perf: async endpoint + startup warmup to reduce ZeroGPU cold-start latency

app.py

@@ -1,5 +1,6 @@
 import spaces
 import os
+import threading
 import gradio as gr
 from transformers import pipeline
 from fastapi import HTTPException
@@ -66,7 +67,6 @@ demo = gr.Interface(
 )
 
 # ── Launch Gradio (non-blocking) — properly initialises ZeroGPU ───────────────
-# demo.launch() returns Gradio's internal FastAPI app which we extend below.
 gr_app, _local_url, _share_url = demo.launch(
     server_name="0.0.0.0",
     server_port=7860,
@@ -76,7 +76,6 @@ gr_app, _local_url, _share_url = demo.launch(
 
 # ── REST API routes added to Gradio's FastAPI app ─────────────────────────────
 
-
 class DetectRequest(BaseModel):
     model_config = ConfigDict(
         json_schema_extra={"example": {"text": "Mage name Kamal, phone 071-1234567"}}
@@ -104,7 +103,7 @@ def health_check():
 
 
 @gr_app.post("/api/detect", response_model=DetectResponse)
-def api_detect_pii(request: DetectRequest):
+async def api_detect_pii(request: DetectRequest):
     """
     Detect PII entities in Sinhala-English code-mixed text.
 
@@ -129,5 +128,16 @@ def api_detect_pii(request: DetectRequest):
     }
 
 
+# ── Startup warmup — pre-warms ZeroGPU so first real request is fast ──────────
+def _warmup():
+    try:
+        print("[warmup] Pre-warming ZeroGPU GPU allocation...")
+        detect_pii("warmup kamal@test.com 0771234567")
+        print("[warmup] Done.")
+    except Exception as e:
+        print(f"[warmup] Skipped: {e}")
+
+threading.Thread(target=_warmup, daemon=True).start()
+
 # ── Keep the process alive ─────────────────────────────────────────────────────
 demo.block_thread()