🎓 大模型微调完整指南

# finetune_qlora.py - 使用Unsloth进行QLoRA微调 from unsloth import FastLanguageModel import torch from datasets import load_dataset from trl import SFTTrainer from transformers import TrainingArguments # ============ 1. 配置参数 ============ max_seq_length = 2048 # 最大序列长度 dtype = None # 自动检测FP16/BF16 load_in_4bit = True # 使用4bit量化 # LoRA配置 lora_r = 16 # rank值，越大效果越好但越慢 lora_alpha = 16 # alpha = r 是常见配置 lora_dropout = 0 # dropout=0时速度更快 # ============ 2. 加载模型 ============ model, tokenizer = FastLanguageModel.from_pretrained( model_name = "unsloth/qwen2.5-7b-bnb-4bit", # 或其他模型 max_seq_length = max_seq_length, dtype = dtype, load_in_4bit = load_in_4bit, ) # ============ 3. 添加LoRA适配器 ============ model = FastLanguageModel.get_peft_model( model, r = lora_r, target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], lora_alpha = lora_alpha, lora_dropout = lora_dropout, bias = "none", use_gradient_checkpointing = True, # 节省显存 random_state = 3407, ) # ============ 4. 准备数据 ============ # 定义提示词模板 alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. ### Instruction: {} ### Input: {} ### Response: {}""" # 格式化函数 def formatting_prompts_func(examples): instructions = examples["instruction"] inputs = examples["input"] outputs = examples["output"] texts = [] for instruction, input_text, output in zip(instructions, inputs, outputs): text = alpaca_prompt.format(instruction, input_text, output) + tokenizer.eos_token texts.append(text) return { "text" : texts, } # 加载数据集 dataset = load_dataset("json", data_files={ "train": "./data/train.jsonl", "test": "./data/val.jsonl" }) # ============ 5. 训练参数配置 ============ trainer = SFTTrainer( model = model, tokenizer = tokenizer, train_dataset = dataset["train"], eval_dataset = dataset["test"], dataset_text_field = "text", formatting_func = formatting_prompts_func, max_seq_length = max_seq_length, packing = False, # 短序列时可设True提速 args = TrainingArguments( # 基础参数 output_dir = "./outputs", num_train_epochs = 3, per_device_train_batch_size = 2, per_device_eval_batch_size = 2, gradient_accumulation_steps = 4, # 有效batch=2*4=8 # 学习率和优化器 learning_rate = 2e-4, warmup_steps = 5, optim = "adamw_8bit", # 8bit优化器省显存 weight_decay = 0.01, lr_scheduler_type = "linear", # 日志和保存 logging_steps = 1, eval_strategy = "steps", eval_steps = 100, save_strategy = "steps", save_steps = 100, save_total_limit = 3, # 精度和性能 fp16 = not torch.cuda.is_bf16_supported(), bf16 = torch.cuda.is_bf16_supported(), seed = 3407, ), ) # ============ 6. 开始训练 ============ print("🚀 开始训练...") trainer_stats = trainer.train() # ============ 7. 保存模型 ============ # 保存LoRA适配器（仅几十MB） model.save_pretrained("./lora_model") tokenizer.save_pretrained("./lora_model") # 可选：合并为完整模型（需要更多显存） # model.save_pretrained_merged("./full_model", tokenizer, save_method="merged_16bit") print("✅ 训练完成！") print(f"📊 训练统计: {trainer_stats}") # ============ 8. 推理测试 ============ FastLanguageModel.for_inference(model) # 开启快速推理 inputs = tokenizer( alpaca_prompt.format( "请解释什么是LoRA", # instruction "", # input "", # output - 留空让模型生成 ), return_tensors = "pt" ).to("cuda") outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7) print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# finetune_peft.py from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training from datasets import load_dataset import torch # 1. 加载基础模型（4bit量化） model_name = "Qwen/Qwen2.5-7B-Instruct" # 量化配置 from transformers import BitsAndBytesConfig bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True, ) model = AutoModelForCausalLM.from_pretrained( model_name, quantization_config=bnb_config, device_map="auto", trust_remote_code=True ) tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) # 2. 准备模型（量化训练必需） model = prepare_model_for_kbit_training(model) # 3. LoRA配置 lora_config = LoraConfig( r=16, lora_alpha=32, target_modules=["q_proj", "k_proj", "v_proj", "o_proj"], lora_dropout=0.05, bias="none", task_type="CAUSAL_LM" ) # 4. 应用LoRA model = get_peft_model(model, lora_config) model.print_trainable_parameters() # 打印可训练参数量 # 5. 数据准备 dataset = load_dataset("json", data_files="./data/train.jsonl") # 6. 训练 training_args = TrainingArguments( output_dir="./outputs", num_train_epochs=3, per_device_train_batch_size=4, learning_rate=2e-4, logging_steps=10, save_steps=100, ) trainer = Trainer( model=model, args=training_args, train_dataset=dataset["train"], ) trainer.train() # 7. 保存 model.save_pretrained("./lora_adapter")