Pre-built Python wheels for packages that only ship as sdist upstream — built against the specific torch / CUDA / cpython combo pinned in evolutionaryscale/oss.
Some Python packages publish only source distributions (.tar.gz) because their compiled output is ABI-locked to a specific torch + CUDA + cpython combination. Building them from sdist on every developer machine and CI runner is slow (minutes per install). This repo hosts the pre-built .whl files behind a stable GitHub release URL so pixi.toml can reference them directly.
Releases are tagged as <package>/<version-tag> so the repo can host multiple package families. Each release attaches one or more .whl files. Use the asset download URL in your pixi.toml:
foo = { url = "https://github.com/evolutionaryscale/wheels/releases/download/<package>/<version-tag>/<wheel-filename>" }| Package | Latest tag | Build matrix |
|---|---|---|
transformer-engine-torch |
transformer-engine-torch/v2.13.0-pt2.8-cu128-cp312 |
torch 2.8 / CUDA 12.8 / cp312 / linux x86_64 |
- Build the wheel inside the matching pixi env so it links against the right torch/CUDA ABI.
- Compute its sha256 and pick a tag that encodes the matrix:
<package>/v<version>-pt<torch>-cu<cuda>-cp<py>. gh release create <tag> --repo evolutionaryscale/wheels --title "..." --notes "..." <path-to-wheel>.- Update the consuming repo's
pixi.tomland re-lock.