Update app.py

app.py

@@ -3,7 +3,25 @@ import os
 # Must be set before torch's OpenMP/MKL thread pools initialize on first use.
 # This is the most reliable way to make sure BLAS/OpenMP actually uses all
 # cores instead of a conservative default.
-_CPU_THREADS = str(os.cpu_count() or 4)
+def _detect_cpu_threads() -> int:
+    # Respect a value the platform/container already set -- on cgroup-limited
+    # containers (like HF Spaces) this is often the *correct* real quota,
+    # whereas os.cpu_count() reports the host machine's full core count and
+    # will cause thread oversubscription if trusted blindly.
+    for var in ("OMP_NUM_THREADS", "MKL_NUM_THREADS"):
+        val = os.environ.get(var)
+        if val and val.isdigit() and int(val) > 0:
+            return int(val)
+    # sched_getaffinity reflects CPU-affinity/cgroup restrictions more
+    # accurately than os.cpu_count() on Linux; fall back to cpu_count if
+    # unavailable (e.g. non-Linux).
+    try:
+        return len(os.sched_getaffinity(0))
+    except AttributeError:
+        return os.cpu_count() or 4
+
+
+_CPU_THREADS = str(_detect_cpu_threads())
 os.environ.setdefault("OMP_NUM_THREADS", _CPU_THREADS)
 os.environ.setdefault("MKL_NUM_THREADS", _CPU_THREADS)
 
@@ -85,8 +103,6 @@ def startup_event():
         else:
             MODEL = Qwen3TTSModel.from_pretrained(MODEL_NAME, **load_kwargs)
 
-        MODEL.eval()  # disables dropout etc.; inference-only mode
-
         SUPPORTED_SPEAKERS = MODEL.get_supported_speakers()
         SUPPORTED_LANGUAGES = MODEL.get_supported_languages()