Fix: use gr.Blocks.launch(), add REST API via demo.app, add warmup GPU func

app.py

@@ -13,6 +13,7 @@ if not hasattr(huggingface_hub, "HfFolder"):
     huggingface_hub.HfFolder = _HfFolder
 
 import gradio as gr
+
 import spaces
 
 logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
@@ -49,7 +50,7 @@ TOOL_DEFS = {
 }
 
 # ============================================================
-# GPU Functions
+# GPU Functions — must be module-level for ZeroGPU detection
 # ============================================================
 
 _LLM_PIPELINE = None
@@ -57,6 +58,12 @@ _LLM_TOKENIZER = None
 _IMAGE_PIPELINE = None
 _CODE_GLOBALS: dict = {}
 
+@spaces.GPU(duration=60)
+def _gpu_warmup() -> str:
+    """Warmup function for ZeroGPU detection."""
+    import torch
+    return f"GPU available: {torch.cuda.is_available()}, device: {torch.cuda.get_device_name(0)}"
+
 @spaces.GPU(duration=120)
 def _gpu_llm_chat(messages_json: str, tools_json: str) -> str:
     """Qwen3-30B inference on H200. Returns JSON with content + tool_calls."""
@@ -350,6 +357,25 @@ def create_app():
 
 if __name__ == "__main__":
     demo = create_app()
-    app = gr.mount_gradio_app(api_app, demo, path="/")
-    import uvicorn
-    uvicorn.run(app, host="0.0.0.0", port=7860)
+    # Add REST API endpoint
+    @demo.app.post("/api/agent")
+    def api_agent(req: dict):
+        agent_type = req.get("agent_type", "web")
+        message = req.get("message", "")
+        if agent_type not in AGENT_REGISTRY:
+            agent_type = "web"
+        messages = []
+        for update in run_agent(message, [], agent_type):
+            if isinstance(update, tuple):
+                chat_history = update[0]
+                status = update[1]
+                if "✅" in status and chat_history:
+                    last = chat_history[-1]
+                    return {"status": "ok", "agent_type": agent_type, "response": last[1] or ""}
+        return {"status": "ok", "agent_type": agent_type, "response": "No response."}
+    
+    @demo.app.get("/api/health")
+    def health():
+        return {"status": "ok", "model": "Qwen3-30B-A3B", "gpu": "H200"}
+    
+    demo.launch(server_name="0.0.0.0", server_port=7860)