Faster model

app.py

@@ -80,11 +80,13 @@ try:
         model_path=model_path,
         n_ctx=3072,
         n_threads=2,
-        n_batch=128,
+        n_batch=64,
         n_gpu_layers=0,
         use_mlock=True,
         f16_kv=True,
         verbose=True,
+        batch_prefill=True,
+        prefill_logits=False,
     )
     logger.info(f"{filename} model loaded")