Pre-built CUDA Python wheels for packages used by ComfyUI custom nodes. Spec-driven build pipeline consuming packages/*.yml configurations.
| Package | Source | Pinned Ref | Custom Nodes |
|---|---|---|---|
| cc_torch | ronghanghu/cc_torch | 446e936 |
pollockjj/ComfyUI-SAM3 |
| flash_attn | Dao-AILab/flash-attention | v2.8.3 |
pollockjj/ComfyUI-SAM3 |
| sageattention | thu-ml/SageAttention | v2.2.0 |
|
| sageattn3 | thu-ml/SageAttention | v2.2.0 |
|
| torch_generic_nms | ronghanghu/torch_generic_nms | fcf02b4 |
pollockjj/ComfyUI-SAM3 |
- Ensure a package spec exists in
packages/<name>.yml - Run:
gh workflow run build.yml --repo pollockjj/wheels -f package=<name> - Download the wheel from the GitHub release
Wheels are served via PEP 503 index at the GitHub Pages URL for this repository.
Create packages/<name>.yml with required fields: name, source_repo, version, build_matrix. See existing specs for format.
See ATTRIBUTION.md for upstream source credits.