Expand support for Collator caller functions

[kmehant]

At this point, fms-acceleration patches only torch_call function, however there are standard collators such as DataCollatorForSeq2Seq which do not implement torch_call funciton however use the standard __call__.

see: https://github.com/huggingface/transformers/blob/4d5b45870411053c9c72d24a8e1052e00fe62ad6/src/transformers/data/data_collator.py#L585

We need to update the torch call patch function.

fms-acceleration/plugins/attention-and-distributed-packing/src/fms_acceleration_aadp/aadp_utils.py

Line 64 in 3cf092a

_old_collate_torch_call = collate_fn.torch_call

I am happy to raise a PR.

[fabianlim]

@kmehant the intent here is only to mimic the recent version of trl.DataCollatorForCompletionOnlyLM which has the remove padding logic

The logic to restrict the two data collators only come from the current implementation of fms-hf-tuning:

• in pretokenized case the collator is DataCollatorForSeq2Seq. in this case we replace with DataCollatorWithFlattening
• otherwise the collator is DataCollatorForCompletionOnlyLM. in this case we activate the remove padding logic

So why do you expect to have other other collator?

[kmehant]

Thanks @fabianlim

Makes sense!