Cap pytest-xdist's auto worker count to the container CPU limit

`auto` prefers psutil.cpu_count(), which reports the machine's cores and sees
neither a cpuset nor a CFS quota, so on an 8 vCPU OSDC pod it created 64 workers
and got test-arm-backend-no-driver OOMKilled.

pytest-parallelism.sh exports PYTEST_XDIST_AUTO_NUM_WORKERS, which xdist checks
ahead of psutil, set to the smaller of the cgroup CPU quota and nproc. nproc is
what pytorch relies on for OMP_NUM_THREADS on the same fleet
(.ci/pytorch/test.sh, USE_ARC block), whose comment says ARC pods carry a real
cpuset; the quota covers the case where the limit is expressed that way
instead, as pytorch documents for their nested-container ROCm fleet. Taking the
smaller needs no assumption about which is in play, and both fall back to the
machine count off OSDC, so this is a no-op elsewhere.

Sourced from the three scripts that run pytest with `auto`. pytorch avoids the
problem differently, by never using `auto` -- NUM_PROCS in
tools/testing/test_selections.py is a hardcoded 1-3.

Authored with Claude Code.

Author: huydhn

.ci/scripts/pytest-parallelism.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Pin pytest-xdist's `auto` worker count. `auto` prefers psutil.cpu_count(),
+# which reports the machine's cores and sees neither a cpuset nor a CFS quota,
+# so on a container it fans out far past the CPU limit and the workers exhaust
+# the memory limit. Take the smaller of the cgroup quota and nproc, so it holds
+# whichever way the limit is expressed.
+
+pytest_num_workers() {
+  local affinity quota period
+  affinity=$(nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 1)
+
+  quota=""
+  if [[ -r /sys/fs/cgroup/cpu.max ]]; then  # cgroup v2
+    read -r quota period < /sys/fs/cgroup/cpu.max
+    [[ "${quota}" == "max" ]] && quota=""
+  elif [[ -r /sys/fs/cgroup/cpu/cpu.cfs_quota_us ]]; then  # cgroup v1
+    quota=$(cat /sys/fs/cgroup/cpu/cpu.cfs_quota_us)
+    period=$(cat /sys/fs/cgroup/cpu/cpu.cfs_period_us)
+    (( quota <= 0 )) && quota=""
+  fi
+
+  if [[ -n "${quota}" ]] && (( period > 0 )); then
+    local from_quota=$(( quota / period ))
+    (( from_quota > 0 && from_quota < affinity )) && affinity="${from_quota}"
+  fi
+
+  (( affinity < 1 )) && affinity=1
+  echo "${affinity}"
+}
+
+if [[ -z "${PYTEST_XDIST_AUTO_NUM_WORKERS:-}" ]]; then
+  export PYTEST_XDIST_AUTO_NUM_WORKERS="$(pytest_num_workers)"
+  echo "PYTEST_XDIST_AUTO_NUM_WORKERS=${PYTEST_XDIST_AUTO_NUM_WORKERS}"
+fi

.ci/scripts/test_backend.sh

@@ -7,6 +7,9 @@
 # LICENSE file in the root directory of this source tree.
 set -eux
 
+# Cap pytest-xdist's `auto` workers to the container's CPU quota.
+source .ci/scripts/pytest-parallelism.sh
+
 SUITE=$1
 FLOW=$2
 ARTIFACT_DIR=$3

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

@@ -7,6 +7,9 @@
 # LICENSE file in the root directory of this source tree.
 set -eux
 
+# Cap pytest-xdist's `auto` workers to the container's CPU quota.
+source .ci/scripts/pytest-parallelism.sh
+
 # Some ARM/TOSA-adjacent tests import modules that require tosa_serializer.
 # Install from a local tosa-tools checkout when available. If absent in this
 # checkout layout, clone the pinned upstream tag and install from there.

backends/arm/test/test_arm_backend.sh

@@ -12,6 +12,9 @@ script_dir=$(cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
 et_root_dir=$(cd ${script_dir}/../../.. && pwd)
 cd "${et_root_dir}"
 pwd
+
+# Cap pytest-xdist's `auto` workers to the container's CPU quota.
+source .ci/scripts/pytest-parallelism.sh
 scratch_dir=${et_root_dir}/examples/arm/arm-scratch
 setup_path_script=${scratch_dir}/setup_path.sh
 _setup_msg="please refer to ${et_root_dir}/examples/arm/setup.sh to properly install necessary tools."