Update app.py

app.py

@@ -555,7 +555,7 @@ def respond(
             flash_attn=True,  
             n_gpu_layers=-1,
                 n_batch=2048,        # increase
-                n_ctx= 32768,          # reduce if you don’t need 8k
+                n_ctx= 16384,          # reduce if you don’t need 8k
                 # n_threads= os.cpu_count(),        # set to your CPU cores
                  n_threads=os.cpu_count(),        # use all cores
                 n_threads_batch=os.cpu_count(),