attention weights:
 [[1.     0.     0.    ]
 [0.3302 0.6698 0.    ]
 [0.2483 0.2483 0.5035]]
last output: [0.7517 0.7517]
future change leaves earlier positions unchanged: True
cached last output: [0.6304 0.8374]
max error: 0.0
query positions: full=4, cached=1; KV positions=4
