Pre-warm LLM into CPU RAM at startup (avoids first-call GPU timeout)

app.py

@@ -22,7 +22,10 @@ MOCK = os.environ.get("TINYWORLD_MOCK", "0") == "1"
 if not MOCK and os.environ.get("TINYWORLD_INFER", "modal").lower() == "local":
     try:
         import inference  # noqa: F401  (registers ZeroGPU GPU tasks)
-        print("[app] ZeroGPU inference module loaded")
+        import threading
+        # warm the model into CPU RAM in the background so startup isn't blocked
+        threading.Thread(target=inference.warmup, daemon=True).start()
+        print("[app] ZeroGPU inference module loaded; warming model in background")
     except Exception as e:
         print(f"[app] inference preimport failed: {e}")
 

inference.py

@@ -40,6 +40,17 @@ def _load_llm():
     return _llm, _tok
 
 
+def warmup():
+    """Download + load the LLM into CPU RAM ahead of time (no GPU needed) so the
+    first @spaces.GPU call only moves to CUDA and generates — avoids the cold 8GB
+    load racing the ZeroGPU duration limit. Safe to call from a background thread."""
+    try:
+        _load_llm()
+        print("[inference] LLM warmed (CPU RAM)")
+    except Exception as e:
+        print(f"[inference] warmup failed: {e}")
+
+
 def _strip_think(text):
     text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL).strip()
     if "<think>" in text: