Integrate Updates to BeamSearchScorer

[sramshetty]

CoCa generate() fails on transformers>=4.31.0 due to fixes in BeamSearchScorer.

Specifically, the following snippet:

model, _, transform = open_clip.create_model_and_transforms(
  model_name="coca_ViT-L-14",
  pretrained="mscoco_finetuned_laion2B-s13B-b90k"
)
PIL_image = Image.open(im_path).convert("RGB")

im = transform(PIL_image).unsqueeze(0)
with torch.no_grad(), torch.cuda.amp.autocast():
  generated = model.generate(im)

results in:

in _generate_beamsearch(self, image_inputs, pad_token_id, eos_token_id, sot_token_id, num_beams, num_beam_groups, min_seq_len, stopping_criteria, logit_processor, logit_warper)
    329 
    330         if num_beams * batch_size != batch_beam_size:
--> 331             raise ValueError(
    332                 f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}."
    333             )

ValueError: Batch dimension of `input_ids` should be 18, but is 6.

This PR just ports over their changes in generation utils.

[rwightman]

@sramshetty will previous releases (but still somewhat recent) ver of transformers fail with this?

[sramshetty]

@rwightman I was able to get it working by downgrading to transformers==4.30.0. The changes were introduced in 4.31.0, so I suspect any version before that would also work, but haven't verified that.