repo for Blackwell GPUs (RTX 6000)

[21tesla]

https://github.com/21tesla/PXDesign/

contains an installation shell file that builds and patches the original PXDesign repo.

It will fail at the filtering stage due to an unresolved incompatibility of PXDesign with later versions of Protenix. Old APIs (rdkit and biotite) used by PXDesign, in turn, require older versions of numpy and others that are not compatible with Blackwell.

[guanjq]

Thanks for pointing this out!

At the moment, we primarily test and support the versions pinned in the official installation instructions. If you’ve identified concrete fixes or compatibility adjustments on the PXDesign side, we’d be very happy to review them as a PR — that would be the most effective way for us to evaluate and discuss what makes sense to support upstream.

Thanks again for taking a look and for the helpful feedback.

[monkollek]

I found the https://github.com/21tesla/PXDesign/ repo extremely helpful.
@21tesla thank you so much for doing this!

Just to pay it forward the script https://github.com/21tesla/PXDesign/blob/main/install_blackwell.sh

Clones the wrong branch
git clone -b dev_design https://github.com/bytedance/Protenix.git

This does not support the model: protenix_mini_tmpl_v0.5.0
which is required for template mediated folding. This is not required for the example as Protenix folds it with RSMD < 2.5 so it's not until you try something else that Protenix doesn't fold well compared to ground truth (GT) that you notice!

The correct branch/tag is listed here in the install script
https://github.com/bytedance/PXDesign/blob/main/install.sh#L244

So if you replace the line in https://github.com/21tesla/PXDesign/blob/main/install_blackwell.sh#L42
with something like

git clone --depth 1 --branch v0.5.0+pxd https://github.com/bytedance/Protenix.git

then you should be sweet! It took me bloody ages to realize and that's why I'm sharing.

Note: The patch may not work then as the code is slightly different in the position it's trying to modify.

[masamune-prog]

I am currently getting pytorch dependencies issues on the blackwell repo

Collecting nvidia-nvtx-cu12==12.8.90.* (from cuda-toolkit[cublas,cudart,cufft,cufile,cupti,curand,cusolver,cusparse,nvjitlink,nvrtc,nvtx]==12.8.1; platform_system == "Linux"->torch)
Downloading nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl (89 kB)
Collecting cuda-pathfinder~=1.1 (from cuda-bindings<13,>=12.9.4->torch)
Using cached cuda_pathfinder-1.2.2-py3-none-any.whl.metadata (3.2 kB)
Collecting numpy (from torchvision)
Using cached numpy-2.5.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl.metadata (6.6 kB)
INFO: pip is looking at multiple versions of torchvision to determine which version is compatible with other requirements. This could take a while.
Collecting setuptools<82 (from torch)
Using cached setuptools-70.2.0-py3-none-any.whl (930 kB)
ERROR: Cannot install torch and torchvision==0.27.0.dev20260407+cu128 because these package versions have conflicting dependencies.

The conflict is caused by:
The user requested torch
torchvision 0.27.0.dev20260407+cu128 depends on torch==2.12.0.dev20260407

Additionally, some packages in these conflicts have no matching distributions available for your environment:
torch

To fix this you could try to:

1. loosen the range of package versions you've specified
2. remove package versions to allow pip to attempt to solve the dependency conflict

ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/topics/dependency-resolution/#dealing-with-dependency-conflicts
(base) root@dgx-spark-tailscale:~/pxdesign_blackwell#

Anybody has faced similar issues