Update app.py

app.py

@@ -396,7 +396,8 @@ class Phi3MiniEducationalLLM(Runnable):
                     repetition_penalty=1.1,
                     pad_token_id=self.tokenizer.eos_token_id,
                     early_stopping=True,
-                    use_cache=True         # Enable cache for performance
+                    use_cache=False
+                    past_key_values=None
                 )
 
             # Decode only new tokens
@@ -467,7 +468,8 @@ class Phi3MiniEducationalLLM(Runnable):
                 "repetition_penalty": 1.2,
                 "pad_token_id": self.tokenizer.eos_token_id,
                 "streamer": streamer,
-                "use_cache": True
+                "use_cache": False
+                "past_key_values": None
             }
 
             # Start generation in background