Scaling PyTorch Model Training With Minimal Code Changes

[@stanislawcronberg3271]

Love the straightforward video, didn't know about Fabric for quickly upgrading existing PyTorch code

[@user-wr4yl7tx3w]

Wow, great presentation.

[@SebastianRaschka]

Thanks :)

[@dinabandhub]

Great tutorial. 🎉

[@SebastianRaschka]

Thank you! 😊

[@user-wr4yl7tx3w]

this is really awesome content.

[@nguyenhuuuc2311]

Thanks so much for the tutorial. I learned a lot from you!!!! I have a question: What modifications should be made to the code fabric.setup(model, optimizer) if I use a learning rate scheduler?

[@nguyenhuuuc2311]

And just personal feedback for an awesome tutorial It would be great if you could consider including a gentle reminder that running code on multiple GPUs often requires the use of scripts rather than executing them directly in a notebook. Sorry if I missed any mention of this information already being included in the tutorial.

[@SebastianRaschka]

thanks, and great question. Since normal schedulers don't have any parameters you can use it as usual (no need to put it into fabric.setup). But using fabric.setup also doesn't hurt. I added a quick example here: github.com/rasbt/cvpr2023/blob/main/07_fabric-vit-mixed-fsdp-with-scheduler.py

[@SebastianRaschka]

@@nguyenhuuuc2311 Good point. Yeah, notebook (or interactive) environments are generally incompatible with multi-GPU training due to their multiprocessing limitations. Hah, I take it for granted these days but definitely a good time to mention that as a reminder!

[@nguyenhuuuc2311]

@@SebastianRaschka Thanks for spending time on my question and the quick answer with a notebook ❤

[@hamzawi2752]

Thank you so much. Impressive presentation! Do you think it is worth learning lightning, I am a PhD student and I am comfortable with Pytorch. Does lightning have all capabilities like Pytorch? I know that lightening to Pytorch like keras to Tensorflow

[@SebastianRaschka]

Good question @hamzawi2752. Fabric (covered in this video) is basically an add-on to PyTorch. It's basically useful for tapping into more advanced features like multi-GPU training, mixed-precision training etc. with minimal code changes. It's essentially a wrapper around PyTorch features, but doing this in pure PyTorch is definitely more work. So, I'd say it's worth it.