From 772dc32ca451c342b9359533b57bbf9304ce990f Mon Sep 17 00:00:00 2001 From: ncaq Date: Tue, 10 Nov 2026 23:39:05 +0900 Subject: [PATCH] ci: use default Nix parallelism in the Python check jobs Drop `--max-jobs 1 --cores 2` from `python-runtime-deps` and `extra-python-packages`. The cap was introduced in 909f18e for the CUDA Docker image build, where it works around OOM. It was later copied to these check jobs without a recorded OOM there. Hosted runners have 4 cores, and the cap makes the source builds of torch and onnxruntime the bottleneck. In run 42780073869 the CPU `extra-python-packages` job hit the 6 hour job limit, with torch alone taking 154 minutes at `-j2`. The uncapped `lint` job already builds the same torch and onnxruntime concurrently at `-j4` each. It finished in 4h49m without OOM in run 42500297518 on main. The cap on the CUDA Docker image build is kept. --- .github/workflows/build.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index d7fa813..a7d7f00 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -112,7 +112,7 @@ jobs: run: | nix build ".#checks.x86_64-linux.${{ matrix.check }}" \ --no-link \ - --print-build-logs --max-jobs 1 --cores 2 + --print-build-logs # Realize the extended runtime for every backend so PRs cannot replace the # backend-specific Torch/NumPy stack or ship dependencies that fail to import. @@ -163,7 +163,7 @@ jobs: - name: Build ${{ matrix.check }} run: | nix build .#checks.x86_64-linux.${{ matrix.check }} \ - --print-build-logs --max-jobs 1 --cores 2 + --print-build-logs # Full nix flake check (includes package build) — only on main/tags #