mlx_lm sync-eval decode 370.4 tok/s async 468.7 tok/s titan_fast sync-eval decode 380.3 tok/s async 479.0 tok/s titan_ident sync-eval decode 384.6 tok/s async 475.9 tok/s titan_i32 sync-eval decode 365.2 tok/s titan_map32 sync-eval decode 352.2 tok/s async 469.4 tok/s dtype of mlx-lm indices: mlx.core.uint32