Build custom_rasterizer for torch 2.8.0 cuda12.8

[DaniilVdovin]

whl was build and test on:
Win11 x64
Python 3.12.9
CUDA 12.8
PyTorch 2.8.0

[GilHacco]

The current setup works well with PyTorch 2.6 + Cuda 12.6 on an RTX 3090. Well done, and thank you.

When trying on an RTX 5090, I ran into complications. The PyTorch 2.6 + Cuda 12.6 combination does not seem to work on the 5090 (which requires sm_120 support). When trying to update to PyTorch 2.8 + Cuda 12.8, I hit a dead end (per pull request).

Are there plans to update for PyTorch 2.8 + Cuda 12.8 ?

[GilHacco]

The current setup works well with PyTorch 2.6 + Cuda 12.6 on an RTX 3090. Well done, and thank you.

When trying on an RTX 5090, I ran into complications. The PyTorch 2.6 + Cuda 12.6 combination does not seem to work on the 5090 (which requires sm_120 support). When trying to update to PyTorch 2.8 + Cuda 12.8, I hit a dead end (per pull request).

Are there plans to update for PyTorch 2.8 + Cuda 12.8 ?

Disregard, I found a solution:

• install cuda 12.8
• install pytorch 2.8.0
• install DaniilVdovin's 280 wheel: [https://github.com/DaniilVdovin/ComfyUI-Hunyuan3DWrapper/blob/9021bbc11ac8af1c219ab7a967869b57837f7903/wheels/custom_rasterizer-0.1.0%2Btorch280.cuda128-cp312-cp312-win_amd64.whl]

Thank you @DaniilVdovin !

[tritiumnox]

I tried @GilHacco's steps, installing CUDA 12.8, and pytorch 2.8.0, but I get the following error when trying to install @DaniilVdovin's .whl:

ERROR: custom_rasterizer-0.1.0+torch280.cuda128-cp312-cp312-win_amd64.whl is not a supported wheel on this platform.

Does anyone know what to try next, please?

[GilHacco]

I tried @GilHacco's steps, installing CUDA 12.8, and pytorch 2.8.0, but I get the following error when trying to install @DaniilVdovin's .whl:

ERROR: custom_rasterizer-0.1.0+torch280.cuda128-cp312-cp312-win_amd64.whl is not a supported wheel on this platform.

Does anyone know what to try next, please?

Are you using Python 3.12 ?
What OS? (I'm on Win10)
Try using the "--verbose" flag for more details.
pip install {wheel} --verbose

Might also try:
pip cache purge
pip install {wheel} --force-reinstall

[DaniilVdovin]

@dangerousmrd @GilHacco
Hi! whl was build and test on:
Win11 x64
Python 3.12.9
CUDA 12.8
PyTorch 2.8.0

[tritiumnox]

Thanks a lot for helping me sense check this, it looks though I didn't check Python version - I was running 3.10. Can confirm wheel working with this environment:

ComfyUI Desktop
Win 11
Python 3.12.11
CUDA 12.9
PyTorch 2.8.0
RTX 5090

Thanks again! @GilHacco @DaniilVdovin

[vvhitevvizard]

@DaniilVdovin, thx! could u please re-build the wheel for torch 2.9+cuda 13.0+python 3.13? :)

[DaniilVdovin]

@DaniilVdovin, thx! could u please re-build the wheel for torch 2.9+cuda 13.0+python 3.13? :)

@vvhitevvzad Unfortunately, I don't have any hardware that will work correctly on 2.9+cuda, as I understand it, this is a series of 50 cards.