●Change8

Migrating to Transformers v5.0.0rc2

Version v5.0.0rc2 introduces 3 breaking changes. This guide details how to update your code.

Released: 1/8/2026

3
Breaking Changes
5
Migration Steps
9
Affected Symbols

⚠️ Check Your Code

If you use any of these symbols, you need to read this guide:

AutoTokenizerTokenizersBackendBatchFeatureConvNeXtImageProcessorEfficientNetImageProcessorMobileViTImageProcessorconvert_segmentation_map_to_binary_masks_init_weightspost_init

Breaking Changes

●Issue #1

The default image interpolation method for ConvNeXt image processor has changed, potentially affecting image preprocessing results.

✓Solution

If you were relying on the previous default interpolation for ConvNeXt, explicitly set the desired interpolation method (e.g., 'BILINEAR') when initializing the processor.

●Issue #2

The default image interpolation method for EfficientNet image processor has changed, potentially affecting image preprocessing results.

✓Solution

If you were relying on the previous default interpolation for EfficientNet, explicitly set the desired interpolation method (e.g., 'BILINEAR') when initializing the processor.

●Issue #3

The default image interpolation method for MobileViT image processor has changed, potentially affecting image preprocessing results.

✓Solution

If you were relying on the previous default interpolation for MobileViT, explicitly set the desired interpolation method (e.g., 'BILINEAR') when initializing the processor.

Migration Steps

  1. 1
    Update your library to the latest version.
  2. 2
    Review any code that uses `AutoTokenizer` as the library now enforces `tokenizer_class` more strictly when using `TokenizersBackend`.
  3. 3
    If you are using image processors for ConvNeXt, EfficientNet, or MobileViT and notice changes in output due to interpolation, explicitly set the `interpolation` argument during processor initialization to match your previous behavior (e.g., set it to 'BILINEAR' if you expect the old behavior, or 'BICUBIC' if you want the new default).
  4. 4
    If you are working with MoEs, be aware that performance optimizations have been introduced via batched/grouped experts and optimized decoding, which should generally improve performance without code changes.
  5. 5
    Review any custom weight initialization logic, as there are core changes enforcing `post_init` calls and better handling of buffers and tied weights during initialization.

Release Summary

This release focuses on fixing AutoTokenizer enforcement, optimizing MoE performance with batched implementations, and significantly improving model loading speeds via meta device initialization.

Need More Details?

View the full release notes and all changes for Transformers v5.0.0rc2.

View Full Changelog