Upload folder using huggingface_hub

app.py

@@ -198,9 +198,11 @@ def _duration(screenshot, instruction, previous_actions="", enable_thinking=True
         budget = int(max_new_tokens)
     except (TypeError, ValueError):
         budget = 768
-    # Measured on the live Space: ~28 tok/s decode, ~4 s of prefill and image
-    # encoding per call. Sized to the observed worst case plus ~15%.
-    return int(min(150, 10 + budget * 0.042))
+    # Measured on the live Space: ~3.5 s of prefill + image encoding and ~42
+    # tok/s decode (6.9-8.6 s for 130-220-token replies on 1920x1080 frames).
+    # Sized to that rate plus cold-worker weight streaming and a small margin,
+    # deliberately tight so the visitor's ZeroGPU quota check stays cheap.
+    return int(min(90, 12 + budget * 0.028))
 
 
 @spaces.GPU(duration=_duration)
@@ -353,7 +355,7 @@ CSS = """
 .dark .gradio-container { color: var(--body-text-color); }
 """
 
-with gr.Blocks(title="UI-Mate-9B", theme=gr.themes.Citrus(), css=CSS) as demo:
+with gr.Blocks(title="UI-Mate-9B") as demo:
     gr.Markdown(
         """
         # 🖱️ UI-Mate-9B — next GUI action from a screenshot
@@ -448,4 +450,7 @@ with gr.Blocks(title="UI-Mate-9B", theme=gr.themes.Citrus(), css=CSS) as demo:
     )
 
 if __name__ == "__main__":
-    demo.queue(max_size=12).launch(mcp_server=True)
+    # Gradio 6 moved `theme` / `css` from the Blocks constructor to launch().
+    demo.queue(max_size=12).launch(
+        theme=gr.themes.Citrus(), css=CSS, mcp_server=True
+    )