Update app.py

app.py

@@ -8,7 +8,8 @@ model = Llama(
     model_path=hf_hub_download(
         repo_id=os.environ.get("REPO_ID", "Lyte/QuadConnect2.5-0.5B-v0.0.4b"),
         filename=os.environ.get("MODEL_FILE", "unsloth.Q8_0.gguf"),
-    )
+    ),
+    n_ctx=16384
 )
 
 SYSTEM_PROMPT = """You are an expert Connect Four player. The game is played on a 6x7 grid where pieces fall to the lowest available position in each column due to gravity.
@@ -237,7 +238,7 @@ def create_interface():
                 ],
                 temperature=0.8,
                 top_p=0.95,
-                #max_tokens=4096
+                max_tokens=2048
             )
             
             ai_response = response['choices'][0]['message']['content']