7B模型LoRA微调显存压至16G：Qwen2.5指令遵循提升42%

Python 3.10.12
torch 2.1.2+cu118
transformers 4.38.2
peft 0.9.0
datasets 2.17.0
accelerate 0.27.0
bitsandbytes 0.43.1

用户：我想挂明天下午王医生的号
意图：挂号预约
槽位：{"科室": "心血管内科", "时间": "明天下午", "医生": "王医生"}

from datasets import Dataset
from transformers import AutoTokenizer

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

def format_example(example):
    # 按Qwen2.5的chat模板格式化
    messages = [
        {"role": "user", "content": example["query"]},
        {"role": "assistant", "content": f"意图：{example['intent']}\n槽位：{example['slots']}"}
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
    return {"text": text}

dataset = Dataset.from_json("train.jsonl").map(format_example)

def tokenize_function(examples):
    model_inputs = tokenizer(examples["text"], truncation=True, max_length=512, padding="max_length")
    # 关键：labels和input_ids相同，但padding部分要设为-100
    labels = []
    for i, input_ids in enumerate(model_inputs["input_ids"]):
        # 找到第一个eos_token的位置，后面的padding都算-100
        eos_pos = input_ids.index(tokenizer.eos_token_id) if tokenizer.eos_token_id in input_ids else len(input_ids)
        labels.append([-100 if j >= eos_pos else input_id for j, input_id in enumerate(input_ids)])
    model_inputs["labels"] = labels
    return model_inputs

tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset.column_names)
tokenized_dataset.save_to_disk("train_tokenized")

from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from transformers import TrainingArguments, Trainer, AutoModelForCausalLM
import torch

# 加载模型（LoRA用BF16，QLoRA用4bit）
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    use_cache=False  # 必须关，否则gradient checkpointing冲突
)

# QLoRA额外加这两行
# from transformers import BitsAndBytesConfig
# bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.bfloat16)
# model = AutoModelForCausalLM.from_pretrained(..., quantization_config=bnb_config)
# model = prepare_model_for_kbit_training(model)

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()  # 应该输出7B参数里只训练了800万

training_args = TrainingArguments(
    output_dir="./qwen7b-lora-checkpoints",
    per_device_train_batch_size=8,
    gradient_accumulation_steps=16,  # 等效batch_size=8*16=128
    num_train_epochs=3,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_steps=200,
    logging_steps=50,
    save_steps=500,
    evaluation_strategy="steps",
    eval_steps=500,
    fp16=False,  # 用bf16，A100上fp16会溢出
    bf16=True,
    gradient_checkpointing=True,
    save_total_limit=2,
    load_best_model_at_end=True,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    eval_dataset=eval_tokenized,
)
trainer.train()