5xxx Series support via CU128

[mspacemedia]

This build is missing support for CU128 - attempting with limited knowledge to change results in many requirements breaking.

[YanWenKun]

It's a nigthmare to compile those libs under cu128. Even the upstream struggles to build the pt27+cu128 binaries on Windows.

For now, I can only advise installing CUDA Toolkit 12.8.0 to 12.9.1 (required for 50 series GPUs) and trying to run the package without upgrading anything.

If this doesn't work, I strongly recommend stopping further attempts. Either waiting for the AI community to catch up, or simply getting an RTX 40 series GPU.

[archay0]

Hey, @YanWenKun pytorch now has support for sm_120 blackwell architecture gpus.
The conflict im facing is just with nvdiffrast, Will it be possible for an update to the portable pack?

[YanWenKun]

@archay0 It's near (finally).

PyTorch 2.8.0 + xFormers 0.0.32.post2 is buggy on Blackwell.

PyTorch 2.9.0 + xFormers 0.0.33.post1 works, but PyTorch 2.9.1 (fixed memory issues) is just released.

I'm waiting for the xFormers to release a new version to pair with PyTorch 2.9.1. Then I'll compile the dependency libs for 3D-Pack, probably under CUDA 13.0 or 12.8, and Python 3.12.

If it takes too long, I may use self-built binary of xFormers on PyTorch 2.9.1 (not ideal since issues are not really fixed).

xFormers was used by 3D-Pack in few ways. I believe the performance gains are barely noticeable nowadays, but it still provides compatibility for some legacy code in 3D-Pack.

[Kamranahm]

@YanWenKun Any updates on this? I've been eagerly waiting for a while now hoping for a 5090 compatible version =*(.