--no-warmup

app.py

@@ -93,6 +93,7 @@ def start_llama_server():
         "-c", "2048", 
         "-t", "8", 
         "--port", "8081",
+        "--no-warmup",
         "-ngl", "999"  # <--- CRUCIAL: GPU offload instruction
     ]
 
@@ -101,7 +102,7 @@ def start_llama_server():
     print(f"Llama-server process started with PID {llama_process.pid}")
 
     # You might need a small sleep here to wait for the server to initialize
-    # time.sleep(5) 
+    time.sleep(5) 
 
 
 custom_theme = ResearchMonochrome()