38 lines
1.4 KiB
JSON
38 lines
1.4 KiB
JSON
{
|
||
// VSCode 调试配置。用法：左侧"运行和调试"(Ctrl+Shift+D) → 顶部下拉选配置 → F5 启动
|
||
"version": "0.2.0",
|
||
"configurations": [
|
||
{
|
||
// 调试预训练脚本（用小参数，方便单步跟）
|
||
"name": "调试: train_pretrain (小参数)",
|
||
"type": "debugpy",
|
||
"request": "launch",
|
||
"program": "/home/qiaohui/Project/minimind/trainer/train_pretrain.py",
|
||
"python": "/home/qiaohui/Project/minimind/.venv/bin/python",
|
||
// cwd 设为 trainer 目录，这样 ../dataset/ 这类相对路径才找得到
|
||
"cwd": "/home/qiaohui/Project/minimind/trainer",
|
||
"console": "integratedTerminal",
|
||
// justMyCode=false → 允许单步进入 torch/transformers/模型内部代码
|
||
"justMyCode": false,
|
||
"args": [
|
||
"--data_path", "../dataset/pretrain_t2t_mini.jsonl",
|
||
"--epochs", "1",
|
||
"--batch_size", "2",
|
||
"--num_workers", "0", // ★关键：必须设0，否则 __getitem__ 里的断点打不上
|
||
"--log_interval", "1",
|
||
"--accumulation_steps", "1"
|
||
]
|
||
},
|
||
{
|
||
// 调试数据管线小脚本（不碰GPU，最适合先理解数据流）
|
||
"name": "调试: 当前打开的文件",
|
||
"type": "debugpy",
|
||
"request": "launch",
|
||
"program": "${file}",
|
||
"python": "/home/qiaohui/Project/minimind/.venv/bin/python",
|
||
"console": "integratedTerminal",
|
||
"justMyCode": false
|
||
}
|
||
]
|
||
}
|