Feature: Automatically clear the JIT cache if the cache fails to load

[woct0rdho]

Please help if you have some idea on how to do this.

A common cause of ImportError: DLL load failed while importing cuda_utils is that the user tries to load a cache compiled with a different version of Python/Triton/C compiler/CUDA. Although in the latest version of triton-windows I'm trying keep the ABI compatibility, I still need detailed error reports from more people to understand this issue.

In the meanwhile, we can just clear the caches and redo the compilation when the error happens. Both Triton's cache and TorchInductor's cache need to be cleared:

C:\Users\<your username>\.triton\cache\
C:\Users\<your username>\AppData\Local\Temp\torchinductor_<your username>\

The most naive idea is to just delete these two folders and redo the compilation when the error is caught here. However, it looks strange if Triton manages TorchInductor's cache (TorchInductor depends on Triton, not vice-versa), and I haven't checked how TorchInductor calls Triton's compilation API and manages the cache, so we need a good idea on how properly do this.

The re-compilation should be done only once. If the cache still fails to load, then it's a problem somewhere else.

For app developers like ComfyUI: You may also consider clearing the caches when the app starts.