What does this PR do?

BioGptForSequenceClassification is the last decoder classification head still using the legacy pooling index:

sequence_length = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1)
pooled_logits = logits[torch.arange(batch_size), sequence_length]

(num_non_pad - 1) equals the index of the last real token only when padding is on the right. With left padding the real tokens are shifted to the end, so this selects a token in the front of the sequence and the head pools the wrong position. The classification logits and loss are then silently wrong, with no error or warning. Right padding is unaffected.

#35911 moved every other ForSequenceClassification decoder (gpt2, llama, ...) onto a left/right-padding-safe selection, but biogpt was missed. This PR brings biogpt in line by using the same block as gpt2:

non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32)
last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)

which takes the rightmost non-pad index and is correct for both padding sides. This is the full block from gpt2, so it also adds the same raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") guard for pad_token_id is None with batch_size > 1 (previously biogpt silently pooled index -1); this matches every other decoder after #35911. The edit is in modular_biogpt.py, regenerated into modeling_biogpt.py with the modular converter.

biogpt is the only remaining model with this pattern (grep "torch.ne(input_ids, self.config.pad_token_id).sum" matches biogpt only).

import torch
from transformers import BioGptConfig, BioGptForSequenceClassification, set_seed

torch.set_grad_enabled(False)
cfg = BioGptConfig(vocab_size=100, hidden_size=32, num_hidden_layers=2,
                   num_attention_heads=2, intermediate_size=64, num_labels=3, pad_token_id=0)
set_seed(0)
model = BioGptForSequenceClassification(cfg).eval()

set_seed(1)
real = torch.randint(1, cfg.vocab_size, (1, 5))          # 5 real tokens
input_ids = torch.cat([torch.zeros(1, 3, dtype=torch.long), real], dim=1)  # 3 left pads
mask = (input_ids != 0).long()

pooled = model(input_ids, attention_mask=mask).logits     # head's pooled logits
hidden = model.biogpt(input_ids, attention_mask=mask).last_hidden_state
per_pos = model.score(hidden)                             # same-forward per-position logits

last_real = input_ids.shape[1] - 1                        # index 7
print("d(pooled, last_real) =", (pooled - per_pos[:, last_real]).abs().max().item())

d(pooled, last_real) = 0.23832...

d(pooled, last_real) = 0.0

Added regression test test_biogpt_sequence_classification_left_padding: it asserts pooled_logits equals the last-real-token logits from the same forward. It fails on the current code and passes with the fix. Full tests/models/biogpt/test_modeling_biogpt.py passes.

Code Agent Policy

• I confirm that this is not a pure code agent PR.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Did you write any new necessary tests?

Who can review?

cc @Rocketknight1