Newbie - Block size limitation pytorch - Intel ARC a770 16GB

[kodiconnect]

First, thank you for building this!

I'm getting this error: "Current platform can NOT allocate memory block with size larger than 4GB! Tried to allocate 7.03 GiB (GPU 0; 15.91 GiB total capacity; 2.29 GiB already allocated; 2.50 GiB reserved in total by PyTorch)" I suspect it can be resolved with something like: "export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512"

Has anyone run into this? If so, what was your solution other than accepting the limit?

[kodiconnect]

Well, ignore me as I answered my own question in the question... Setting this as an env variable in the docker file corrected the issue:

Set environment variable inside the container

ENV PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512

[simonlui]

That won't solve the core issue, I refer you to intel/intel-extension-for-pytorch#325 where the issue is still being discussed. If you have Battlemage and above, it will work without hassle. But on Alchemist and similar 1st generation Xe architecture, there are changes that make it work technically but practically, for ComfyUI and video/image generation purposes, there are still issues. No real ETA for when they fix it and if they do, it is going to land inside a Pytorch nightly build first before an IPEX build.

[kodiconnect]

Thank you sir! It looks like the core issue may be resolved in the next release 🤞 I’ll follow that thread for updates.

…

On Sun, Jan 26, 2025 at 14:41 Simon Lui ***@***.***> wrote: That won't solve the core issue, I refer you to intel/intel-extension-for-pytorch#325 <intel/intel-extension-for-pytorch#325> where the issue is still being discussed. If you have Battlemage and above, it will work without hassle. But on Alchemist and similar 1st generation Xe architecture, there are changes that make it work technically but practically, for ComfyUI and video/image generation purposes, there are still issues. No real ETA for when they fix it and if they do, it is going to land inside a Pytorch nightly build first before an IPEX build. — Reply to this email directly, view it on GitHub <#11 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AGGTP5VPKRWAG3HRA3VO5732MVJCBAVCNFSM6AAAAABV42NLE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMMJUGYYDEMRVGA> . You are receiving this because you authored the thread.Message ID: ***@***.***>