Skip to content

Cannot install on RTX 5070 (Blackwell sm_120) - torch-points3d / Minkowski Engine incompatibility on Windows WSL2 and Kaggle #38

Description

@SyedTuhinAli

Environment:

Laptop: Windows 11, WSL2 Ubuntu 24.04, RTX 5070 (Blackwell, sm_120), CUDA 12.8, Python 3.12
Also attempted: Kaggle (T4 GPU environment)

Problem with the laptop:
The RTX 5070 uses the Blackwell architecture (compute capability sm_120), which is not yet supported by the PyTorch builds that torch-points3d and Minkowski Engine require. Installation of the Docker image pulls correctly, but inference fails at the CUDA kernel compilation step. Minkowski Engine in particular, does not compile against sm_120.
Problem on Kaggle:
Kaggle's pre-built PyTorch environment conflicts with the torch-points3d dependency versions. Installing torch-points3d from source on the Kaggle T4 instance fails due to version mismatches in torch-scatter and torch-sparse.
Questions:
Is there a known working PyTorch + CUDA version combination for the Docker image that avoids the Minkowski Engine compilation issue?
Is there a plan to update the dependencies to support sm_120 (Blackwell) GPUs?
Has anyone successfully run SegmentAnyTree on Kaggle, and if so, which environment/version was used?

I am working with MLS dataset (urban) and am trying to use SegmentAnyTree for individual tree instance segmentation.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions