transformers

625ad844 - Switch gradient checkpointing default to `use_reentrant=False` (PyTorch recommended) (#43203)

Switch gradient checkpointing default to `use_reentrant=False` (PyTorch recommended) (#43203) * Switch gradient checkpointing default to use_reentrant=False (PyTorch recommended) * All testers * revert * skip for the right reason * skip for right reason and this one is fixed * skip for the right reason * fix up to blip2 excluded * reactivate all tests * skip non trainable and xfail know issues * audio encoder after lm initialization and add xfail tests for gradient checkpointing issues * continue fixing tests * some last cases * fix check_training_gradient_checkpointing * ignore and x fail other tests * this module doesn't support training * forgot one * mra * style * fix recurrent gemma when enable input require grads with gradient checkpointing * mark training test as expected failure for reentrant compatibility * style * reverting, moshi won't work anyway * remove skip training tests for Siglip2VisionModel --------- Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>