Initial support for Metal Performance Shaders (MPS) #98
Conversation
|
Thanks a lot @dameikle ! Quite curious to see what we can achieve on Apple silicon chips. |
|
@dameikle, I managed to run this fine on an M3 Macbook Air 👌 A few questions:
|
|
Thanks for taking a look @francoishernandez.
I normally use Virtual Envs created from Homebrew installed Python to avoid clashes with the macOS installed version. I can put together some documentation.
I was mainly doing things by hand as well as using others models I'd trained elsewhere, so wasn't trying them out to be honest. Can have a wee look at them and see if we can make them work across both.
Sounds like a plan. Will pull information together. |
…PS to avoid users having to use PYTORCH_ENABLE_MPS_FALLBACK
|
Thanks for the updates @dameikle! |
|
@francoishernandez - I think it's good to merge now. I was starting to work on the docs and taking a look at the recipies last night but like you say, can do that in seperate PR. |
Makes adjustments to provide initial support for Apple Silicon Metal Performance Shaders (MPS) to provide a base for further optimisation.
This has been tested on basic training and inference on both M1 and M3 Macs. It has also been used for inference of larger production grade models that have been trained on CUDA.