Decouple deterministic smoke from GPU quota

app.py

@@ -633,8 +633,7 @@ def recent_failure_reports():
     return [str(path) for path in sorted(OUTPUT_DIR.glob("*_failure.json"), key=lambda item: item.stat().st_mtime, reverse=True)[:10]]
 
 
-@GPU_FUNCTION
-def generate(image: Image.Image, extra_files, backend: str, guess: str, demo_mode: bool, progress=gr.Progress()):
+def _generate_impl(image: Image.Image, extra_files, backend: str, guess: str, demo_mode: bool, progress=gr.Progress()):
     if image is None:
         raise gr.Error("Please add an input image first.")
     started = time.perf_counter()
@@ -840,6 +839,18 @@ def generate(image: Image.Image, extra_files, backend: str, guess: str, demo_mod
         ) from exc
 
 
+@GPU_FUNCTION
+def _generate_gpu(image: Image.Image, extra_files, backend: str, guess: str, demo_mode: bool, progress=gr.Progress()):
+    return _generate_impl(image, extra_files, backend, guess, demo_mode, progress)
+
+
+def generate(image: Image.Image, extra_files, backend: str, guess: str, demo_mode: bool, progress=gr.Progress()):
+    """Keep the deterministic export diagnostic independent of GPU quota."""
+    if demo_mode:
+        return _generate_impl(image, extra_files, backend, guess, demo_mode, progress)
+    return _generate_gpu(image, extra_files, backend, guess, demo_mode, progress)
+
+
 with gr.Blocks(title="Simam3D") as demo:
     gr.Markdown("# Simam3D\nSingle-image depth and 3D reconstruction workbench")
     gr.Markdown(