Skip to content

Blackwell GPUs are incompatible with the current CUDA 12.6 / PyTorch stack #44

@pablovela5620

Description

@pablovela5620

Wanted to follow up on #18 as I saw it was closed. I wanted to try again on my 5090 (and all other blackwell gpus) a minimum of cuda 12.8 is required to run.

This mostly affects the following baselines

  • dpvo
  • droidslam
  • mast3rslam
  • monogs
  • vggt
  • vggtslam

in particular because they require lietorch/torch-scatter/ and other precompiled conda packages.

I tried out orbslam2/3 and pycuvslam and those seemed to work on the 5090. All others froze.

Along with this it seems that they dont fail in an obvious way but rather get stuck and an error message never surfaces which makes it hard to understand whats going on

So I think the fix is bumping 12.8 and recompiling all the underlying packages with 12.8 support (or 12.9? it looks like the conda linux package only has 12.9 when I was looking https://prefix.dev/channels/conda-forge/packages/pytorch)
@Tobias-Fischer

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