Patch demo.launch to inject FastAPI routes into Gradio server

app.py

@@ -105,6 +105,23 @@ def _get_runtime_specs():
     }
 
 
+# ---------------------------------------------------------------------------
+# FastAPI application with discovery routes
+# ---------------------------------------------------------------------------
+
+api = FastAPI(title=API_SPECS["name"], version=API_SPECS["version"])
+
+
+@api.get("/api/resources")
+def resources():
+    return JSONResponse({"resources": API_RESOURCES})
+
+
+@api.get("/api/specs")
+def specs():
+    return JSONResponse({**API_SPECS, "runtime": _get_runtime_specs()})
+
+
 # ---------------------------------------------------------------------------
 # Model cache
 # ---------------------------------------------------------------------------
@@ -156,23 +173,6 @@ def generate_audio(prompt, duration, steps, cfg_scale, seed, model_name):
     return output_path
 
 
-# ---------------------------------------------------------------------------
-# FastAPI application
-# ---------------------------------------------------------------------------
-
-api = FastAPI(title=API_SPECS["name"], version=API_SPECS["version"])
-
-
-@api.get("/api/resources")
-def resources():
-    return JSONResponse({"resources": API_RESOURCES})
-
-
-@api.get("/api/specs")
-def specs():
-    return JSONResponse({**API_SPECS, "runtime": _get_runtime_specs()})
-
-
 # ---------------------------------------------------------------------------
 # Gradio UI
 # ---------------------------------------------------------------------------
@@ -217,9 +217,23 @@ with gr.Blocks(title="Respite API") as demo:
         outputs=audio_output,
     )
 
+
+# ---------------------------------------------------------------------------
+# Patch launch() so the HF Spaces runner picks up our FastAPI routes.
+#
+# The Gradio SDK runner calls demo.launch() directly, which normally creates
+# its own FastAPI app. By intercepting that call we inject our routes so
+# /api/resources and /api/specs are served alongside the Gradio UI.
+# ---------------------------------------------------------------------------
+
 demo.queue(max_size=4, default_concurrency_limit=1)
 
-# Mount Gradio UI under the FastAPI app.
-# Routes are defined on `api` first, then Gradio is mounted at "/".
-# FastAPI evaluates routes top-down, so /api/* routes are served directly.
-app = gr.mount_gradio_app(api, demo, path="/")
+_original_launch = demo.launch
+
+
+def _patched_launch(*args, **kwargs):
+    kwargs.setdefault("app_kwargs", {})["app"] = api
+    return _original_launch(*args, **kwargs)
+
+
+demo.launch = _patched_launch