transformers

3283d5f7 - [fix] Update stale golden values and fix expected_logits shape in FlavaForPreTraining integration tests (#48639)

[fix] Update stale golden values and fix expected_logits shape in FlavaForPreTraining integration tests (#48639) - test_inference: update mmm_text/mmm_image/loss golden values (drifted since 2025-01-25) - test_inference_with_itm_labels: fix expected_logits shape from [2, 2] → [1, 2] to match the model's correct output (only positive pairs selected when itm_labels=[1,0]). The old [2,2] tensor accidentally passed via torch.allclose broadcasting (both rows were identical), but torch.testing.assert_close correctly catches the shape mismatch. Also update stale loss golden values. Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>