Install seemingly impossible as of November 2025: version conflict between torchsparse and mmvc

[elisehoi]

Hello,

After multiple attempts to install TorchSparse, I found that the versions currently available on their GitHub page do not include 1.4.0, which is recommended for use with LaserMix.

I was able to successfully install TorchSparse 2.0.0 with CUDA 11.8 (it does not work with CUDA 11.3, which is the version currently advised). However, LaserMix also requires MMCV<2.1.0 (I received the following error during training: AssertionError: MMCV==2.2.0 is used but incompatible. Please install mmcv>=2.0.0rc4, <2.1.0.). Installing this version of MMCV seems to require an older CUDA version, which conflicts with the version needed for TorchSparse.

At the moment, it appears that setting up all the dependencies required to run LaserMix is not feasible.

If possible, I would greatly appreciate any guidance or an update to resolve this dependency issue. I am very enthusiastic about LaserMix, as I am currently writing a review of machine learning frameworks for classifying automobile LiDAR data (due February 2026).

Kind regards

[L-Reichardt]

torchsparse v1.4.0

To my knowledge, MMDet3D also supports spconv or minkowskiengine as backends. You could try those, but probably some settings have to be changed.

[elisehoi]

torchsparse v1.4.0

To my knowledge, MMDet3D also supports spconv or minkowskiengine as backends. You could try those, but probably some settings have to be changed.

Thank you for your feedback ! I believe I meant pre-built version for torchsparse 1.4.

I had issues with building the 1.4 repository, but managed to do it by downgrading my system cuda to 11.3, with pytorch 1.12.1. However, am still facing issues with the lasermix installation because of mismatches between my nvidia driver (535) and cuda 11.3. Older drivers do not seem to be available with my setup (ubuntu 24.04) .

If anyone has a successful experience running lasermix on a newer system, I would greatly appreciate to know about it