使用Unsloth高效微调Gemma-3模型：完整实践指南

# 设置Hugging Face镜像加速下载
export HF_ENDPOINT=https://hf-mirror.com

# 指定GPU设备
export CUDA_VISIBLE_DEVICES=1

from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name="/media/do/llmhub/modelhub/gemma-3-4b-it",
    dtype=None,
    max_seq_length=1024,
    load_in_4bit=True,      # 4-bit量化，显著减少内存使用
    load_in_8bit=False,
    full_finetuning=False   # 使用PEFT而非全量微调
)

model = FastModel.get_peft_model(
    model,
    finetune_vision_layers=False,
    finetune_language_layers=True,
    finetune_attention_modules=True,   # 注意力模块对GRPO训练有益
    finetune_mlp_modules=True,
    
    r=8,                    # 秩参数：越大精度越高，但可能过拟合
    lora_alpha=8,          # 推荐 alpha == r
    lora_dropout=0,
    bias="none",
    random_state=3407,
)

from unsloth.chat_templates import get_chat_template

tokenizer = get_chat_template(
    tokenizer,
    chat_template="gemma-3",  # 使用Gemma-3专用模板
)

def formatting_prompts_func(examples):
    convos = examples["conversations"]
    texts = [
        tokenizer.apply_chat_template(
            convo, 
            tokenize=False, 
            add_generation_prompt=False
        ).removeprefix('<bos>') 
        for convo in convos
    ]
    return {"text": texts}

# 加载并格式化数据集
dataset = load_dataset("mlabonne/FineTome-100k", split="train")
dataset = standardize_data_formats(dataset)
dataset = dataset.map(formatting_prompts_func, batched=True)

from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    eval_dataset=None,
    args=SFTConfig(
        dataset_text_field="text",
        output_dir="./training_outputs",
        logging_dir="./training_logs",
        
        # 批次配置
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,  # 梯度累积模拟更大批次
        
        # 训练步数
        max_steps=30,                   # 快速验证设置
        warmup_steps=5,
        
        # 优化器配置
        learning_rate=2e-4,            # 短训练用2e-4，长训练建议2e-5
        optim="adamw_8bit",           # 8-bit优化器节省内存
        weight_decay=0.01,
        lr_scheduler_type="linear",
        
        # 日志和保存
        logging_steps=1,
        save_steps=10,
        save_strategy="steps",
        
        # 监控配置
        report_to="tensorboard",
        run_name="gemma3-finetune",
        
        # 性能优化
        dataloader_num_workers=2,
        seed=3407,
    ),
)

from unsloth.chat_templates import train_on_responses_only

trainer = train_on_responses_only(
    trainer,
    instruction_part="<start_of_turn>user\n",
    response_part="<start_of_turn>model\n",
)

import torch

# 解决dynamo重编译限制问题
torch._dynamo.config.cache_size_limit = 64
torch._dynamo.config.accumulated_cache_size_limit = 256

# 备选方案：完全禁用torch编译
# torch._dynamo.config.disable = True

#!/bin/bash
# 检查并安装TensorBoard
if ! command -v tensorboard &> /dev/null; then
    echo "安装TensorBoard..."
    pip install tensorboard
fi

# 创建必要目录
mkdir -p ./training_logs
mkdir -p ./training_outputs

# 启动TensorBoard
tensorboard --logdir=./training_logs --port=6006 --host=0.0.0.0

# 安装依赖
pip install -r requirements.txt

# 启动TensorBoard监控
./start_tensorboard.sh

# 运行微调脚本
./run-unsloth-finetune.sh