I just wanted to update this to mention that there are a lot of custom low level performance improvements for CPU based inferencing in Llamafile: https://justine.lol/matmul/