perf: optimize CPU inference via thread limiting and inference mode usage

app.py

@@ -21,7 +21,14 @@ from df import config
 from df.enhance import enhance, init_df, load_audio, save_audio
 from df.io import resample
 
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+# Optimize PyTorch CPU execution for faster inference
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+else:
+    device = torch.device("cpu")
+    # Limit intra-op thread count to avoid scheduling overhead on multi-core environments
+    torch.set_num_threads(4)
+
 model, df, _ = init_df("./DeepFilterNet2", config_allow_defaults=True)
 model = model.to(device=device).eval()
 
@@ -184,7 +191,8 @@ def demo_fn(
 
     logger.info("Start denoising audio")
     # Call enhance with attenuation limit (atten_lim_db) to prevent over-silencing & artifacts
-    enhanced = enhance(model, df, sample, atten_lim_db=atten_lim_db)
+    with torch.inference_mode():
+        enhanced = enhance(model, df, sample, atten_lim_db=atten_lim_db)
     logger.info("Denoising finished")
 
     # Dry/wet blending: mix original 'sample' back into 'enhanced' to restore voice texture