Verified, then logged.
Sun 22 Nov 2026 · 00:14 KST / 15:14 UTC한국어 | ENSearch /

Breaking02 / IT

Transformers v5.20.0 breaks code with deprecated cache calls

Code that calls the removed cache methods needs changes, and the release notes do not say how far expert-parallel inference has slowed.

Past the blurred back of a developer in warm lamplight, a monitor shows code with changed lines marked in red and green beside an open workstation holding several graphics cards. AI-generated image

Hugging Face released Transformers v5.20.0 on 16 Oct 2026, changing cache and configuration compatibility. Users of deprecated cache methods have to change their code, the release notes said.

New models include Meta's Muse Spark Open, InclusionAI's Ling 3.0 and Aleph Alpha's Kolibri 1, according to the release notes. The export tools now support quantisation directly. Quantisation is a technique for making models lighter.

Compatibility. The release notes said deprecated methods have been removed from cache_utils.py. Code that called them has to move to the current cache API. The minimum kernels version is now 0.17.2.

Names imported through other model files should now come from their original modules, the release notes said. Custom nested configurations may need updating to match SubConfigSpec. The LUKE and mLUKE tokenizers now treat entity tokens as special tokens.

Expert parallelism (EP), which spreads a model's experts across devices, carries a separate warning. The release notes said "inference with EP has regressions in speed" and that a fix is under way.

Runtime. v5.20.1, released the same day, fixed a loading problem on older torch versions, according to its release notes. Distributed checkpointing now requires torch 2.8 or later. The version also adds Reflection AI's Beam, described as a coding model for running on a team's own hardware.

The two sets of release notes do not give the licence or commercial terms of the new models, or the size of the EP slowdown.