Add support for greedy decoding
#2
by
adityastomar
- opened
The current implementation of sampling only uses torch.multinomial and does not support greedy decoding when temperature is 0.0 / top-k is 0 / top-p is 1.0. This PR adds support for greedy decoding.