As explained here #10343 (comment), we would like to keep the CPU implementations inside the CPU backend. The AMX backend was created mainly because at the time we didn't support runtime weight repacking. Since now this functionality is supported, we should merge the AMX backend into the CPU backend.

The rough plan to achieve that is outlined here: #10350 (reply in thread)

The plan to reintegrate the AMX backend would be to create a new buffer type that converts the weights to the layout that the AMX backend needs them, and then check in the matrix multiplication the buffer type to determine if the AMX matrix multiplication code should be used. Basically extending the same that is done in #9921 for the aarch64 types.