1
# 更新系统
2
sudo apt update && sudo apt upgrade -y
3

4
# 安装基础工具
5
sudo apt install -y build-essential curl git python3-pip
6

7
# 配置NVIDIA官方仓库（Ubuntu 22.04）
8
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
9
sudo dpkg -i cuda-keyring_1.1-1_all.deb
10
sudo apt update
11

12
# 一键安装CUDA（包含驱动、toolkit、cuDNN等）
13
sudo apt install -y cuda

1
echo 'export CUDA_HOME=/usr/local/cuda' >> ~/.bashrc
2
echo 'export PATH=$CUDA_HOME/bin:$PATH' >> ~/.bashrc
3
echo 'export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
4
source ~/.bashrc
5

6
# 验证安装
7
nvcc --version

1
python3 -m pip install --user virtualenv
2
python3 -m virtualenv ~/llm_env
3
source ~/llm_env/bin/activate

1
# 安装PyTorch（CUDA 12.1版本）
2
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
3

4
# 安装Hugging Face生态和Unsloth
5
pip install transformers datasets accelerate peft
6
pip install unsloth

1
python -c "import torch; print(torch.cuda.is_available())"

1
{
2
  "instruction": "用甄嬛体描述今天的天气",
3
  "input": "今天气温28度，晴天",
4
  "output": "今儿的天气倒是个好日子，暖风和煦，晴空万里，正是踏青的好时节。"
5
}

1
pip install unsloth

1
from unsloth import FastLanguageModel
2
import torch
3
from trl import SFTTrainer
4
from transformers import TrainingArguments
5

6
# 1. 加载模型和分词器（4位量化，大幅降低显存）
7
model, tokenizer = FastLanguageModel.from_pretrained(
8
    model_name="unsloth/Qwen2.5-3B-Instruct",
9
    max_seq_length=2048,
10
    load_in_4bit=True,
11
)
12

13
# 2. 添加LoRA适配器
14
model = FastLanguageModel.get_peft_model(
15
    model,
16
    r=16,  # LoRA秩，建议8-128，越大越精细但显存需求越高
17
    lora_alpha=16,  # 缩放因子，通常设为和r相同或2倍
18
    lora_dropout=0,  # 设为0避免过拟合
19
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
20
)
21

22
# 3. 准备数据集（格式：instruction-input-output）
23
dataset = [
24
    {"instruction": "用甄嬛体描述今天的天气", "input": "气温28度，晴天", "output": "今儿个日头正好，暖风和煦，晴空万里，正是踏青的好时节。"},
25
    {"instruction": "用甄嬛体表示拒绝", "input": "我不想帮忙", "output": "这事儿怕是不妥，本宫心有余而力不足呢。"},
26
    # 添加更多数据...
27
]
28

29
def format_example(example):
30
    return f"### 指令：{example['instruction']}\n### 输入：{example['input']}\n### 回答：{example['output']}"
31

32
# 4. 配置训练参数
33
trainer = SFTTrainer(
34
    model=model,
35
    tokenizer=tokenizer,
36
    train_dataset=dataset,  # 用格式化后的数据
37
    args=TrainingArguments(
38
        per_device_train_batch_size=2,
39
        gradient_accumulation_steps=4,
40
        warmup_steps=5,
41
        max_steps=60,
42
        learning_rate=2e-4,
43
        fp16=not torch.cuda.is_bf16_supported(),
44
        bf16=torch.cuda.is_bf16_supported(),
45
        logging_steps=10,
46
        output_dir="outputs",
47
    ),
48
)
49

50
# 5. 开始训练
51
trainer.train()
52

53
# 6. 保存模型
54
model.save_pretrained("qwen3b_zhinv_style")

1
python train.py

1
# 安装依赖
2
!pip install unsloth trl transformers datasets accelerate peft
3

4
# 后续代码与第6节相同

1
!pip install unsloth trl transformers datasets accelerate peft

1
# 合并LoRA权重
2
merged_model = model.merge_and_unload()
3

4
# 导出为GGUF格式（用于llama.cpp/Ollama）
5
from unsloth import export_to_gguf
6
export_to_gguf(merged_model, tokenizer, "my_model_q4_K_M.gguf", quantization="q4_K_M")

1
# 安装Ollama
2
curl -fsSL https://ollama.com/install.sh | sh
3

4
# 创建Modelfile
5
echo 'FROM ./my_model_q4_K_M.gguf' > Modelfile
6
ollama create my_model -f Modelfile
7

8
# 运行
9
ollama run my_model

1
# 编译llama.cpp
2
git clone https://github.com/ggerganov/llama.cpp
3
cd llama.cpp
4
make -j
5

6
# 运行
7
./main -m my_model_q4_K_M.gguf -p "用甄嬛体说今天天气真好" -n 128