transformers
Fix contrastive search to correctly handle input with padding
#33507

Merged

Commits

fix: handle padding in contrastive search for decoder-only models

ducviet00 committed 1 year ago
fix: handle padding in contrastive search for encoder-decoder models

ducviet00 committed 1 year ago
tests: move padding contrastive test to test_util, add t5 test

ducviet00 committed 1 year ago
fix: handle if model_kwargs["decoder_attention_mask"] is None

ducviet00 committed 1 year ago
refactor: improve padding input contrastive search generation tests

ducviet00 committed 1 year ago
chore: _ranking_fast to use LongTensor for cosine_matrix_mask

ducviet00 committed 1 year ago