[davidwalter2]

all-unit-tests runs the 41 collected test files one after the other, and every fit job (fitting, sparse-fits, alternative-fits, regularization, bsm, likelihoodscans, abcd-fits) waits for it. In #208's last CI run it took 56 of the 63.5 min; everything after it finished within 7 min.

The files now run 12 at a time:

• Still one process per file. XLA keeps compiled code mapped until the process exits, and the suite in one process exceeds vm.max_map_count. That is also why this uses xargs -P and not pytest-xdist, whose workers live across files.
• Each process is capped at 4 threads (OMP_NUM_THREADS, OPENBLAS_NUM_THREADS, MKL_NUM_THREADS, TF_NUM_INTRAOP_THREADS; 2 inter-op threads). Uncapped, every TensorFlow process would start one thread per core of the 256-core runner host, which the WRemnants CI shares. 12 × 4 threads is 48 cores. rabbit sets no thread counts itself, so the caps also reach the rabbit_fit.py subprocesses some tests start.
• Logs. Each file's output goes to its own log and is printed in one group after all files have run, as before. A line per file as it ends shows progress meanwhile.
• Timeout. A file still running after 30 min is killed and counts as failed (exit 124), instead of holding the job for the default 6 h.
• Unchanged: collection over tests/ decides what runs (test_ci_registration passes on the new workflow), every file runs even if another fails, and the job fails if any did.

Running files in parallel is safe: the tests write into their own temporary directories, and the one that uses $RABBIT_OUTDIR (test_global_asym_impacts.py) writes to its own subdirectory.

Checked on a 64-core submit node, with the step's script and env taken from this workflow and run under bash -e:

• the whole suite passes in 412 s;
• a toy suite with a passing, a failing and a timed-out file reports exit 0, 1 and 124, lists the two failures, and fails the step.

Expected on the runner host: per file it ran ~1.7× slower than the submit node, which puts all-unit-tests at ~8–10 min and the whole CI at ~15–17 min instead of ~63. The floor is the slowest file, test_mcstat_kfold.py (7.6 min in CI). This PR's own CI run shows the real timing.

Not changed: the fit jobs still wait for all-unit-tests. They only need make-tensor's files, so dropping that wait would overlap them with the unit tests (CI ~10 min), at the cost of running fits for PRs whose unit tests fail.

🤖 Generated with Claude Code

https://claude.ai/code/session_014i9SdYoFtmw6ig5a4kzAx2

[github-actions]

Issues Found

No issues found.

Summary

Total issues: 0 critical, 0 important, 0 minor.

Overall Verdict

Status: Patch is correct

Explanation: The xargs-based parallelization is mechanically sound: filenames are passed safely as single argv entries, per-file exit codes are captured in .rc files and correctly aggregated (including the missing-file and timeout-exit-124 cases), and the collect-only line still satisfies the repo's workflow-guard tests in test_ci_registration.py (whole-directory coverage token present, no narrowing flags). Log-name flattening is injective because tests/ is flat, the glob-empty edge case still ends in job failure via failed="collection", and all fallible commands are guarded against GitHub's bash -e shell. Thread caps, per-file timeout, and deferred grouped output are documented, intentional design choices.

Review details

• Model: cern/Kimi-K3-FLATIRON

Review metrics: 4 turns · 6 tool calls · 14m 48s

• Tokens: 71.7K input · 41.2K cached (58%) · 23.6K output