Generate: nudge towards do_sample=False when temperature=0.0

[gante]

What does this PR do?

Related issue: meta-llama/llama#687

Improves the error message when temperature=0.0, which asymptotically corresponds to greedy decoding... except that it results in numerical problems :D

---

test run:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
model = AutoModelForCausalLM.from_pretrained("distilgpt2")

inputs = tokenizer(["The quick brown"], return_tensors="pt")
gen_out = model.generate(**inputs, do_sample=True, temperature=0.0)

yields

ValueError: `temperature` (=0.0) has to be a strictly positive float, otherwise your next token scores will be invalid. If you're looking for greedy decoding strategies, set `do_sample=False`.

[HuggingFaceDocBuilderDev]

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

[ArthurZucker]

We could also just set do_sample = False in case temperature = 0. Will let you decide !

[ArthurZucker]

do you mean that it will be nan?

[gante]

depends on the value the user places here (e.g. a negative float will not generate nans, but make the scores enter uncharted territory), hence the vague message

[gante]

We could also just set do_sample = False in case temperature = 0. Will let you decide !

I agree we should do that! But I'm going to leave that for the generate refactor, as it implies significant code changes to do it right :)