CI: identify hanging tests in macOS unittest job

Summary:
The macOS unittest job has been hitting its timeout for several runs in a
row with the progress bar frozen partway through pytest. Add
`pytest-timeout` so a stuck test fails with a nodeid and per-thread stack
trace, and set `faulthandler_timeout=180` so every xdist worker dumps its
threads every 3 minutes while tests are still running -- this surfaces the
hung test as it develops, not just at termination.

Diagnostic runs identified the hang fingerprint: tests stuck inside
`torch._inductor.package._package.__call__` (line 736), the AOTI-packaged
`.so` invocation. The same stack was observed for tests in
`examples/models/llama3_2_vision/{preprocess,vision_encoder}` and the
local M1 reruns of these tests against the CI-pinned torch wheel all
pass quickly, so the hang is CI-environment-specific.

Experiment: drop the macOS pytest run to `-n 1` to test the hypothesis
that the hang is caused by xdist-worker contention (parallel clang/ld
during AOTI compile, dlopen lock on darwin, libomp oversubscription)
rather than a true deadlock. Bump the job timeout to 90 minutes so a
serial run has room to finish, and keep the per-test `--timeout=1500`.
If the previously-hung AOTI tests now pass under `-n 1`, the right
permanent fix is splitting macOS into a serial AOTI lane plus a parallel
everything-else lane (or simply dropping the parallelism cap).

Co-Authored-By: Claude <noreply@anthropic.com>

Author: SS-JIA

.ci/docker/requirements-ci.txt

@@ -11,6 +11,7 @@ zstd==1.5.5.1
 pandas>=2.2.2; python_version >= '3.10'
 pytest==7.2.0
 pytest-cov==4.1.0
+pytest-timeout==2.2.0
 expecttest==0.1.6
 hypothesis==6.84.2
 parameterized==0.9.0

.ci/scripts/unittest-macos-cmake.sh

@@ -12,8 +12,15 @@ set -eux
 export TORCHINDUCTOR_CACHE_DIR="$(mktemp -d "${RUNNER_TEMP:-/tmp}/torchinductor_cache_XXXXXX")"
 trap 'rm -rf "${TORCHINDUCTOR_CACHE_DIR}"' EXIT
 
-# Run pytest with coverage
-${CONDA_RUN} pytest -n auto --cov=./ --cov-report=xml
+# EXPERIMENT: drop xdist (`-n 1`) on macOS to test whether AOTI hangs are
+# caused by parallel-worker contention (clang/ld, dlopen lock, libomp
+# oversubscription) rather than a true deadlock. AOTI skips removed so we
+# can observe whether the previously-hung tests now pass serially.
+# --timeout surfaces hung tests with a thread dump and faulthandler_timeout
+# periodically dumps every worker's threads while tests are still running.
+${CONDA_RUN} pytest -n 1 --cov=./ --cov-report=xml \
+  --timeout=1500 --timeout-method=thread \
+  -o faulthandler_timeout=180
 # Run gtest
 LLVM_PROFDATA="xcrun llvm-profdata" LLVM_COV="xcrun llvm-cov" \
 ${CONDA_RUN} test/run_oss_cpp_tests.sh

.github/workflows/_unittest.yml

@@ -49,6 +49,7 @@ jobs:
       python-version: '3.11'
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
       script: |
         set -eux
         # This is needed to get the prebuilt PyTorch wheel from S3