Everyone, I have implemented multi-token prediction of InfiniAttention and meta.

[win10ogod]

Is anyone willing to help improve memory consumption?
I will post a PR.
(https://github.com/win10ogod/llama2-InfiniAttention.c/blob/master/model.py)
I commented outmodel_export(raw_model, os.path.join(out_dir, "model.bin"), version=0)
Because I don’t understand C language, there are also errors in export .py

[bityigoss]

if we still tie the weights of lm_head in multi-token prediction, how they output different token predictions? @win10ogod