大模型微调实战：从 HuggingFace 下载到 QLoRA 部署，Java 程序员也能玩转 LLM

import os
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="meta-llama/Llama-3.1-8B-Instruct",
    local_dir="./models/llama-3.1-8b",
    resume_download=True,  # 断点续传
)

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")

messages = [
    {"role": "system", "content": "你是一个专业的代码助手"},
    {"role": "user", "content": "写一个快速排序"},
]

prompt = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True
)

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
你是一个专业的代码助手<|eot_id|><|start_header_id|>user<|end_header_id|>
写一个快速排序<|eot_id|><|start_header_id|>assistant<|end_header_id|>

from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=[
        "q_proj", "v_proj", "k_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()

trainable params: 20,185,088 || all params: 8,062,263,808 || trainable%: 0.2504

from transformers import BitsAndBytesConfig

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

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-70B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
)

model = get_peft_model(model, lora_config)

from transformers import TrainingArguments, Trainer
import wandb

wandb.init(project="llama-finetune", name="qlora-70b-run1")

training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    logging_steps=10,
    eval_steps=100,
    save_steps=500,
    report_to="wandb",
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_data,
    eval_dataset=eval_data,
)

trainer.train()

import asyncio

async def train_async():
    loop = asyncio.get_event_loop()
    await loop.run_in_executor(None, trainer.train)
    return "训练完成"

task = asyncio.create_task(train_async())

./lora-adapter/
├── adapter_config.json
├── adapter_model.safetensors

from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(model, "./lora-adapter")

merged_model = model.merge_and_unload()
merged_model.save_pretrained("./merged-model")

// 从 HuggingFace Hub 下载模型（自动缓存到 ~/.jpy-ml/llm-models/）
LLMModel model = LLMModel.download("Qwen/Qwen2.5-0.5B-Instruct");

// 对话推理
ChatResponse response = model.chat(
    ChatMessage.system("你是一个专业的代码助手"),
    ChatMessage.user("写一个快速排序")
);
System.out.println(response.getContent());
System.out.println("Tokens: prompt=" + response.getPromptTokens() 
    + " completion=" + response.getCompletionTokens());

// LoRA 微调
LLMTrainingResult result = model.fineTune()
    .lora(LoRAConfig.create().rank(16).alpha(32))
    .dataset("training_data.jsonl")
    .config(LLMTrainConfig.create()
        .epochs(3)
        .batchSize(4)
        .gradientAccumulation(4)
        .learningRate(2e-4f)
        .maxSeqLength(2048))
    .run((step, log) -> {
        System.out.println("Step " + step + ": " + log);
    });

// 异步微调
CompletableFuture<LLMTrainingResult> future = model.fineTune()
    .lora(LoRAConfig.create().rank(8).alpha(16))
    .dataset("data.jsonl")
    .config(LLMTrainConfig.create().epochs(2))
    .runAsync((step, log) -> {
        System.out.println("[异步] " + log);
    });

// 合并适配器到基座模型
String mergedPath = LLMModel.mergeAdapter(
    model.getModelPath(),
    result.getAdapterPath(),
    "/path/to/merged-model"
);