LoRA微调7B模型实录：单卡24G从数据构造到推理对比的完整链路

import json
from datasets import load_dataset

# 原始数据示例构造
sample = {
    "instruction": "你是客服工单处理助手。请阅读工单内容，输出JSON，包含category（23类之一）和reply（给用户的回复，不超过80字，不要出现'联系人工客服'）。",
    "input": "我上周三下的单，物流显示已发货但三天没更新了，到底什么时候能到？",
    "output": json.dumps({
        "category": "物流停滞",
        "reply": "您的包裹当前在运输途中，因中转站积压导致更新延迟，预计48小时内会有新物流信息，请留意。"
    }, ensure_ascii=False)
}

with open("train.jsonl", "w", encoding="utf-8") as f:
    for s in raw_samples:  # 你清洗后的样本列表
        f.write(json.dumps(s, ensure_ascii=False) + "\n")

dataset = load_dataset("json", data_files={
    "train": "train.jsonl",
    "validation": "val.jsonl"
})
print(dataset)
# DatasetDict({train: 2560, validation: 320})

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from peft import LoraConfig, get_peft_model, TaskType
from trl import SFTTrainer
from datasets import load_dataset

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

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

lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    bias="none",
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 40,370,176 || all params: 7,655,986,688 || trainable%: 0.5273

def format_sample(example):
    text = (
        f"system\n{example['instruction']}\n"
        f"user\n{example['input']}\n"
        f"assistant\n{example['output']}"
    )
    return {"text": text}

dataset = load_dataset("json", data_files={"train": "train.jsonl", "validation": "val.jsonl"})
dataset = dataset.map(format_sample, remove_columns=dataset["train"].column_names)

training_args = TrainingArguments(
    output_dir="./lora-qwen2.5-7b",
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    num_train_epochs=3,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    eval_strategy="steps",
    eval_steps=50,
    save_strategy="steps",
    save_steps=200,
    bf16=True,
    gradient_checkpointing=True,
    optim="adamw_torch",
    report_to="none",
    seed=42,
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
    tokenizer=tokenizer,
    dataset_text_field="text",
    max_seq_length=1024,
    packing=False,
)

trainer.train()
trainer.save_model("./lora-qwen2.5-7b/final")

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

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "./lora-qwen2.5-7b/final")
model = model.merge_and_unload()  # 合并后就是一个普通7B模型
model.eval()

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
prompt = "system\n你是客服工单处理助手...\nuser\n{工单}\nassistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))