RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the ‘spawn’ start method
| fork | spawn |
|---|
| 分叉 | 产卵 |
| 继承 | 非继承 |
| 从头构建一个子进程，父进程的数据拷贝到子进程的空间中，拥有自己的Python解释器 |
| 变量具有与父进程中相同的id | |
| N | 在每个子进程开始时导入模块 |
| 子进程获取在 name == main 块中定义的变量 | N |
解决方案
if __name__=='__main__':
torch.multiprocessing.set_start_method('spawn')
参考与更多
- 日志提示“RuntimeError: Cannot re-initialize CUDA in forked subprocess”
- https://britishgeologicalsurvey.github.io/science/python-forking-vs-spawn/
- https://www.pythonf.cn/read/65459
- https://stackoverflow.com/questions/64095876/multiprocessing-fork-vs-spawn
