#!/bin/bash
# 启动训练（GPU 模式下跑）
set -e
WORK=/root/autodl-tmp/hv
[ -f "$WORK/train.yaml" ] || { echo "先跑 hv-setup.sh"; exit 1; }
LF=$(cat "$WORK/lf_dir.txt" 2>/dev/null || python3 -c "import os,llamafactory;print(os.path.dirname(os.path.dirname(os.path.abspath(llamafactory.__file__))))")

echo "== 环境自检 =="
nvidia-smi --query-gpu=name,memory.total --format=csv || { echo "没看到 GPU：确认已切到 GPU 模式（不是无卡模式）"; exit 1; }
python3 -c "import torch,transformers;print('torch',torch.__version__,'| transformers',transformers.__version__,'| cuda',torch.cuda.is_available())"

echo "== 开始训练 =="
cd "$LF"
llamafactory-cli train "$WORK/train.yaml"

echo
echo "== 训练结束，打包 LoRA adapter =="
OUT=/root/autodl-tmp/output/qwen35-9b-claude-tone
ls -la "$OUT" || true
cd /root/autodl-tmp/output && zip -qr qwen35-9b-claude-tone-lora.zip "$(basename $OUT)" && ls -lh /root/autodl-tmp/output/qwen35-9b-claude-tone-lora.zip
echo "下载：JupyterLab 左侧文件树 → autodl-tmp/output → 右键 Download"