How to support customized extensions on MPS?

Hi, I have my customized CPU and CUDA kernel, but I want to run them on my MacBook Pro GPU. I wonder if there are some tutorials to write the customized kernel on MPS backend, especially how to load the customized op in PyTorch?

Hi,
I don’t think we have a tutorial for that yet no.
You should be able to write custom Metal kernels though within PyTorch.
I’m not sure if the cpp_extension module works fine for these though. You should feel free to try it out and open issues on github if you encounter any problem with it!

Hello, did you find a solution ?