环境配置指南

# ============================================
# 课程核心依赖（一次性安装）
# ============================================

# 重要：Qwen3 要求 transformers >= 4.51.0
pip install "transformers>=4.51.0" \
    trl \
    peft \
    accelerate \
    bitsandbytes \
    datasets \
    tokenizers

# 推理加速（第4、5课需要）
pip install vllm

# 多模态微调（第5课选做）
pip install llamafactory

# 评估工具
pip install rouge-score nltk scikit-learn

# 可视化与实用工具
pip install matplotlib seaborn wandb gradio

# 数据处理
pip install pandas numpy tqdm

# ============================================
# 模型预下载脚本
# ============================================

# 设置模型缓存目录（建议放在数据盘）
export HF_HOME=/root/autodl-fs/huggingface
mkdir -p $HF_HOME

# 安装 huggingface_hub CLI
pip install huggingface_hub[cli]

# --- 主力模型（第1-4课必需）---
# Qwen3-1.7B（Instruct 版）
huggingface-cli download Qwen/Qwen3-1.7B --local-dir $HF_HOME/Qwen3-1.7B

# Qwen3-1.7B-Base（基座版，第4课 GRPO 需要）
huggingface-cli download Qwen/Qwen3-1.7B-Base --local-dir $HF_HOME/Qwen3-1.7B-Base

# --- 低配备选（T4 / RTX 4090 用户）---
# Qwen3-0.6B
huggingface-cli download Qwen/Qwen3-0.6B --local-dir $HF_HOME/Qwen3-0.6B

# Qwen3-0.6B-Base
huggingface-cli download Qwen/Qwen3-0.6B-Base --local-dir $HF_HOME/Qwen3-0.6B-Base

# --- 第5课量化实验 ---
# Qwen3-8B
huggingface-cli download Qwen/Qwen3-8B --local-dir $HF_HOME/Qwen3-8B

# --- 第5课多模态选做 ---
# Qwen3-VL-2B-Instruct
huggingface-cli download Qwen/Qwen3-VL-2B-Instruct --local-dir $HF_HOME/Qwen3-VL-2B-Instruct

# --- 第4-5课蒸馏对比 ---
# DeepSeek-R1-Distill-Qwen-1.5B（注意：基于 Qwen2.5，非 Qwen3）
huggingface-cli download deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \
    --local-dir $HF_HOME/DeepSeek-R1-Distill-Qwen-1.5B

# ============================================
# 数据集预下载
# ============================================

export DATA_DIR=/root/autodl-fs/datasets

# 第1课 - UltraChat
python -c "
from datasets import load_dataset
ds = load_dataset('HuggingFaceH4/ultrachat_200k', split='train_sft[:10000]')
ds.save_to_disk('$DATA_DIR/ultrachat_10k')
print(f'UltraChat 子集: {len(ds)} 条')
"

# 第3课 - UltraFeedback
python -c "
from datasets import load_dataset
ds = load_dataset('HuggingFaceH4/ultrafeedback_binarized', split='train_prefs[:5000]')
ds.save_to_disk('$DATA_DIR/ultrafeedback_5k')
print(f'UltraFeedback 子集: {len(ds)} 条')
"

# 第4课 - GSM8K
python -c "
from datasets import load_dataset
ds = load_dataset('openai/gsm8k', 'main')
ds.save_to_disk('$DATA_DIR/gsm8k')
print(f'GSM8K 训练集: {len(ds[\"train\"])} 条, 测试集: {len(ds[\"test\"])} 条')
"

#!/bin/bash
# setup.sh - 一键环境配置脚本
# 使用方法: bash setup.sh

set -e

echo "============================================"
echo "  LLM 后训练课程环境配置"
echo "============================================"

# 1. 检测 GPU
echo ""
echo "[1/5] 检测 GPU..."
if command -v nvidia-smi &> /dev/null; then
    GPU_NAME=$(nvidia-smi --query-gpu=name --format=csv,noheader | head -1)
    GPU_MEM=$(nvidia-smi --query-gpu=memory.total --format=csv,noheader | head -1)
    echo "  GPU: $GPU_NAME ($GPU_MEM)"
else
    echo "  警告: 未检测到 GPU！"
    echo "  部分实验将无法运行。"
fi

# 2. 安装 Python 依赖
echo ""
echo "[2/5] 安装 Python 依赖..."
pip install -q "transformers>=4.51.0" \
    trl peft accelerate bitsandbytes \
    datasets tokenizers \
    rouge-score nltk scikit-learn \
    matplotlib seaborn gradio \
    pandas numpy tqdm

# 3. 验证关键库版本
echo ""
echo "[3/5] 验证库版本..."
python -c "
import transformers, trl, peft, bitsandbytes, torch
print(f'  transformers: {transformers.__version__}')
print(f'  trl: {trl.__version__}')
print(f'  peft: {peft.__version__}')
print(f'  bitsandbytes: {bitsandbytes.__version__}')
print(f'  torch: {torch.__version__}')
print(f'  CUDA available: {torch.cuda.is_available()}')
if torch.cuda.is_available():
    print(f'  GPU: {torch.cuda.get_device_name(0)}')
    print(f'  显存: {torch.cuda.get_device_properties(0).total_mem / 1024**3:.1f} GB')

# 检查 transformers 版本
from packaging import version
assert version.parse(transformers.__version__) >= version.parse('4.51.0'), \
    f'transformers 版本过低: {transformers.__version__}, 需要 >= 4.51.0'
print('  transformers 版本检查: 通过')
"

# 4. 设置模型缓存目录
echo ""
echo "[4/5] 配置模型缓存目录..."
if [ -d "/root/autodl-fs" ]; then
    export HF_HOME=/root/autodl-fs/huggingface
    echo "export HF_HOME=/root/autodl-fs/huggingface" >> ~/.bashrc
    echo "  模型缓存: $HF_HOME (数据盘)"
else
    export HF_HOME=~/.cache/huggingface
    echo "  模型缓存: $HF_HOME (默认)"
fi
mkdir -p $HF_HOME

# 5. 快速验证：加载 tokenizer
echo ""
echo "[5/5] 快速验证..."
python -c "
try:
    from transformers import AutoTokenizer
    # 尝试加载 tokenizer（如果模型已下载）
    import os
    hf_home = os.environ.get('HF_HOME', '~/.cache/huggingface')
    local_path = os.path.join(hf_home, 'Qwen3-1.7B')
    if os.path.exists(local_path):
        tok = AutoTokenizer.from_pretrained(local_path)
        print('  Qwen3-1.7B tokenizer 加载: 成功')
    else:
        print('  Qwen3-1.7B 模型尚未下载（请运行模型下载脚本）')
    print('  环境验证: 通过')
except Exception as e:
    print(f'  验证失败: {e}')
"

echo ""
echo "============================================"
echo "  环境配置完成！"
echo "============================================"
echo ""
echo "下一步："
echo "  1. 运行模型下载脚本（如果尚未下载）"
echo "  2. 在 JupyterLab 中打开课程 notebook"
echo "  3. 实验结束后记得关机！"

# verify_env.py - 环境验证脚本

import sys

def check(name, test_fn):
    try:
        result = test_fn()
        print(f"  [PASS] {name}: {result}")
        return True
    except Exception as e:
        print(f"  [FAIL] {name}: {e}")
        return False

print("=" * 50)
print("环境验证")
print("=" * 50)

passed = 0
total = 0

# 1. PyTorch + CUDA
total += 1
if check("PyTorch + CUDA", lambda: (
    __import__('torch'),
    f"torch {__import__('torch').__version__}, "
    f"CUDA {'可用' if __import__('torch').cuda.is_available() else '不可用'}"
)[-1]):
    passed += 1

# 2. transformers 版本
total += 1
if check("transformers >= 4.51.0", lambda: (
    __import__('transformers'),
    f"v{__import__('transformers').__version__}"
)[-1]):
    passed += 1

# 3. TRL
total += 1
if check("TRL", lambda: f"v{__import__('trl').__version__}"):
    passed += 1

# 4. PEFT
total += 1
if check("PEFT", lambda: f"v{__import__('peft').__version__}"):
    passed += 1

# 5. bitsandbytes
total += 1
if check("bitsandbytes", lambda: f"v{__import__('bitsandbytes').__version__}"):
    passed += 1

# 6. datasets
total += 1
if check("datasets", lambda: f"v{__import__('datasets').__version__}"):
    passed += 1

# 7. GPU 信息
total += 1
import torch
if torch.cuda.is_available():
    if check("GPU", lambda: (
        f"{torch.cuda.get_device_name(0)}, "
        f"{torch.cuda.get_device_properties(0).total_mem / 1024**3:.1f} GB"
    )):
        passed += 1
else:
    print("  [WARN] GPU: 不可用（CPU 模式）")

print(f"\n结果: {passed}/{total} 项通过")
if passed == total:
    print("环境配置完成，可以开始实验！")
else:
    print("请修复失败项后重试。")