"4-bit finetuning” is ambiguous

[jzhang38]

Dear authors, can I ask what 4-bit finetuning exactly refer to?
A. Use 16-bit mixed precision to fine-tune LoRA on a 4-bit base model.
B. Use 4-bit mixed precision to fine-tune LoRA on a 4-bit base model.
C. Use 4-bit mixed prevision to fine-tune a full model.

[jzhang38]

and maybe another slightly dumb qn: how do you store the 4-bit parameters in the GPU memory? Because 4-bit is not directly addressable.