activation quantization of v_proj and o_proj

Thanks for sharing the code for such nice work!

I have one question about the activation quantization configuration.
It seems that the quantization granularity of both the output of `v_proj` and the input of `o_proj` is always set to `head_dim` in `eval_utils/main.py`, which is not consistent with the conventional setting in other quantization works such as QuaRot.

Could you clarify whether the results reported in the paper are from this head-wise quantization setting or the per-token quantization setting as mentioned in the paper?

Thanks.