pixal3d-cuda-wheels

Python wheels for the three CUDA extensions the Pixal3D Space uses, compiled for RTX 3090 (sm_86), RTX 4090 (sm_89) and Blackwell (sm_120). The wheels the Space used before carry sm_120 kernels only and fail with "no kernel image is available" on 30- and 40-series cards.

Wheel	Source (MIT)	Commit
flex_gemm 1.0.0	JeffreyXiang/FlexGEMM	6dd94a859c26ee8246888502eada3dd8ad85532e
cumesh 0.0.1	JeffreyXiang/CuMesh	12289e1062f0603f2f0d0771b02e1395d247f26f
o_voxel 0.0.1	microsoft/TRELLIS.2 o-voxel/	75fbf0183001ed9876c8dbb35de6b68552ee08bd

Build: CUDA 13.0.3 devel (Ubuntu 24.04, image digest sha256:b7ae301dea2c162444795462ce17a05f6a516e5a75944b57af5b88540a1a2266), Python 3.12, torch 2.11.0+cu130, TORCH_CUDA_ARCH_LIST="8.6;8.9;12.0". Recipe: V-Sekai-fire/pixal3d-cuda-wheels.

Each commit to this repository is one build; SHA256SUMS lists the wheels and BUILD.txt records where they were built. Pin a commit in requirements.txt:

https://huggingface.co/ifire/pixal3d-cuda-wheels/resolve/<commit>/flex_gemm-1.0.0-cp312-cp312-linux_x86_64.whl