You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Nov 21, 2026. It is now read-only.
Repository navigation
This repository was archived by the owner on Nov 21, 2026. It is now read-only.
CI: examples send every package lookup to download.pytorch.org, so a PyTorch index outage fails unrelated jobs #401
The examples workflow (.github/workflows/test-python-examples.yml) sets UV_INDEX=https://download.pytorch.org/whl/cpu with UV_INDEX_STRATEGY=unsafe-best-match so that torch gets its CPU build. With that strategy uv asks the PyTorch index about every package it resolves, so a PyTorch index outage fails jobs that never needed torch. Two jobs failed that way on PR #391 and PR #395 with HTTP status server error (503 Service Unavailable) for url (https://download.pytorch.org/whl/cpu/rich/) and the same for .../h11/, neither of which is a torch package.
Proposed fix: write a small throwaway uv project in $RUNNER_TEMP that lists the example packages and pins only torch to the PyTorch CPU index ([tool.uv.sources] plus an index with explicit = true), and run the examples with uv run --project <dir> --with <wheel>. --no-project cannot do this because uv reads [tool.uv.sources] only from a project, and a requirements file cannot mark an index explicit. Measured locally on the same requirements: the resolved versions are identical (67 packages, no differences), the PyTorch index is asked about 1 package (torch) instead of 66, and numpy, rich, h11 and requests come from PyPI. I also raise UV_HTTP_RETRIES above its default of 3 for short outages, and update bindings/python/docs/development/ci-setup.md.
Activity
ci: pin only torch to the PyTorch index in the examples job