Skip to content

Cannot run on RTX 6000 / Blackwell (sm_120): pinned CUDA 12.6 + JAX 0.4.29 stack is incompatible #54

Description

@sbonner0

I wanted to report a similar issue to #36

The current environment cannot run on Blackwell GPUs (e.g. RTX PRO 6000, sm_120). Blackwell requires CUDA 12.8+, but the project pins a CUDA 12.6 PyTorch/JAX stack:

  • torch==2.7.0+cu126
  • PyG wheels built against cu126
  • jax==jaxlib==0.4.29

These builds do not include sm_120 kernels, so they fail on Blackwell.

Upgrading PyTorch to a CUDA 12.8 build is straightforward, but upgrading JAX to a Blackwell-compatible release exposes a second issue: ColabDesign still uses deprecated JAX APIs (jax.tree_map, jax.tree_flatten, etc.), which were removed in newer JAX versions. As a result, AF2 evaluation/reward paths fail at runtime with AttributeError/TypeError, while pure PyTorch workflows continue to work.

The repository already suppresses the associated deprecation warnings, but trying to use a newer JAX version results in errors.

I did have a quick go at a JAX upgrade, however found several additional issues that lead me to believe that it won't be super simple to do.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions