LoRA微调7B大模型全流程：从数据构建到推理效果对比

GPU: RTX 4090 24G
CUDA: 12.1
PyTorch: 2.3.1+cu121
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": "如何配置灰度发布规则？",
  "input": "",
  "output": "1. 登录控制台，进入\"发布管理\"。\n2. 选择目标服务，点击\"新建规则\"。\n3. 设置灰度比例（建议5%-10%）。\n4. 指定灰度用户标签或IP段。\n5. 保存并启用。"
}

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

model_name = "Qwen/Qwen2.5-7B-Instruct"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16
)

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()
# 输出: trainable params: 20,185,088 || all params: 7,635,000,000 || trainable%: 0.264

from datasets import load_dataset
from transformers import TrainingArguments, Trainer, DataCollatorForSeq2Seq

dataset = load_dataset("json", data_files={"train": "train.json", "validation": "val.json"})

def format_example(example):
    prompt = f"user\n{example['instruction']}\n{example['input']}\nassistant\n"
    answer = example['output'] + tokenizer.eos_token
    return {"prompt": prompt, "answer": answer}

def tokenize_fn(example):
    prompt_ids = tokenizer(example["prompt"], add_special_tokens=False)["input_ids"]
    answer_ids = tokenizer(example["answer"], add_special_tokens=False)["input_ids"]
    input_ids = prompt_ids + answer_ids
    labels = [-100] * len(prompt_ids) + answer_ids
    return {"input_ids": input_ids, "labels": labels, "attention_mask": [1]*len(input_ids)}

train_ds = dataset["train"].map(format_example).map(tokenize_fn, remove_columns=dataset["train"].column_names)
val_ds = dataset["validation"].map(format_example).map(tokenize_fn, remove_columns=dataset["validation"].column_names)

training_args = TrainingArguments(
    output_dir="./qwen2.5-7b-lora",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    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,
    optim="paged_adamw_8bit",
    report_to="none",
    save_total_limit=2,
    gradient_checkpointing=True
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_ds,
    eval_dataset=val_ds,
    data_collator=DataCollatorForSeq2Seq(tokenizer, padding=True, return_tensors="pt")
)

trainer.train()

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "./qwen2.5-7b-lora/checkpoint-563")
model = model.merge_and_unload()  # 合并LoRA权重，推理更快

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

def ask(question):
    messages = [{"role": "user", "content": question}]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.3,
        repetition_penalty=1.1,
        do_sample=True
    )
    return tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

print(ask("灰度发布怎么回滚？"))