IntelPython / dpnp / 41471777437

83%

push

github

Speed up the Sphinx docs workflow (#3144)

This PR makes two changes to `build-sphinx.yml`, one per commit.

**Skip the docs build when a PR is closed.** Closing a PR triggers the
workflow so the `clean` job can remove the PR's docs from `gh-pages`.
But `clean` had `needs: build-and-deploy`, and that job had no
condition, so every close first ran a full library and docs build. The
last 4 PR-close runs took 41-47 min, followed by a 2 min `clean`. The
build job is now skipped on `closed`, and `clean` no longer depends on
it.

Until now, the cancel step inside the build job is what stopped an
in-progress run when its PR closed. Without a replacement, that run
could finish and republish `pull/<N>` after `clean` removed it. The PR
replaces `styfle/cancel-workflow-action` with a native `concurrency`
group per PR (per ref on `master` pushes), so the close still cancels
it, and drops the `actions: write` permission the action needed.

**Install only the NVIDIA driver library.** CuPy is imported when
generating the comparison table, and `import cupy` needs `libcuda.so.1`
even without a GPU. The conda environment already provides the CUDA
libraries CuPy uses (cudart, nvrtc, cuBLAS and others, CUDA 13.4), so
the step now installs only `libnvidia-compute-580` from the Ubuntu
restricted archive. It no longer adds the graphics-drivers PPA or
installs `libnvidia-gl-550`, the full `nvidia-cuda-toolkit` and
`clinfo`. The old step took ~164 s.

1604 of 2994 branches covered (53.57%)

Branch coverage included in aggregate %.

29438 of 34570 relevant lines covered (85.15%)

17680.32 hits per line