Multi GPU for Style Transfer

Hi I’m keen to test out Multi-GPU scaling for style transfer as per the docs (basically at the bottom of the page) - https://github.com/ProGamerGov/neural-style-pt It looks like there is a machine with 2x A100s on gradient. Would this be the best option to go with using a PyTorch1.10 runtime?

I’ve been having a reasonable amount of success with the same git using google colab, but run into issues when the resolution gets to a size that their single GPU can’t process. I’m guessing the above is the most similar to what google are using?

Hey @jamahun First off, welcome :hugs:

Gradient runs a similar setup to Colab which is a docker image with the major frameworks, drivers, etc. preinstalled. The environments are not identical – you may need to update some packages – but it shouldn’t be difficult to get going.

FYI the linked repo is a bit out of date (I see most of the commits are from about 2 years ago) so you may need to downgrade the PyTorch version. Or you can pull an older version e.g. from Dockerhub using the Advanced option of the notebook create page.

The A100x2 is a great option but any multi-GPU option should work.

@nmb may be able to test/assist.

1 Like

Amazing thanks so much for the quick response. Is there a way I can test it with the free GPU version to get everything working and then just run the A100x2 when I have it working on the free GPU?

Yeah absolutely. You can toggle the instance anytime after you’ve run your tests on the free GPU. Just click this dropdown.

1 Like

Hi @jamahun,

Daniel tagged me on this as I have been using our multi-GPUs on Gradient. Let me know if you have any questions on this thread.

Nick (nmb)