Add position_ids in XLMRobertaXLForCausalLM.prepare_inputs_for_generation

[ydshieh]

What does this PR do?

The model component XLMRobertaXLEmbeddings will perform

position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length)

if position_ids is not passed to the model (say XLMRobertaXLForCausalLM).

However, during the decoding, even if past_key_values_length is passed, the create_position_ids_from_input_ids will compute the wrong position_ids when the input_ids is only the single last token. This is because create_position_ids_from_input_ids relies the whole input_ids in order to get the number of padding tokens:

mask = input_ids.ne(padding_idx).int()

(since past_key_values_length doesn't contain this information.)

In general, we need to prepare position_ids in prepare_inputs_for_generation, like what has been done in

GenerationMixin.prepare_inputs_for_generation

This PR adds this for XLMRobertaXLForCausalLM (whose prepare_inputs_for_generation is already overwritten from the parent class)

---

As a bonus:

tests/models/xlm_roberta_xl/test_modeling_xlm_roberta_xl.py::XLMRobertaXLModelTest::test_assisted_decoding_matches_greedy_search_1_same

is also no longer flaky (running 10000 times) while it was failing 5% of the time before this PR.

[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]

Oh yeah, we have some inconsistencies with how position ids are created in models. I've been trying to create position ids in all models in the same way, but it got too big and complicated. We'll need to come back to position ids later at some point

For now LGTM, thanks!

[ArthurZucker]

Looks good to me, let's cleanup a bit before merging 😉