Fix FastAPI route mounting with proper callback

app.py

@@ -411,10 +411,31 @@ with gr.Blocks(title="Router Model API - ZeroGPU") as gradio_app:
         inputs=[prompt_input, max_tokens_input, temp_input, top_p_input],
         outputs=output,
     )
-    
-    # Mount FastAPI routes to Gradio's underlying FastAPI app
-    gradio_app.app.mount("/v1", fastapi_app)
-    gradio_app.app.mount("/gradio", fastapi_app)
+
+# Mount FastAPI routes to Gradio's underlying FastAPI app after Blocks context
+# Use a callback to mount routes when the app is ready
+def mount_fastapi_on_load():
+    """Mount FastAPI routes when Gradio app loads."""
+    try:
+        # Access Gradio's FastAPI app and mount our routes
+        gradio_app.app.mount("/v1", fastapi_app)
+        gradio_app.app.mount("/gradio", fastapi_app)
+    except (AttributeError, Exception) as e:
+        print(f"Note: FastAPI routes mounting deferred: {e}")
+
+# Try to mount immediately, and also set up for when app launches
+try:
+    mount_fastapi_on_load()
+except:
+    # If mounting fails now, it will be handled during launch
+    pass
+
+# Override launch to mount routes
+original_launch = gradio_app.launch
+def launch_with_mount(*args, **kwargs):
+    mount_fastapi_on_load()
+    return original_launch(*args, **kwargs)
+gradio_app.launch = launch_with_mount
 
 # Set app to Gradio Blocks for Spaces - ZeroGPU requires Gradio SDK
 app = gradio_app