Prebuilt CUDA wheels consumed by the comfyui-runpod images.
This repo holds no source and no code. It exists only so those images can
ADD --checksum a wheel from a release asset: comfyui-runpod is private, and a
private repo's release assets need an Authorization header that a docker build
does not have. Every wheel here is a build of public, permissively licensed
source, so nothing is exposed by hosting it publicly.
Wheels are never committed to git — they are release assets only, so cloning this repo stays cheap.
Tag grammar: <package>-<short source commit>-<cuNNN>-torch<version>.
| Tag | Package | Links against |
|---|---|---|
sage-40d3586-cu130-torch2.13.0 |
SageAttention 2.2.0 | torch 2.13.0 + cu130, cp310, linux_x86_64 |
Every release's notes must record enough to rebuild it exactly:
- the asset's sha256
- the full upstream source commit, and the repo it came from
- the torch version and index URL it links, and the python tag
- the
TORCH_CUDA_ARCH_LISTactually used - which arches were verified how (cubins present vs. a real kernel launch)
- the build machine's GPU and the date
The comfyui-runpod Dockerfiles pin these URLs by checksum. A deleted or retagged release breaks image builds. Supersede with a new tag instead.
By tools/build_sage_wheel.sh in comfyui-runpod, following the runbook at
tools/BUILD_SAGE_WHEEL.md there. Not by CI: these builds need more cores and
RAM than a GitHub-hosted runner has, and a GPU to prove the result launches.