Blackwell support

[sistar2020]

Blackwell GPU is supported by cuda version >= 12.8.
OpenDDE currently supports cuda 12.6.
Could you upgrade OpenDDE to use cuda 12.8?

[mrzzmrzz]

We are currently adapting to the GPU on the Blackwell architecture and building the new Docker image. I estimate that the new compatible image will be released in 1-2 days. Thank you for your patience.

[WillHua127]

We are working on it, plan to get it done ASAP

[GriffinYoung]

Was this completed?

[WillHua127]

we are getting close, running internal tests with NVIDIA team

[mrzzmrzz]

@sistar2020, @GriffinYoung

Thank you for your patience. OpenDDE now supports inference on NVIDIA Blackwell GPUs through a Python package installation using the PyTorch CUDA 12.8 backend.

Please install OpenDDE with:

uv venv --python 3.11
uv pip install --python .venv --torch-backend cu128 opendde

At present, Blackwell inference uses the standard PyTorch triangle kernels. Please explicitly select them when running inference:

--triatt_kernel torch --trimul_kernel torch

cuEquivariance (cueq) acceleration is not yet available for OpenDDE on Blackwell GPUs. We are following upstream development closely and will enable cueq acceleration after the required Blackwell support is complete and has passed our validation.

If you encounter any installation or inference issues, please share your GPU model, NVIDIA driver version, the output of opendde doctor, the command you ran, and the complete error log.

[GriffinYoung]

Ok will do. Thanks a lot!