ChatRWKV icon indicating copy to clipboard operation
ChatRWKV copied to clipboard

No time_shift use in ChatRWKV?

Open 3outeille opened this issue 2 years ago • 1 comments

Hi,

Why is time_shift not applied in ChatRWKV on x before computing x * self.time_mix_k + xx * (1 - self.time_mix_k) while in RWKV V4, it is the case. Any idea ?

3outeille avatar Mar 06 '23 09:03 3outeille

state[5*i+1] is the x in previous iteration, see here. It's exactly what time_shift do.

Blealtan avatar Mar 06 '23 14:03 Blealtan