Use /respite/ prefix with working Gradio mount structure

app.py

@@ -7,6 +7,7 @@ import torch
 import torchaudio
 import gradio as gr
 import spaces
+from fastapi import FastAPI
 from fastapi.responses import JSONResponse
 from einops import rearrange
 from huggingface_hub import login
@@ -70,11 +71,11 @@ API_SPECS = {
     "name": "Respite API",
     "version": "1.0.0",
     "description": "General-purpose AI API server with audio generation capabilities.",
-    "base_path": "/api",
+    "base_path": "/respite",
     "authentication": "none",
     "content_types": ["application/json", "audio/wav"],
-    "resources_endpoint": "/api/resources",
-    "specs_endpoint": "/api/specs",
+    "resources_endpoint": "/respite/resources",
+    "specs_endpoint": "/respite/specs",
     "limits": {
         "max_concurrent_requests": 1,
         "max_queue_size": 4,
@@ -133,6 +134,23 @@ def generate_audio(prompt, duration, steps, cfg_scale, seed, model_name):
     return output_path
 
 
+# ---------------------------------------------------------------------------
+# FastAPI application with discovery routes
+# ---------------------------------------------------------------------------
+
+api = FastAPI(title=API_SPECS["name"], version=API_SPECS["version"])
+
+
+@api.get("/respite/resources")
+def resources():
+    return JSONResponse({"resources": API_RESOURCES})
+
+
+@api.get("/respite/specs")
+def specs():
+    return JSONResponse({**API_SPECS, "runtime": _get_runtime_specs()})
+
+
 # ---------------------------------------------------------------------------
 # Gradio UI
 # ---------------------------------------------------------------------------
@@ -171,43 +189,9 @@ with gr.Blocks(title="Respite API") as demo:
     )
 
 
-# ---------------------------------------------------------------------------
-# Inject discovery routes into Gradio's own FastAPI server.
-#
-# The HF Spaces Gradio runner calls demo.launch() directly.  Our FastAPI
-# routes must be registered on Gradio's internal FastAPI app AFTER the
-# server is created.  We monkey-patch Blocks.launch to intercept that call
-# and inject the routes before the server begins accepting traffic.
-# ---------------------------------------------------------------------------
-
-_original_launch = gr.Blocks.launch
-
-
-def _inject_respite_routes(block_demo):
-    """Add Respite API discovery routes to the running Gradio FastAPI app."""
-    fa_app = getattr(block_demo, "app", None)
-    if fa_app is None:
-        _log("Warning: could not access Gradio FastAPI app for route injection")
-        return
-
-    @fa_app.get("/api/resources")
-    def _resources():
-        return JSONResponse({"resources": API_RESOURCES})
-
-    @fa_app.get("/api/specs")
-    def _specs():
-        return JSONResponse({**API_SPECS, "runtime": _get_runtime_specs()})
-
-    _log("Respite API routes injected into Gradio server")
-
-
-def _patched_launch(self, *args, **kwargs):
-    result = _original_launch(self, *args, **kwargs)
-    if self is demo:
-        _inject_respite_routes(self)
-    return result
-
-
-gr.Blocks.launch = _patched_launch
-
+# Mount Gradio under the FastAPI app.
 demo.queue(max_size=4, default_concurrency_limit=1)
+app = gr.mount_gradio_app(api, demo, path="/")
+
+if __name__ == "__main__":
+    demo.launch()