Update start.sh

start.sh

@@ -1,8 +1,35 @@
+#!/usr/bin/env bash
+# set -euo pipefail
+
+# # ensure checkpoint directory exists; model will auto-download in app init if missing
+# mkdir -p /tmp
+
+# # run app
+# exec gunicorn --bind 0.0.0.0:7860 --workers 1 --threads 4 --timeout 120 app:app
+
+
 #!/usr/bin/env bash
 set -euo pipefail
 
-# ensure checkpoint directory exists; model will auto-download in app init if missing
+# Ensure checkpoint directory exists
 mkdir -p /tmp
 
-# run app
-exec gunicorn --bind 0.0.0.0:7860 --workers 1 --threads 4 --timeout 120 app:app
+# Set CPU threading limits
+export OMP_NUM_THREADS=4
+export MKL_NUM_THREADS=4
+export OPENBLAS_NUM_THREADS=4
+
+# Run app with gunicorn
+# --workers 1: Single worker to avoid loading model multiple times
+# --threads 4: Allow 4 threads per worker for concurrent requests
+# --timeout 120: Allow 2 minutes for inference on CPU
+# --preload: Load application code before forking workers (more efficient)
+exec gunicorn \
+    --bind 0.0.0.0:7860 \
+    --workers 1 \
+    --threads 4 \
+    --timeout 120 \
+    --preload \
+    --access-logfile - \
+    --error-logfile - \
+    app:app