the int8 kernel looks poorly-performing, it doesn't use vectorised ops.
furthermore, this has nothing to do with bitsandbytes, it doesn't have NF4 or any of the standalone ops that are expected by downstream integrators.
this has no shared functionality with bitsandbytes.
it doesn't share the ABI of bitsandbytes.
it just seems to serve the needs of the spam campaign you're currently running on Github.