Fix: Remove HTML/console endpoints, use APIRouter for clean API integration

app.py

@@ -294,6 +294,7 @@ def _generate_with_gpu(
 
 
 def healthcheck() -> dict[str, str]:
+    """Health check endpoint handler."""
     return {
         "status": "ok",
         "model": MODEL_ID,
@@ -307,6 +308,7 @@ def warm_start() -> None:
 
 
 def generate_endpoint(payload: GeneratePayload) -> GenerateResponse:
+    """Generate endpoint handler."""
     try:
         text = _generate_with_gpu(
             prompt=payload.prompt,
@@ -480,43 +482,27 @@ with gr.Blocks(
         fn=lambda: ("", "", STATUS_IDLE),
         outputs=[prompt_input, output, status_display],
     )
-    
-    # Add API routes using Gradio's load event to ensure they're registered correctly
-    def add_api_routes():
-        """Add API routes after Gradio is fully initialized."""
-        try:
-            from fastapi.responses import JSONResponse
-            from starlette.routing import Route
-            from starlette.requests import Request
-            
-            async def health_handler(request: Request):
-                """Handle GET /health requests."""
-                return JSONResponse(content=healthcheck())
-            
-            async def generate_handler(request: Request):
-                """Handle POST /v1/generate requests."""
-                try:
-                    data = await request.json()
-                    payload = GeneratePayload(**data)
-                    result = generate_endpoint(payload)
-                    return JSONResponse(content={"text": result.text})
-                except Exception as exc:
-                    from fastapi import HTTPException
-                    raise HTTPException(status_code=500, detail=str(exc))
-            
-            # Add routes directly to Gradio's router
-            # Insert at position 0 to ensure they're checked before Gradio's catch-all routes
-            gradio_app.app.router.routes.insert(0, Route("/health", health_handler, methods=["GET"]))
-            gradio_app.app.router.routes.insert(0, Route("/v1/generate", generate_handler, methods=["POST"]))
-            print(f"API routes added via load event. Total routes: {len(gradio_app.app.routes)}")
-        except Exception as e:
-            print(f"Warning: Could not add API routes: {e}")
-            import traceback
-            traceback.print_exc()
-    
-    # Register the load event handler
-    gradio_app.load(add_api_routes)
 
+# Attach API routes directly onto Gradio's FastAPI instance using APIRouter
+api_router = APIRouter()
+
+
+@api_router.get("/health")
+def api_health() -> dict[str, str]:
+    """API health check endpoint."""
+    return healthcheck()
+
+
+@api_router.post("/v1/generate", response_model=GenerateResponse)
+def api_generate(payload: GeneratePayload) -> GenerateResponse:
+    """API generate endpoint."""
+    return generate_endpoint(payload)
+
+
+# Include the router in Gradio's FastAPI app
+gradio_app.app.include_router(api_router)
+
+# Call warm start
 warm_start()
 
 # Enable queued execution so ZeroGPU can schedule GPU work reliably