Understanding torch.compile()

torch.compile
1

Hello,

Does Torch.compile() generate kernels at the Python level (quite abstract), at the C++ level (close to the metal), or using “assembly language” the current device uses (really close to the metal)?

Thanks!

2

torch.compile generally stays at python level, and usually emits OpenAI triton code for kernels, which then ultimately lowers to assembly (e.g. CUDA PTX for cuda kernels).

3

Thanks, Matt. That helps!