Add temperature, top-k and top-p logits processors

[SimJeg]

Hello,

In outlines.text.generate.sample, only greedy and multinomial methods are available. Would it be possible to implement some classical features in the multinomial function such as top_k, top_p or temperature ?

Thanks for your work,
Simon

[brandonwillard]

Yes, definitely!

[brandonwillard]

Also, we need to add beam search.

[rlouf]

Also, we need to add beam search.

Tracked by #258

[rlouf]

#533 introduces a small refactor of the samplers that would allow to pass the top_k and top_p arguments to the multinomial sampler. We could pass the temperature using the same mechanism as well

[rlouf]

top-k was added in #642, and I opened separate issues for the other requests: #645 and #643.