perf: cap torch threads per camera worker

Default TORCH_NUM_THREADS=1 so 14-20 workers do not each spawn a
28-thread BLAS pool.

audio_cleanup.py

@@ -27,6 +27,7 @@ The preprocessor is created once per worker process and reused.
 from __future__ import annotations
 
 import logging
+import os
 import time
 
 import numpy as np
@@ -36,6 +37,20 @@ log = logging.getLogger("audio.cleanup")
 SAMPLE_RATE = 16000
 
 
+def apply_torch_thread_limits(n: int = 1) -> None:
+    """Cap BLAS/Torch pools so N camera workers do not oversubscribe."""
+    n = max(1, int(n))
+    os.environ.setdefault("OMP_NUM_THREADS", str(n))
+    os.environ.setdefault("MKL_NUM_THREADS", str(n))
+    os.environ.setdefault("OPENBLAS_NUM_THREADS", str(n))
+    try:
+        import torch
+        torch.set_num_threads(n)
+        torch.set_num_interop_threads(1)
+    except Exception:  # noqa: BLE001
+        pass
+
+
 class AudioCleaner:
     """Stateful audio cleaner. Call .process(...) -> int16 mono.
 
@@ -48,7 +63,8 @@ class AudioCleaner:
                  sample_rate: int = SAMPLE_RATE,
                  chunk_s: float = 1.0,
                  hop_s: float = 0.2,
-                 beamform: str = "auto"):
+                 beamform: str = "auto",
+                 torch_num_threads: int = 1):
         self.mode = mode
         self.beamform_mode = beamform
         self.sample_rate = sample_rate
@@ -61,6 +77,7 @@ class AudioCleaner:
         self._torch = None
         self.backend = "off"
         self.last_dt_ms = -1.0
+        self.torch_num_threads = max(1, int(torch_num_threads))
 
         self._load_torch()
         if beamform == "force":
@@ -92,6 +109,7 @@ class AudioCleaner:
         self.backend = "+".join(parts) if parts else "off"
 
     def _load_torch(self) -> None:
+        apply_torch_thread_limits(self.torch_num_threads)
         try:
             import torch
             self._torch = torch