●Change8

Migrating to Transformers v4.55.3

Version v4.55.3 introduces 1 breaking change. This guide details how to update your code.

Released: 8/21/2025

1
Breaking Changes
4
Migration Steps
5
Affected Symbols

⚠️ Check Your Code

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

FlashAttention2FSDPGPT-OSSMambaswiglu_limit

Breaking Changes

●Issue #1

The preparation logic for FlashAttention keyword arguments (kwargs) has been reverted due to a regression introduced in a previous version.

✓Solution

If you were relying on specific kwargs preparation behavior for FlashAttention, you may need to adjust how you pass arguments, as the logic has been rolled back to a more stable state.

Migration Steps

  1. 1
    1. Update your library dependencies to version 4.55.3 or later.
  2. 2
    2. If you are using FlashAttention, review any custom logic related to passing keyword arguments, as the preparation step has been reverted.
  3. 3
    3. If you are using FSDP with generic-task models, this patch resolves sharding/runtime issues, so no code changes are typically required, but testing is recommended.
  4. 4
    4. If you are using MXFP4 integration with GPT-OSS models, ensure that the swiglu_limit is being passed correctly, as this patch fixes its propagation.

Release Summary

Patch release 4.55.3 focuses on stability improvements for FlashAttention-2 on Ascend NPU, FSDP sharding fixes, and critical bug fixes for GPT-OSS and Mamba models.

Need More Details?

View the full release notes and all changes for Transformers v4.55.3.

View Full Changelog