transformers

52297584 - fix: Load model with co-located adapter from local path (Granite Speech) (#43781)

fix: Load model with co-located adapter from local path (Granite Speech) (#43781) * fix: Only overwrite the pretrained_model_name_or_path if needed with adapter The check is based on the assumption that if the current value is a path on disk and there is a `config.json` present in that path, the path points to a full model checkpoint with an embedded adapter. https://github.com/huggingface/transformers/issues/43746 Branch: PeftLocalCheckpoint-43746 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * test: Add a unit test to validate that the path in the adapter config does not override https://github.com/huggingface/transformers/issues/43746 Branch: PeftLocalCheckpoint-43746 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * feat(peft, pipelines): Apply on-disk conditional to direct peft integration and pipelines init https://github.com/huggingface/transformers/issues/43746 Branch: PeftLocalCheckpoint-43746 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * test: Add unit tests for peft_integration and pipeline local embedded adapters https://github.com/huggingface/transformers/issues/43746 Branch: PeftLocalCheckpoint-43746 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> AI-usage: draft * styld: Comment improvements based on review feedback https://github.com/huggingface/transformers/issues/43746 Branch: PeftLocalCheckpoint-43746 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> AI-usage: none * Apply suggestion from @Rocketknight1 --------- Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>