●Breaking
HuggingFace transformers v5.15.0 — Four API Breaks: Linear Attention, Cache Cropping, T5 Attention Backend, Multimodal Processor Helpers
What changed
Four behavior changes shipped in v5.15.0 that break existing code in specific conditions:
- Linear attention kernels — previously selected automatically when a linear attention model was loaded; now require explicit opt-in. Code that relied on automatic kernel selection silently falls back to a different (potentially slower) backend.
- Cache cropping API —
cache_croppingnow accepts only negative values for relative offsets. Positive offset values that previously worked will raise an error.
- T5 family models — SDPA (Scaled Dot-Product Attention) and alternative attention backends are now enabled by default. Any T5 code that assumed the original attention backend will use a different implementation, which can affect numerical precision in edge cases.
- Multimodal processors — private helper functions removed from multimodal processor classes. Any code calling internal
_helpermethods on multimodal processor objects will raiseAttributeError.
TL;DR
transformers v5.15.0 removes automatic linear attention kernel selection, restricts cache cropping offsets to negative values, flips T5 to SDPA by default, and removes private multimodal processor helpers — four targeted breaks that affect specific code paths, not the majority of users.
Developer signal
Before running
pip install transformers==5.15.0, grep your codebase for: (1) any model class using linear attention — check whether you're passing kernel parameters explicitly or relying on auto-selection; (2) any cache_cropping calls with positive offset values — flip to negative equivalents; (3) T5ForConditionalGeneration or any T5 variant in production with precision-critical outputs — run a quick regression test on your attention outputs; (4) any private method calls on multimodal processor objects (pattern: processor._something) — rewrite using the public API, which exposes the same functionality. None of these breaks affect the majority of transformers users (transformer-style attention, non-T5 models, no custom cache manipulation, no internal processor methods), but all four are silent or error-raising rather than deprecation-warned. Full new features in the Tooling section.Affects you ifYou use linear attention models, custom cache cropping logic, T5-family models with precision-critical outputs, or internal multimodal processor helpers.EffortModerate (audit the four affected code patterns before upgrading; otherwise a straightforward version bump).
HuggingFace / GitHub | Date: August 10, 2026 | Link: https://github.com/huggingface/transformers/releases/tag/v5.15.0https://github.com/huggingface/transformers/releases/tag/v5.15.0