Docs fix: Multinomial sampling decoding needs "num_beams=1", since by default it is usually not 1.

[manueldeprada]

Fix error in docs: multinomial sampling decoding strategy

As indicated in the library source code:
https://github.com/huggingface/transformers/blob/228792a9dc0c36f1e82ab441e1b1991d116ee0a0/src/transformers/generation/utils.py#LL1364-L1367

Multinomial sampling needs num_beams=1. However, this is not indicated in the docs, potentially leading to execute beam-search multinomial sampling instead of the intended multinomial sampling.

This deviation from the expected behaviour happens quite often, since a lot of models have in their generation_config.json the parameter num_beams set to something higher than 1. This happens, for example, in the majority of top translation models from the Hub.

Also, I have included "ancestral sampling" as another name for multinomial sampling, since it is the most common name in the decoding algorithms literature.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).

Who can review?

Original authors of this piece of documentation: @gante, @sgugger, @stevhliu and @MKhalusova

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[gante]

LGTM, thanks for the correction!