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.

.env.example

@@ -34,6 +34,8 @@ ENHANCE_HOP_S=0.2
 # auto (default): beamform if speechbrain loads, else average channels
 # force: require DelaySum_Beamformer, off: average L/R
 BEAMFORM_MODE=auto
+# One BLAS/Torch thread per worker (20 workers vs 28 cores)
+TORCH_NUM_THREADS=1
 
 # Misc
 PUBLISH_TIMEOUT_S=60

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

config.py

@@ -40,6 +40,7 @@ class Config:
     enhance_hop_s: float = 0.2
     # SpeechBrain Delay-and-Sum (GCC-PHAT) on C920 stereo before enhance
     beamform_mode: str = "auto"  # auto | force | off
+    torch_num_threads: int = 1
     # misc
     publish_timeout_s: float = 60.0
     log_level: str = "INFO"
@@ -86,6 +87,7 @@ def load_config(env_path: Optional[Path] = None) -> Config:
         enhance_chunk_s=float(os.environ.get("ENHANCE_CHUNK_S", "1.0")),
         enhance_hop_s=float(os.environ.get("ENHANCE_HOP_S", "0.2")),
         beamform_mode=os.environ.get("BEAMFORM_MODE", "auto").lower(),
+        torch_num_threads=int(os.environ.get("TORCH_NUM_THREADS", "1")),
         publish_timeout_s=float(os.environ.get("PUBLISH_TIMEOUT_S", "60")),
         log_level=os.environ.get("LOG_LEVEL", "INFO").upper(),
         cameras=_parse_cameras(os.environ.get("CAMERAS", "all")),
@@ -96,4 +98,6 @@ def load_config(env_path: Optional[Path] = None) -> Config:
     if cfg.beamform_mode not in ("auto", "force", "off"):
         raise ValueError(
             f"BEAMFORM_MODE must be auto|force|off, got {cfg.beamform_mode}")
+    if cfg.torch_num_threads < 1:
+        raise ValueError("TORCH_NUM_THREADS must be >= 1")
     return cfg

run.py

@@ -43,6 +43,7 @@ def worker_args_for(cfg, cam) -> list[str]:
         "--enhance-chunk-s", str(cfg.enhance_chunk_s),
         "--enhance-hop-s", str(cfg.enhance_hop_s),
         "--beamform-mode", cfg.beamform_mode,
+        "--torch-num-threads", str(cfg.torch_num_threads),
     ]
 
 

run_worker.py

@@ -38,6 +38,7 @@ def main() -> int:
     ap.add_argument("--enhance-chunk-s", type=float, default=1.0)
     ap.add_argument("--enhance-hop-s", type=float, default=0.2)
     ap.add_argument("--beamform-mode", default="auto")
+    ap.add_argument("--torch-num-threads", type=int, default=1)
     args = ap.parse_args()
 
     cam = json.loads(args.camera_json)
@@ -47,8 +48,9 @@ def main() -> int:
     )
 
     from config import Config  # noqa: F401  (ensures .env is loaded)
+    from audio_cleanup import AudioCleaner, apply_torch_thread_limits
+    apply_torch_thread_limits(args.torch_num_threads)
     import livekit.rtc as rtc
-    from audio_cleanup import AudioCleaner
 
     identity = cam["identity"]
     video_dev = cam["video_device"]
@@ -62,6 +64,7 @@ def main() -> int:
             chunk_s=args.enhance_chunk_s,
             hop_s=args.enhance_hop_s,
             beamform=args.beamform_mode,
+            torch_num_threads=args.torch_num_threads,
         )
         log.info("audio backend: %s", cleaner.backend)
 

tests/test_config.py

@@ -9,3 +9,7 @@ def test_parse_cameras_list():
 
 def test_default_beamform_mode_auto():
     assert Config().beamform_mode == "auto"
+
+
+def test_torch_num_threads_default_one():
+    assert Config().torch_num_threads == 1