debug + gpu optimisation

app.py

@@ -11,6 +11,8 @@ from controllable_blender import ControllableBlender
 from huggingface_hub import snapshot_download
 from huggingface_hub import login
 
+torch.set_default_dtype(torch.float16)
+
 token = os.environ.get("Token1")
 
 login(token=token)
@@ -38,6 +40,7 @@ def init_world(cefr, inference_type):
     opt = agent_opt.copy()
     opt["rerank_cefr"] = cefr
     opt["inference"] = inference_type
+    opt["gpu"]
 
     # Settings for rerank methods (not used if "inference" == "vocab")
     opt["rerank_tokenizer"] = "distilroberta-base"        # Tokenizer from Huggingface Transformers. Must be compatible with "rerank_model"
@@ -66,8 +69,12 @@ def chat(user_input, cefr, inference_type, history):
         conversation_state["world"] = world
         conversation_state["human_agent"] = human_agent
 
+        print("🔥 Warming up...")
+        conversation_state["human_agent"].msg = "Hello"
+        conversation_state["world"].parley()
+        print("✅ Warmup complete.")
+    
     conversation_state["human_agent"].msg = user_input
-
     conversation_state["world"].parley()
 
     bot_reply = conversation_state["world"].acts[1].get("text", "")