何使用Qwen2.5-3B-Instruct模型进行LoRA微调并构建完整的推理流程。
一、环境准备
1.1 硬件要求
• GPU: 至少16GB显存(如NVIDIA RTX 3090/A10)
• 内存: 32GB以上
• 存储: 50GB可用空间
开门见山首先给出完整微调代码和完整推理代码推理代码需要名为A1的jsonl文件文件内部结构如下完整微调代码
from modelscope import AutoModelForCausalLM, AutoTokenizer
import torch
import os
from datasets import load_dataset, Dataset
import pandas as pd
from trl import SFTTrainer, SFTConfig
from transformers import TextStreamer, TrainerCallback
from peft import LoraConfig, TaskType, get_peft_model, PeftModel
import gc
from tqdm.auto import tqdm
import re
import time# 环境设置
os.environ[CUDA_VISIBLE_DEVICES] 0
os.environ.pop(LOCAL_RANK, None)print(fGPU设备数: {torch.cuda.device_count()})
if torch.cuda.is_available():print(f当前GPU: {torch.cuda.get_device_name(0)})gpu_mem torch.cuda.get_device_properties(0).total_memory / 1024**3print(f可用GPU内存: {gpu_mem:.2f} GB)
else:print(未检测到GPU将使用CPU运行训练会非常缓慢)# 模型路径
model_name /root/autodl-tmp/Qwen/Qwen2.5-3B-Instruct# 加载分词器
tokenizer AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token tokenizer.eos_token# 模型加载
max_retries 3
model None
for attempt in range(max_retries):try:model AutoModelForCausalLM.from_pretrained(model_name,torch_dtypetorch.bfloat16 if torch.cuda.is_available() else torch.float32,device_mapauto if torch.cuda.is_available() else None,low_cpu_mem_usageTrue,trust_remote_codeTrue,load_in_8bitTrue,max_memory{0: f{int(gpu_mem * 0.7)}GB})model.enable_input_require_grads()breakexcept Exception as e:print(f模型加载尝试 {attempt 1}/{max_retries} 失败: {str(e)[:400]}...)if attempt max_retries - 1:raisetime.sleep(5)# LoRA配置
config LoraConfig(task_typeTaskType.CAUSAL_LM,target_modules[q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj],inference_modeFalse,r4,lora_alpha8,lora_dropout0.1,
)
model get_peft_model(model, config)# 修改后的数据集加载部分
try:# 先加载完整数据集full_ds load_dataset(bespokelabs/Bespoke-Stratos-17k, cache_dir./data/reasoning, splittrain)# 随机采样2000条ds full_ds.shuffle(seed3407).select(range(2000))
except Exception as e:print(f数据集加载失败: {e})# 回退方案full_ds load_dataset(bespokelabs/Bespoke-Stratos-17k, cache_dir./data/reasoning, splittrain,download_modereuse_cache_if_exists)ds full_ds.shuffle(seed3407).select(range(2000))def generate_a_format_data(examples):formatted_data []for conv in examples[conversations]:user_msg conv[0][value]assistant_msg conv[1][value]# 清理用户问题user_msg user_msg.replace(Return your final response within \\boxed{}., ).strip()# 提取思考部分think_part if |begin_of_thought| in assistant_msg:think_part assistant_msg.split(|begin_of_thought|)[1].split(|end_of_thought|)[0].strip()# 提取答案answer_part if \\boxed{ in assistant_msg:answer_part assistant_msg.split(\\boxed{)[1].split(})[0]elif |begin_of_solution| in assistant_msg:answer_part assistant_msg.split(|begin_of_solution|)[1].split(|end_of_solution|)[0].strip()# 构建标准格式formatted_data.append([{role: user, content: user_msg},{role: assistant, content: fthink{think_part}/think answer{answer_part}/answer}])return {conversations: formatted_data}# 数据处理添加随机性
processed_ds ds.map(generate_a_format_data,batchedTrue,batch_size5,remove_columnsds.column_names
).shuffle(seed3407) # 关键修改确保数据随机性# 格式验证
def validate_training_data(dataset, num_samples3):print(f\n 格式验证前 {num_samples} 条 )for i in range(min(num_samples, len(dataset))):content dataset[i][conversations][1][content]has_think think in content and /think in contenthas_answer answer in content and /answer in contentprint(f样本 {i1}: {✅ if has_think and has_answer else ❌})if not (has_think and has_answer):print(f问题内容: {content[:100]}...)validate_training_data(processed_ds)# 应用模板
reasoning_conversations tokenizer.apply_chat_template(processed_ds[conversations],tokenizeFalse,add_generation_promptTrue
)# 创建训练集
df pd.DataFrame({text: reasoning_conversations})
train_ds Dataset.from_pandas(df).shuffle(seed3407)# 训练回调
class TQDMCallback(TrainerCallback):def on_train_begin(self, args, state, control, **kwargs):self.total_steps state.max_stepsself.progress_bar tqdm(totalself.total_steps, desc训练进度, bar_format{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}{remaining}, {rate_fmt}])def on_step_end(self, args, state, control, logsNone, **kwargs):self.progress_bar.update(1)if logs and loss in logs:self.progress_bar.set_postfix({loss: f{logs[loss]:.4f}})gc.collect()if torch.cuda.is_available():torch.cuda.empty_cache()def on_train_end(self, args, state, control, **kwargs):self.progress_bar.close()if state.log_history:final_loss state.log_history[-1].get(loss, float(nan))print(f训练完成! 最终损失: {final_loss:.4f})# 训练配置
training_args SFTConfig(output_dir./lora_model_a榜,per_device_train_batch_size1,gradient_accumulation_steps4,warmup_steps3,num_train_epochs2,learning_rate1e-4,logging_steps1,optimadamw_8bit,weight_decay0.01,lr_scheduler_typelinear,seed3407,report_tonone,fp16torch.cuda.is_available(),max_grad_norm0.5,logging_first_stepTrue,save_steps10,max_seq_length1500,
)# 训练器
trainer SFTTrainer(modelmodel,train_datasettrain_ds,eval_datasetNone,callbacks[TQDMCallback()],argstraining_args,
)# 清理内存
gc.collect()
if torch.cuda.is_available():torch.cuda.empty_cache()# 开始训练
print(\n开始训练...)
trainer.train()# 保存模型
print(\n保存模型...)
model.save_pretrained(lora_model_a榜)
tokenizer.save_pretrained(lora_model_a榜)# 清理
del model
gc.collect()
if torch.cuda.is_available():torch.cuda.empty_cache()# 推理测试部分
print(\n 加载微调后的模型进行推理测试 )# 加载模型函数
def load_finetuned_model():base_model AutoModelForCausalLM.from_pretrained(model_name,torch_dtypetorch.bfloat16 if torch.cuda.is_available() else torch.float32,device_mapauto,trust_remote_codeTrue)model PeftModel.from_pretrained(base_model,lora_model_a榜,torch_dtypetorch.bfloat16 if torch.cuda.is_available() else torch.float32,)model.eval()return model# 推理函数
def ask(question, temperature0.1, max_new_tokens2048):print(f\n问题: {question})model load_finetuned_model()messages [{role: user, content: question}]# 应用模板text tokenizer.apply_chat_template(messages,tokenizeFalse,add_generation_promptTrue)# 流式输出streamer TextStreamer(tokenizer, skip_promptTrue)print(\n模型回答:)with torch.no_grad():inputs tokenizer(text, return_tensorspt).to(model.device)outputs model.generate(**inputs,max_new_tokensmax_new_tokens,temperaturetemperature,streamerstreamer,pad_token_idtokenizer.pad_token_id,eos_token_idtokenizer.eos_token_id)full_output tokenizer.decode(outputs[0], skip_special_tokensFalse)# 格式验证if think not in full_output or answer not in full_output:print(\n⚠️ 警告: 输出缺少必要标记)return full_output# 测试蘑菇含水率问题
question (You have 1000 kg of mushrooms with a water content of 99%After drying for a few days, the water content drops to 98%How much water needs to be evaporated?
)answer ask(question)
print(\n完整输出:)
print(answer[:1500] (... if len(answer) 1500 else ))# 测试其他问题示例
test_questions [Calculate the area of a circle with radius 5,Solve for x: 2x 5 15,Explain Newtons first law of motion
]for q in test_questions:ask(q, temperature0.1)完整推理代码
import json
import os
import torch
from typing import Dict, List, Union
from vllm import LLM, SamplingParams
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from dataclasses import dataclassdataclass
class GenerationConfig:max_tokens: inttemperature: float 0.3top_p: float 0.95n: int 1class Infer:DEFAULT_MODEL /root/autodl-tmp/Qwen/Qwen2.5-3B-InstructTEMP_MODEL_DIR merged_temp_modeldef __init__(self, base_model: str DEFAULT_MODEL,lora_path: str lora_model_a榜,device: str auto):self._validate_paths(base_model, lora_path)self.device deviceself.tokenizer self._load_tokenizer(base_model)self.base_model self._load_base_model(base_model)self.merged_model self._merge_lora(lora_path)self.llm self._init_vllm_engine()self.sampling_params self._init_sampling_params()def _validate_paths(self, *paths):for path in paths:if not os.path.exists(path) and not path.startswith((http://, https://)):if / not in path:raise FileNotFoundError(f路径不存在: {path})def _load_tokenizer(self, model_path: str):try:tokenizer AutoTokenizer.from_pretrained(model_path,trust_remote_codeTrue)if tokenizer.pad_token is None:tokenizer.pad_token tokenizer.eos_tokenspecial_tokens [think, /think, answer, /answer]tokenizer.add_special_tokens({additional_special_tokens: special_tokens})return tokenizerexcept Exception as e:raise RuntimeError(f加载tokenizer失败: {str(e)})def _load_base_model(self, model_path: str):try:model AutoModelForCausalLM.from_pretrained(model_path,torch_dtypetorch.bfloat16,device_mapself.device,trust_remote_codeTrue)model.resize_token_embeddings(len(self.tokenizer))return modelexcept Exception as e:raise RuntimeError(f加载基础模型失败: {str(e)})def _merge_lora(self, lora_path: str):try:print(开始合并LoRA权重...)lora_model PeftModel.from_pretrained(self.base_model,lora_path,torch_dtypetorch.bfloat16)merged_model lora_model.merge_and_unload()test_input Calculate the area of a circle with radius 5messages [{role: user, content: test_input}]input_text self.tokenizer.apply_chat_template(messages,tokenizeFalse,add_generation_promptTrue)inputs self.tokenizer(input_text, return_tensorspt).to(self.base_model.device)with torch.no_grad():outputs merged_model.generate(**inputs,max_new_tokens2000,temperature0.1,pad_token_idself.tokenizer.pad_token_id,eos_token_idself.tokenizer.eos_token_id)output_text self.tokenizer.decode(outputs[0], skip_special_tokensFalse)print(\n 合并后模型测试输出 )print(output_text)os.makedirs(self.TEMP_MODEL_DIR, exist_okTrue)merged_model.save_pretrained(self.TEMP_MODEL_DIR)self.tokenizer.save_pretrained(self.TEMP_MODEL_DIR)print(✅ LoRA权重合并成功)return merged_modelexcept Exception as e:raise RuntimeError(f合并LoRA失败: {str(e)})def _init_vllm_engine(self):try:return LLM(modelself.TEMP_MODEL_DIR,tokenizerself.TEMP_MODEL_DIR,trust_remote_codeTrue,max_model_len2048,gpu_memory_utilization0.8,enforce_eagerTrue)except Exception as e:raise RuntimeError(f初始化vLLM引擎失败: {str(e)})def _init_sampling_params(self) - Dict[str, SamplingParams]:base_params {temperature: 0.3,top_p: 0.90,repetition_penalty: 1.5,stop: [/answer],skip_special_tokens: False}return {math: SamplingParams(**{**base_params, max_tokens: 1024}),choice: SamplingParams(**{**base_params, max_tokens: 1024}),code-generate: SamplingParams(**{**base_params, max_tokens: 2048, n: 3}),generic-generate: SamplingParams(**{**base_params, max_tokens:1024})}def _validate_output_format(self, text: str) - bool:required_components [think,/think,answer,/answer,]return all(comp in text for comp in required_components)def infer(self, data_file: str A1.jsonl) - Dict:if not os.path.exists(data_file):raise FileNotFoundError(f数据文件不存在: {data_file})try:with open(data_file, r, encodingutf-8) as f:data [json.loads(line) for line in f]except Exception as e:raise RuntimeError(f读取数据文件失败: {str(e)})results {result: {results: []}}for item in data:try:messages [{role: user, content: item[prompt]}]prompt self.tokenizer.apply_chat_template(messages,tokenizeFalse,add_generation_promptTrue)outputs self.llm.generate(prompt, self.sampling_params[item[type]])if item[type] code-generate:generated_text [o.text for output in outputs for o in output.outputs][:3]else:generated_text outputs[0].outputs[0].textresults[result][results].append({id: item[id],content: generated_text})except Exception as e:print(f处理样本 {item.get(id)} 失败: {str(e)})results[result][results].append({id: item.get(id),error: str(e)})return resultsdef cleanup(self):if os.path.exists(self.TEMP_MODEL_DIR):import shutilshutil.rmtree(self.TEMP_MODEL_DIR)if __name__ __main__:infer Nonetry:print(开始初始化推理引擎...)infer Infer(base_model/root/autodl-tmp/Qwen/Qwen2.5-3B-Instruct,lora_pathlora_model_a榜)results infer.infer(data_fileA1.jsonl)with open(res1.json, w, encodingutf-8) as f:json.dump(results, f, ensure_asciiFalse, indent2)print(✅ 推理完成结果已保存到res.json)except Exception as e:print(f运行失败: {str(e)})finally:if infer is not None:infer.cleanup()核心库安装
pip install torch transformers datasets peft trl vllm -U
pip install modelscope -U二、模型加载与配置
2.1 模型与分词器加载
from modelscope import AutoModelForCausalLM, AutoTokenizer
import torchmodel_name /root/autodl-tmp/Qwen/Qwen2.5-3B-Instruct分词器加载
tokenizer AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token tokenizer.eos_token # 设置填充token模型加载(8bit量化节省显存)
model AutoModelForCausalLM.from_pretrained(model_name,torch_dtypetorch.bfloat16,device_mapauto,load_in_8bitTrue,trust_remote_codeTrue
)2.2 LoRA配置
from peft import LoraConfig, TaskType, get_peft_modellora_config LoraConfig(task_typeTaskType.CAUSAL_LM,target_modules[q_proj, k_proj, v_proj, o_proj],inference_modeFalse,r8, # LoRA秩lora_alpha16,lora_dropout0.05,
)model get_peft_model(model, lora_config)
model.print_trainable_parameters() # 查看可训练参数占比三、数据处理与训练
3.1 数据集准备
我们使用Bespoke-Stratos-17k数据集包含17k条结构化问答数据
from datasets import load_datasetdataset load_dataset(bespokelabs/Bespoke-Stratos-17k, splittrain)3.2 数据格式化
将原始数据转换为标准对话格式
def format_conversation(example):user_msg example[conversations][0][value]assistant_msg example[conversations][1][value]# 提取思考过程和答案think_part assistant_msg.split(|begin_of_thought|)[1].split(|end_of_thought|)[0]answer_part assistant_msg.split(|begin_of_solution|)[1].split(|end_of_solution|)[0]return {conversations: [{role: user, content: user_msg},{role: assistant, content: fthink{think_part}/think answer{answer_part}/answer}]}formatted_ds dataset.map(format_conversation, batchedFalse)3.3 训练配置
from trl import SFTTrainer, SFTConfigtraining_args SFTConfig(output_dir./qwen_lora,per_device_train_batch_size2,gradient_accumulation_steps4,num_train_epochs3,learning_rate2e-5,optimadamw_8bit,max_seq_length1024,logging_steps10,save_steps100
)trainer SFTTrainer(modelmodel,train_datasetformatted_ds,dataset_text_fieldconversations,tokenizertokenizer,argstraining_args
)3.4 训练过程监控
自定义回调函数监控训练进度
from tqdm.auto import tqdm
from transformers import TrainerCallbackclass ProgressCallback(TrainerCallback):def on_train_begin(self, args, state, control, **kwargs):self.progress_bar tqdm(totalstate.max_steps, descTraining)def on_step_end(self, args, state, control, **kwargs):self.progress_bar.update(1)if loss in state.log_history[-1]:self.progress_bar.set_postfix({loss: state.log_history[-1][loss]})def on_train_end(self, args, state, control, **kwargs):self.progress_bar.close()trainer.add_callback(ProgressCallback())四、模型推理与部署
4.1 基础推理函数
from transformers import TextStreamerdef generate_response(prompt, model, tokenizer, max_length1024):messages [{role: user, content: prompt}]input_text tokenizer.apply_chat_template(messages,tokenizeFalse,add_generation_promptTrue)streamer TextStreamer(tokenizer, skip_promptTrue)inputs tokenizer(input_text, return_tensorspt).to(model.device)outputs model.generate(**inputs,max_new_tokensmax_length,streamerstreamer,temperature0.7,top_p0.9)return tokenizer.decode(outputs[0], skip_special_tokensTrue)4.2 使用vLLM加速推理
from vllm import LLM, SamplingParamsclass VLLMInference:def __init__(self, model_path, lora_pathNone):self.llm LLM(modelmodel_path,enable_lora(lora_path is not None),lora_modules[{name: lora, path: lora_path}] if lora_path else None,max_model_len2048)self.sampling_params SamplingParams(temperature0.7,top_p0.9,max_tokens1024)def generate(self, prompt):return self.llm.generate(prompt, self.sampling_params)4.3 批量推理处理
import json
from typing import List, Dictdef batch_inference(questions: List[str], inferencer) - Dict[str, str]:results {}for q in questions:output inferencer.generate(q)results[q] output[0].outputs[0].textreturn results五、实战案例数学问题求解
5.1 测试样例
test_questions [Calculate the area of a circle with radius 5,Solve for x: 2x 5 15,You have 1000 kg of mushrooms with water content 99%. After drying, the water content drops to 98%. How much water was evaporated?
]5.2 执行推理
inferencer VLLMInference(model_name,lora_pathqwen_lora
)results batch_inference(test_questions, inferencer)for q, ans in results.items():print(fQ: {q}\nA: {ans}\n{*50})六、性能优化技巧
6.1 显存优化
梯度检查点
model.gradient_checkpointing_enable()8bit Adam优化器
training_args.optim adamw_8bit梯度累积
training_args.gradient_accumulation_steps 46.2 推理加速
使用Flash Attention
model AutoModelForCausalLM.from_pretrained(model_name,torch_dtypetorch.bfloat16,use_flash_attention_2True
)vLLM的连续批处理
sampling_params SamplingParams(batch_size8, # 同时处理8个请求max_num_batched_tokens4096
)七、常见问题解答
Q1: 训练时出现OOM(显存不足)怎么办
A: 尝试以下方法
减小per_device_train_batch_size增加gradient_accumulation_steps启用load_in_4bit代替8bit
Q2: 模型生成内容重复怎么办
A: 调整生成参数
generation_config {
“repetition_penalty”: 1.5,
“no_repeat_ngram_size”: 3,
“do_sample”: True
}
Q3: 如何评估模型效果
建议使用以下指标
准确率(Accuracy)BLEU分数人工评估(最重要)
结语
通过本文的实践指南您已经掌握了使用LoRA微调Qwen2.5-3B-Instruct的核心技术。关键要点
LoRA可大幅降低训练成本(仅微调0.1%参数)数据质量比数量更重要vLLM可显著提升推理速度
完整代码已开源在[GitHub仓库链接]
下一步建议
• 尝试不同的LoRA秩(r4,8,16)比较效果
• 探索QLoRA等4bit微调方法
• 构建Web demo展示成果
