LLM微调框架对比2025：LoRA、QLoRA、PEFT与Unsloth深度解析

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

# 1. 加载 4-bit 量化模型
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B",
    load_in_4bit=True,
    device_map="auto"
)

# 2. 准备模型用于训练
model = prepare_model_for_kbit_training(model)

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

# 4. 包装为 PEFT 模型
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# 输出: trainable params: 20M || all params: 8B || trainable%: 0.25%

from unsloth import FastLanguageModel

# 加载模型 —— 自动启用所有优化
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Meta-Llama-3.1-8B",
    max_seq_length=2048,
    dtype=None,           # 自动选择 BF16/FP16
    load_in_4bit=True,    # QLoRA 模式
)

# 添加 LoRA adapter
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    lora_alpha=16,
    use_gradient_checkpointing="unsloth",  # Unsloth 优化版本
)

# 训练 —— 比标准 PEFT 快 2-5 倍
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    max_seq_length=2048,
    args=TrainingArguments(per_device_train_batch_size=2, num_train_epochs=3),
)
trainer.train()

# 导出到 GGUF（一行代码）
model.save_pretrained_gguf("output", tokenizer, quantization_method="q4_k_m")

[
  {
    "instruction": "请将以下中文翻译成英文",
    "input": "欢迎使用我们的服务",
    "output": "Welcome to our service"
  }
]

# 1. 安装
# !pip install unsloth transformers datasets trl

# 2. 加载模型
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Meta-Llama-3.1-8B-bnb-4bit",
    max_seq_length=2048,
    load_in_4bit=True,
)

# 3. 添加 LoRA
model = FastLanguageModel.get_peft_model(model, r=16)

# 4. 准备数据
from datasets import load_dataset
dataset = load_dataset("json", data_files="train.json", split="train")

# 5. 训练
from trl import SFTTrainer
from transformers import TrainingArguments

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=2048,
    args=TrainingArguments(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        warmup_steps=10,
        max_steps=100,
        learning_rate=2e-4,
        logging_steps=10,
        optim="adamw_8bit",
        seed=3407,
        output_dir="outputs",
    ),
)
trainer.train()

# 6. 保存并导出
model.save_pretrained("lora_adapter")
model.save_pretrained_merged("merged_model")

from peft import AutoPeftModelForCausalLM
model = AutoPeftModelForCausalLM.from_pretrained("lora_adapter")
model = model.merge_and_unload()  # 合并为一个完整模型
model.save_pretrained("merged_model")

python convert_hf_to_gguf.py --outfile model.gguf merged_model/

python -m vllm.entrypoints.openai.api_server \
  --model merged_model \
  --tensor-parallel-size 1 \
  --port 8000