Cache: don't show warning in forward passes when past_key_values is None

[gante]

What does this PR do?

Because of the transition from tuple of tuples to Cache instances, we were throwing a warning when converting past_key_values to the new cache format in the forward passes.

One of those situations was when use_cache=True and past_key_values is None... but there is nothing to convert there. In fact, most of the times, the user didn't even specify the argument (see test script below). Moreover, after the transition is complete, we want to keep the default past_key_values=None argument.

As such, this PR removes the warning when past_key_values=None.

Fixes #33489

---

Test script:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B")
model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3.1-8B")

inputs = tokenizer(["The quick brown"], return_tensors="pt")
gen_out = model(**inputs)

Before:

We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/internal/generation_utils#transformers.Cache)

Now: no warning :)

[gante]

bumped the deprecation to v4.47, we some key models like T5 are still missing

[gante]

copy/paste from llama

(on some models, this pattern was slightly different)

[gante]

Note: not self.training was removed.

If we are training and we pass past_key_values as tuple of tuples, we definitely want to see the warning -- the code will break in the near future

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[zucchini-nlp]

Thanks for fixing, this is way much better than checking for self.training

[LysandreJik]

Thanks Joao!

[LysandreJik]

(nit not really related to the PR but to the link which was already here before)

Linking to the Cache class is cool but you have to scroll down a bit to see an example. Would it be possible to link to a migration doc/example showcasing how a previously written code with past key values as a tuple of tuples can be adapted to be sent to the model?

The more copy-pastable the example, the less friction there will be here

[gante]

@LysandreJik good point!

I've added a tiny section to our cache docs about the legacy cache and how to convert it to/from the new format, with an example (cc @zucchini-nlp). This warning now points to that section in the docs.

(will merge after confirming the docs with the doc builder)

EDIT: for some reason, the doc builder is not updating its contents, despite the doc job being successful 🤔 I'm going to merge and double-check the merged results

EDIT2: it worked :) https://huggingface.co/docs/transformers/main/en/kv_cache#legacy-cache-format