Transformers 4.35: safetensors by default, Distil-Whisper, Fuyu
Hugging Face Transformers v4.35 made safetensors the default serialization format and added models including Distil-Whisper (6× faster than Whisper) and Adept's Fuyu-8B.
What happened
Transformers v4.35.0 switched to safetensors as the default way to save models and added several new models.
The technical details
From the release notes:
- Distil-Whisper: a distilled Whisper that’s 6× faster, 49% smaller, and within 1% word error rate on out-of-distribution data. It copies Whisper’s entire encoder, keeping its robustness, but only 2 decoder layers, which speeds up token-by-token generation. It’s MIT licensed, with Flash Attention 2 and speculative decoding support.
- Fuyu-8B (from Adept): a decoder-only multimodal model. A linear layer embeds image patches, which are then treated like text tokens, with a special image-newline token marking row ends.
- Also added: Kosmos-2, SeamlessM4T and Owl-v2.
Why it matters
The safetensors docs describe the format as a safe alternative to pickle-based checkpoints, which can execute code when loaded. Making it the default is a security improvement for everyone downloading models. I ran into the same issue during my NLP phase, when a pickle-format model download was blocked and I switched to plain-text GloVe vectors.
#hugging-face#transformers#release
Sources
- Transformers v4.35.0 release notes — GitHub · huggingface/transformers
- Gandhi, von Platen & Rush (2023), Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling — arXiv
- Safetensors documentation — Hugging Face
This is a summary written from the linked public sources. Figures and numbers are as reported there.