v5.12.0
📦 transformersView on GitHub →
✨ 3 features🐛 5 fixes🔧 6 symbols
Summary
This release introduces three major new models: MiniMax-M3-VL, PP-OCRv6, and Parakeet-RNNT. It also includes numerous bug fixes, particularly around CI stability and model generation logic.
Migration Steps
- When running a local-directory `custom_generate`, you must now explicitly require `trust_remote_code=True`.
✨ New Features
- Added support for the MiniMax-M3-VL vision-language model, featuring a CLIP-style vision tower, 3D rotary position embeddings, and a mixed dense/sparse Mixture-of-Experts decoder.
- Integrated PP-OCRv6, a lightweight OCR system with architectural innovations like a unified MetaFormer-style building block and structural reparameterization, available in medium, small, and tiny tiers.
- Introduced ParakeetForRNNT, featuring a Fast Conformer Encoder and an RNN-T (RNN Transducer) decoder using an LSTM prediction network.
🐛 Bug Fixes
- Fixed stop string matching for byte-fragment tokens.
- Made DiffusionGemma trainable.
- Removed unnecessary expand_as in get_placeholder_mask across VLMs.
- Fixed hf_hub_download not placing file in current dir for url_to_local_path.
- Ensured AutoModel honors a concrete dtype for composite checkpoints.