Debug monkey-patch with error logging

app.py

@@ -3,6 +3,7 @@ import platform
 import sys
 import shutil
 import tempfile
+import traceback
 import torch
 import torchaudio
 import gradio as gr
@@ -134,23 +135,6 @@ 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
 # ---------------------------------------------------------------------------
@@ -189,9 +173,49 @@ with gr.Blocks(title="Respite API") as demo:
     )
 
 
-# Mount Gradio under the FastAPI app.
-demo.queue(max_size=4, default_concurrency_limit=1)
-app = gr.mount_gradio_app(api, demo, path="/")
+# ---------------------------------------------------------------------------
+# Inject discovery routes into Gradio's internal FastAPI server.
+#
+# The HF Spaces Gradio runner calls demo.launch() directly, so our own
+# FastAPI app is never served.  We monkey-patch Blocks.launch so that
+# AFTER Gradio creates its server, we inject routes onto that server.
+# ---------------------------------------------------------------------------
 
-if __name__ == "__main__":
-    demo.launch()
+_original_blocks_launch = gr.Blocks.launch
+
+
+def _patched_blocks_launch(self, *args, **kwargs):
+    _log("Intercepted Blocks.launch() - calling original...")
+    try:
+        result = _original_blocks_launch(self, *args, **kwargs)
+    except Exception:
+        _log("ERROR in original launch:")
+        _log(traceback.format_exc())
+        raise
+
+    if self is demo:
+        try:
+            fa_app = getattr(self, "app", None)
+            _log(f"demo.app is: {fa_app}")
+            if fa_app is not None:
+                @fa_app.get("/respite/resources")
+                def _resources():
+                    return JSONResponse({"resources": API_RESOURCES})
+
+                @fa_app.get("/respite/specs")
+                def _specs():
+                    return JSONResponse({**API_SPECS, "runtime": _get_runtime_specs()})
+
+                _log("Respite API routes injected successfully")
+            else:
+                _log("Warning: demo.app is None after launch")
+        except Exception:
+            _log("ERROR injecting routes:")
+            _log(traceback.format_exc())
+
+    return result
+
+
+gr.Blocks.launch = _patched_blocks_launch
+
+demo.queue(max_size=4, default_concurrency_limit=1)