torch icon indicating copy to clipboard operation
torch copied to clipboard

Published 20 hours ago • verified publisher icon mlverse

Non-cuda GPU support

Open cregouby opened this issue 4 years ago • 8 comments

Hello,

Is there a chance that lantern get compiled and provide support for running on non-nvidia GPU cards through PlaidML & OpenCL 1.2 or any alternative ?

cregouby avatar Jan 26 '21 21:01 cregouby

This is not possible because LibTorch does not support other backends, although they seem to be working on a AMD Rocm device and support for new M1 macs.

I think this comment is a good reference in this subject: https://github.com/pytorch/pytorch/issues/47702#issuecomment-737682371

dfalbel avatar Jan 28 '21 22:01 dfalbel

Ok, thanks for this refinement.

cregouby avatar Jan 29 '21 17:01 cregouby

This is a reality now, so hopefully pulling in version 1.8 will open new doors.

jaredlander avatar Mar 26 '21 03:03 jaredlander

Thanks @jaredlander for this callback

cregouby avatar Apr 03 '21 16:04 cregouby

I am starting to update to libtorch 1.8 here: https://github.com/mlverse/torch/pull/513 and we will probably be able to use ROCm, however note that currently there are no binaries for LibTorch with ROCm support available for download in Pytorch's website, so users will probably need to build liobtorch from source..

dfalbel avatar Apr 03 '21 16:04 dfalbel

What if we setup GitHub actions to build a bunch of different binaries?

jaredlander avatar Apr 03 '21 17:04 jaredlander

Yeah, we could do that! I have a workflow here to build LibTorch https://github.com/dfalbel/declarations/blob/master/.github/workflows/main.yaml

We would need to figure out how to build it with ROCm support. Probably following something like: https://lernapparat.de/pytorch-rocm/

dfalbel avatar Apr 03 '21 17:04 dfalbel

Tweeted by PyTorch:

We’re excited to announce support for GPU-accelerated PyTorch training on Mac! Now you can take advantage of Apple silicon GPUs to perform ML workflows like prototyping and fine-tuning. Learn more: https://pytorch.org/blog/introducing-accelerated-pytorch-training-on-mac/

Some very impressive speedups shared in that post! I wonder if {torch} will benefit.

bearloga avatar May 18 '22 16:05 bearloga

Support for M1 GPU's is now done :)

I don't plan adding support for AMD GPU's as I don't have hardware to try it. Happy to help if someone is interested in adding it though.

dfalbel avatar Oct 14 '22 18:10 dfalbel

I'm your man for AMD GPU ! I'll open a dedicated issue for that !

cregouby avatar Oct 15 '22 12:10 cregouby