Fix /generate route being shadowed by Gradio's own catch-all -- mount Gradio at /ui instead of adding routes post-hoc to demo.app

app.py

@@ -4,8 +4,17 @@ chatPDB inference API — HuggingFace Space (Gradio SDK, ZeroGPU)
 Exposes POST /generate (SSE) consumed by the Flask PTY app on the droplet.
 Gradio SDK is required for ZeroGPU (confirmed live 2026-07-23: requesting ZeroGPU hardware for a
 Docker-SDK Space returns "ZeroGPU Spaces only work with Gradio SDK") -- the Gradio UI itself is
-just a minimal landing page; a custom FastAPI route is mounted on Gradio's own underlying app to
-keep the exact same /generate contract chat_remote.py already expects.
+just a minimal landing page.
+
+Real fix (found + fixed 2026-07-23): mounting a custom route via `@demo.app.post(...)` on the
+Blocks object's own `.app` attribute does NOT reliably take effect on HF Spaces -- confirmed live
+against the actual deployed Space: `curl -X POST .../generate` returned 405 with `allow: GET`,
+meaning Gradio's own catch-all SPA route claimed the path instead (Gradio's launch machinery
+appears to reconstruct/finalize its FastAPI app at launch time, after the route was added to the
+earlier `demo.app` reference). Fixed using Gradio's own documented pattern for combining custom
+FastAPI routes with a Gradio UI: build a plain FastAPI `app` first, register routes on it, then
+mount Gradio as a sub-application at a distinct path (routes registered directly on the outer
+`app` are not shadowed by Gradio's own routing, which only applies within its mounted sub-path).
 
 Cold-start note: first request after idle downloads the GGUF and allocates the GPU
 (~60-120 s). Subsequent requests within the same GPU lease are fast.
@@ -16,7 +25,8 @@ import json
 
 import gradio as gr
 import spaces
-from fastapi import Request
+import uvicorn
+from fastapi import FastAPI, Request
 from fastapi.responses import StreamingResponse
 from huggingface_hub import hf_hub_download
 
@@ -71,20 +81,12 @@ def _generate_tokens(
     return tokens
 
 
-# -- Gradio UI (minimal -- required for Gradio SDK / ZeroGPU) --
-
-with gr.Blocks(title="chatPDB API") as demo:
-    gr.Markdown(
-        "## 🧬 chatPDB Inference API\n\n"
-        "Internal endpoint for [chatpdb.mdeller.com](https://chatpdb.mdeller.com). "
-        "Use `POST /generate` — returns `text/event-stream` of token chunks.\n\n"
-        "**Cold start:** first request after idle takes ~60-120 s (GGUF download + GPU alloc)."
-    )
+# -- Plain FastAPI app owns the real API routes --
 
+app = FastAPI(title="chatPDB API")
 
-# -- Custom FastAPI route mounted on Gradio's app --
 
-@demo.app.post("/generate")
+@app.post("/generate")
 async def generate(request: Request):
     body = await request.json()
     prompt = body.get("prompt", "")
@@ -102,10 +104,24 @@ async def generate(request: Request):
     return StreamingResponse(event_stream(), media_type="text/event-stream")
 
 
-@demo.app.get("/health")
+@app.get("/health")
 async def health():
     return {"status": "ok"}
 
 
+# -- Gradio UI (minimal -- required for Gradio SDK / ZeroGPU), mounted at a sub-path so it can't
+#    shadow the routes registered directly on `app` above --
+
+with gr.Blocks(title="chatPDB API") as demo:
+    gr.Markdown(
+        "## 🧬 chatPDB Inference API\n\n"
+        "Internal endpoint for [chatpdb.mdeller.com](https://chatpdb.mdeller.com). "
+        "Use `POST /generate` — returns `text/event-stream` of token chunks.\n\n"
+        "**Cold start:** first request after idle takes ~60-120 s (GGUF download + GPU alloc)."
+    )
+
+app = gr.mount_gradio_app(app, demo, path="/ui")
+
+
 if __name__ == "__main__":
-    demo.launch()
+    uvicorn.run(app, host="0.0.0.0", port=7860)

requirements.txt

@@ -3,3 +3,4 @@ llama-cpp-python
 gradio>=4.0
 spaces
 huggingface_hub
+uvicorn