Background

In PR #319 (fix/whisper-vulkan-linux), transcribeOpts was updated to set maxThreads: cpus().length so Whisper utilises all available logical CPU cores on the host.

Problem

cpus().length (Node's os.cpus()) reflects the host's logical CPU count and is unaware of Linux cgroup CPU quotas. In containerised environments (Docker, Kubernetes, LXC, etc.) the container may be limited to a fractional number of CPUs via cpu.cfs_quota_us / cpu.cfs_period_us (cgroups v1) or cpu.max (cgroups v2), yet cpus().length will still return the full host count. Spawning more threads than the container is allotted leads to CPU throttling and can degrade transcription throughput.

Proposed fix

Read the effective CPU limit from the cgroup interface files at runtime and clamp maxThreads to Math.max(1, Math.floor(cgroupQuota)). Fall back to cpus().length if the cgroup files are absent (bare-metal, macOS, Windows) or if the quota is set to -1 (unlimited).

Rough sketch:

function effectiveCpuCount(): number {
  // cgroups v2
  try {
    const raw = fs.readFileSync('/sys/fs/cgroup/cpu.max', 'utf8').trim();
    const [quota, period] = raw.split(' ');
    if (quota !== 'max') return Math.max(1, Math.floor(Number(quota) / Number(period)));
  } catch { /* not v2 */ }
  // cgroups v1
  try {
    const quota  = Number(fs.readFileSync('/sys/fs/cgroup/cpu/cpu.cfs_quota_us',  'utf8'));
    const period = Number(fs.readFileSync('/sys/fs/cgroup/cpu/cpu.cfs_period_us', 'utf8'));
    if (quota > 0) return Math.max(1, Math.floor(quota / period));
  } catch { /* not v1 */ }
  return cpus().length;
}

Context

• Raised as a P2 follow-up during review of PR fix(whisper): enable Vulkan GPU backend on Linux #319 by @kimbotai1337.
• The hasVulkanBackend() and useGpu fixes shipped in that PR; this is the only remaining item.

References

• PR fix(whisper): enable Vulkan GPU backend on Linux #319: fix(whisper): enable Vulkan GPU backend on Linux #319