Update app.py

app.py

@@ -298,6 +298,14 @@ def run_model(
         .numpy()
     )
 
+    # Free GPU memory immediately — with 3 models running back to
+    # back on ZeroGPU's shared allocation, not doing this can push
+    # peak memory over budget and cause a silent worker kill with
+    # no Python traceback.
+    del tensor, prediction
+    if DEVICE == "cuda":
+        torch.cuda.empty_cache()
+
     # --------------------------------------------------------
     # Back to EXACT original dimensions.
     # --------------------------------------------------------
@@ -610,6 +618,10 @@ def run_vitmatte_refine(
 
     alpha = out.alphas[0, 0].float().cpu().numpy()
 
+    del inputs, out
+    if DEVICE == "cuda":
+        torch.cuda.empty_cache()
+
     # ViTMatte processor pads to a stride-32 canvas internally;
     # resize back to the exact crop size.
     alpha = cv2.resize(
@@ -1593,7 +1605,7 @@ def finalize_alpha(alpha):
 # ============================================================
 
 
-@spaces.GPU(duration=180)
+@spaces.GPU(duration=280)
 def remove_background(
     image
 ):
@@ -1858,4 +1870,4 @@ with gr.Blocks(
 
 demo.queue().launch(
     css=css
-)
+)