Verified, then logged.
Sun 22 Nov 2026 · 00:20 KST / 15:20 UTC한국어 | ENSearch /

Breaking02 / IT

Transformers v5.21.0 forces changes to custom parallel plans

Custom tensor-parallel and mixture-of-experts plans face changes, while the release notes give no licence or commercial terms for the new models.

Identical accelerator cards sit side by side in an open server linked by thick cables, while an out-of-focus engineer seen from behind moves one of the connections. AI-generated image

Hugging Face released Transformers v5.21.0 on 27 Oct 2026, changing how parallelisation plans behave, so developers with custom plans may need to revise them, according to the release notes.

The release notes introduce new models including MiniCPM-OCR, for reading documents, and the language model Flint. Pocket TTS, a text-to-speech model, supports voice cloning from a recording. Gradium Phonon TTS voices come only as precomputed latents, numerical encodings of a voice prepared in advance, and cannot be cloned from a recording.

Compatibility. The release notes describe a dedicated embedding style for tensor parallelism (TP), which splits model weights across devices. ColwiseParallel and RowwiseParallel now apply only to linear layers. Custom TP plans that used them on embeddings must move to the new style.

The notes also record a change for mixture-of-experts (MoE) models, which divide work among specialised sub-networks called experts. TP and expert parallelism (EP) plans now all split the experts across devices. Custom and downstream MoE plans must add this split to pass a new test.

The release notes state that "no user action is needed" after a fix to Gemma4 processor imports.

Runtime. The minimum Python version has risen to 3.11, according to the release notes. Older environments need a Python upgrade first. Gradium Phonon TTS is presented as a family of on-device models.

The release notes do not give licence or commercial terms for the new models.