LoRA与QLoRA微调7B大模型：从数据准备到推理效果对比的完整记录

import json
from datasets import Dataset

def build_dataset(path):
    records = []
    with open(path, 'r', encoding='utf-8') as f:
        for line in f:
            item = json.loads(line)
            # 走Qwen的chat template
            messages = [
                {"role": "system", "content": "你是一名专业的电商售后客服。"},
                {"role": "user", "content": item["instruction"] + 
                 (("\n" + item["input"]) if item.get("input") else "")},
                {"role": "assistant", "content": item["output"]}
            ]
            records.append({"messages": messages})
    return Dataset.from_list(records)

train_ds = build_dataset("data/train.jsonl")
eval_ds = build_dataset("data/val.jsonl")
print(f"train: {len(train_ds)}, eval: {len(eval_ds)}")
# train: 11000, eval: 1000

import torch
from transformers import (
    AutoModelForCausalLM, AutoTokenizer,
    BitsAndBytesConfig, TrainingArguments
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer

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

# 1. 4-bit量化配置（QLoRA核心）
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
model = prepare_model_for_kbit_training(model)
model.config.use_cache = False  # 训练时必须关，否则和gradient checkpointing冲突

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

# 3. 训练参数
training_args = TrainingArguments(
    output_dir="./output/qwen2.5-7b-lora",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,   # 等效batch = 16
    num_train_epochs=3,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_strategy="epoch",
    eval_strategy="epoch",
    bf16=True,
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",
    max_grad_norm=0.3,
    report_to="none",
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_ds,
    eval_dataset=eval_ds,
    peft_config=lora_config,
    tokenizer=tokenizer,
    max_seq_length=1024,
    packing=False,
)

trainer.train()
trainer.model.save_pretrained("./output/qwen2.5-7b-lora/final")
tokenizer.save_pretrained("./output/qwen2.5-7b-lora/final")

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base, "./output/qwen2.5-7b-lora/final")
merged = model.merge_and_unload()
merged.save_pretrained("./output/qwen2.5-7b-merged")