DeepSpeed
DeepSpeed copied to clipboard
Refine quantizer for supporting larger hidden-dim and group size
Can one of the admins verify this patch?
Changes fixed under later memory refactor.