transformers

67265ef5 - Fix `num_items_in_batch` over-counting for causal LM losses (#46204)

Fix `num_items_in_batch` over-counting for causal LM losses (#46204) * Fix `num_items_in_batch` over-counting for causal LM losses * Address review: use LOSS_MAPPING, honor pre-shifted labels, add tests - Inspect the actual loss function via LOSS_MAPPING instead of matching loss_type strings (catches CsmForConditionalGeneration etc.). - If the data collator already provides `shift_labels`, count over that tensor directly instead of slicing labels again. - Add unit tests for `_get_num_items_in_batch` covering the causal LM path (with and without pre-shifted labels) and the non-causal-LM path. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix test_train_and_predict_loss_parity --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>