Megatron-DeepSpeed

2004ef1a - [Feature] Porting bitsandbytes to meg-deepspeed (#144)

[Feature] Porting bitsandbytes to meg-deepspeed (#144) * Port bitsandbytes for meg-deepspeed * Cleanup commit * Add functional test * fix test * refactor the new test into the existing one * fix args * deal with custom optimizer * more refactoring Co-authored-by: Stas Bekman <stas@stason.org>