Serve llama.cpp web UI on port 7860 via gateway proxy

gateway.py was the only process exposed on 7860 (HF Spaces only route
one port), but it had no route for "/" or static assets, and
llama-server was started with --no-webui, so every UI request 404'd.
Enable the built-in webui in llama-server and add a catch-all reverse
proxy in gateway.py that forwards any unmatched request to it.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XvAfE7bsSLrQm1VVBswYnx

.gitignore

@@ -0,0 +1,2 @@
+__pycache__/
+*.pyc

gateway.py

@@ -356,4 +356,71 @@ async def cancel_task(job_id: str) -> dict[str, Any]:
             "UPDATE jobs SET status='cancelled', updated_at=? WHERE id=?",
             (time.time(), job_id),
         )
-    return {"task_id": job_id, "status": "cancelled"}
+    return {"task_id": job_id, "status": "cancelled"}
+
+
+# Everything below is not part of this service's own API. It falls through to
+# llama.cpp's built-in web UI (served on LLAMA_URL, reachable only from
+# localhost inside the container), so that visiting the Space's public URL
+# renders the llama.cpp chat UI instead of a 404.
+_HOP_BY_HOP_HEADERS = {
+    "connection",
+    "keep-alive",
+    "proxy-authenticate",
+    "proxy-authorization",
+    "te",
+    "trailers",
+    "transfer-encoding",
+    "upgrade",
+    "content-length",
+    "host",
+}
+
+
+@app.api_route(
+    "/{full_path:path}",
+    methods=["GET", "HEAD", "POST", "PUT", "PATCH", "DELETE", "OPTIONS"],
+)
+async def proxy_to_llama_webui(full_path: str, request: Request) -> Any:
+    url = f"{LLAMA_URL}/{full_path}"
+    headers = {
+        key: value
+        for key, value in request.headers.items()
+        if key.lower() not in _HOP_BY_HOP_HEADERS
+    }
+    body = await request.body()
+
+    client = httpx.AsyncClient(timeout=REQUEST_TIMEOUT)
+    try:
+        upstream_request = client.build_request(
+            request.method,
+            url,
+            params=request.query_params,
+            headers=headers,
+            content=body,
+        )
+        upstream_response = await client.send(upstream_request, stream=True)
+    except httpx.HTTPError:
+        await client.aclose()
+        raise HTTPException(502, "llama.cpp server is unavailable")
+
+    response_headers = {
+        key: value
+        for key, value in upstream_response.headers.items()
+        if key.lower() not in _HOP_BY_HOP_HEADERS
+    }
+
+    async def stream_and_close():
+        try:
+            async for chunk in upstream_response.aiter_raw():
+                yield chunk
+        finally:
+            await upstream_response.aclose()
+            await client.aclose()
+
+    return StreamingResponse(
+        stream_and_close(),
+        status_code=upstream_response.status_code,
+        headers=response_headers,
+        media_type=upstream_response.headers.get("content-type"),
+    )

start.sh

@@ -21,8 +21,7 @@ THREADS="${THREADS:-2}"
   --cache-type-v q8_0 \
   --flash-attn auto \
   --cont-batching \
-  --metrics \
-  --no-webui &
+  --metrics &
 
 LLAMA_PID=$!