Keep production startup independent of benchmarks

servers/unraid/tail2-pose-server/app/tail2/main.py

@@ -157,7 +157,8 @@ def inference_worker(loop: asyncio.AbstractEventLoop) -> None:
     threshold = float(os.getenv("DETECTION_THRESHOLD", ".35")); max_people = int(os.getenv("MAX_PEOPLE", "4"))
     try:
         models.load()
-        run_offline_benchmark(models, threshold, max_people)
+        if os.getenv("RUN_STARTUP_BENCHMARK", "false").lower() == "true":
+            run_offline_benchmark(models, threshold, max_people)
         state.models_ready = True; LOG.info("official detector and pose checkpoints loaded")
     except Exception as exc:
         state.inference_fault = True; state.error_code = "MODEL_LOAD_FAILED"

servers/unraid/tail2-pose-server/app/tail2/models.py

@@ -7,6 +7,7 @@ from typing import Any
 
 import numpy as np
 import torch
+import cv2
 
 from .core import normalize_bbox, normalize_keypoints
 
@@ -45,6 +46,7 @@ class Models:
         return digest.hexdigest()
 
     def load(self) -> None:
+        cv2.setNumThreads(int(os.getenv("CV2_NUM_THREADS", "2")))
         torch.backends.cudnn.benchmark = True
         torch.backends.cuda.matmul.allow_tf32 = True
         torch.backends.cudnn.allow_tf32 = True

servers/unraid/tail2-pose-server/compose.yaml

@@ -111,6 +111,11 @@ services:
       SAVE_FRAMES: "false"
       SAVE_CROPS: "false"
       LOG_COORDINATES: "false"
+      RUN_STARTUP_BENCHMARK: "false"
+      OMP_NUM_THREADS: "2"
+      MKL_NUM_THREADS: "2"
+      OPENBLAS_NUM_THREADS: "2"
+      CV2_NUM_THREADS: "2"
       CUDA_VISIBLE_DEVICES: "1"
       DETECTOR_NAME: yolox-m-humanart-coco-640x640
       DET_CONFIG: ""