$ the-wire · showcase
transformers clears out deprecated mask, cache, and rotary paths
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
10 people shipped this
Cyril Vallez
@Cyrilvallez
2 cited
albertvillanova
@albertvillanova
1 cited
Ashutosh Singh
@AshutoshSinghIntel
1 cited
Elad Segal
@eladsegal
1 cited
assafvayner
@assafvayner
Behrooz Azarkhalili
@behroozazarkhalili
adithya-s-k
@adithya-s-k
Roman Kazantsev
@rkazants
mpaulitsch
@mpaulitsch
remi-or
@remi-or
A coordinated sweep of removal PRs in transformers deletes long-deprecated internals, while TRL opens assistant-only loss to vision SFT and OpenVINO picks up Ministral-3 export support.
Remove deprecation cycle on mask and cache (#49231) huggingface/transformers
The mask and cache deprecation cycle is over: the backward-compatibility shims are gone rather than merely warned about, so any code still passing the old mask or cache arguments will break at import or call time. This is a breaking change, and it lands alongside two companion cleanups in the same release window.
Remove rotaries deprecation cycle (#49228) huggingface/transformers
Cyril Vallez removed the rotary embedding deprecation cycle in full. Code that referenced the deprecated rotary paths through the old names or aliases has nothing left to fall back on.
Support assistant-only loss for vision datasets in SFT huggingface/trl
SFTTrainer now accepts assistant_only_loss=True for vision datasets, with transformers>=5.18.0 required. Previously the trainer refused the setting because processors did not return usable assistant masks once image placeholders were expanded; the vision data collator now tokenizes conversational examples through the processor's apply_chat_template.
Support per-layer cache configuration (#48178) huggingface/transformers
Per-layer cache configuration lands, allowing heterogeneous cache layouts within a single model rather than one layout for every layer. The PR carries compatibility tests for the mixed configuration.
Add ministral3 (Ministral-3-3B-Reasoning-2512) OpenVINO export support (#1986) huggingface/optimum-intel
optimum-intel added OpenVINO export support for Ministral-3-3B-Reasoning-2512 via optimum-cli export openvino -m mistralai/Ministral-3-3B-Reasoning-2512 --task image-text-to-text. The PR ships an inference script using OVModelForVisualCausalLM. The long tail also includes an FLUX.2-klein fix that keeps the text encoder sequence length dynamic in OVDiffusionPipeline.reshape(), an OpenVINO Z-Imag...
Action items
- → Audit transformers code for deprecated mask, cache, and rotary arguments before upgrading past the removal release huggingface/transformers [plan]
