** 基于LoRA与QLoRA微调Qwen2.5-7B：数据构建、显存优化与推理效果全记录

pip install torch==2.3.1 transformers==4.44.2 peft==0.12.0 bitsandbytes==0.43.3 trl==0.9.6 accelerate==0.33.0 datasets==2.20.0

{
  "instruction": "客户询问退货政策中关于拆封商品的处理方式，请给出准确回答。",
  "input": "商品已拆封但未使用，能否退货？",
  "output": "根据平台退货政策，商品拆封但未使用且配件齐全的情况下，支持7天无理由退货。您需要在订单页面提交换货申请，并保证商品外包装完整。"
}

from datasets import load_dataset
from transformers import AutoTokenizer

model_path = "Qwen/Qwen2.5-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

def format_sample(sample):
    prompt = f"### 指令:\n{sample['instruction']}\n\n### 输入:\n{sample['input']}\n\n### 回答:\n{sample['output']}"
    return {"text": prompt + tokenizer.eos_token}

dataset = load_dataset("json", data_files="data/train.json", split="train")
dataset = dataset.map(format_sample, remove_columns=dataset.column_names)

def tokenize_fn(examples):
    tokens = tokenizer(
        examples["text"],
        truncation=True,
        max_length=1024,
        padding=False,
    )
    tokens["labels"] = tokens["input_ids"].copy()
    return tokens

tokenized = dataset.map(tokenize_fn, batched=True, remove_columns=["text"])

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training

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_path,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
model = prepare_model_for_kbit_training(model)

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()

from transformers import TrainingArguments
from trl import SFTTrainer

training_args = TrainingArguments(
    output_dir="./output/qwen2.5-7b-lora",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    num_train_epochs=3,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_strategy="epoch",
    evaluation_strategy="epoch",
    bf16=True,
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",
    report_to="none",
    save_total_limit=2,
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=tokenized,
    tokenizer=tokenizer,
    max_seq_length=1024,
    packing=False,
)

trainer.train()

model.config.use_cache = False
model.gradient_checkpointing_enable()

export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    model_path, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "./output/qwen2.5-7b-lora/checkpoint-2250")
model.eval()

prompt = "### 指令:\n客户询问退货政策中关于拆封商品的处理方式，请给出准确回答。\n\n### 输入:\n商品已拆封但未使用，能否退货？\n\n### 回答:\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))