Update Configs to Keyword Arguments to Prevent Release v5.4.0 Crashes
Hugging Face · Platform Update · · notable
Briefing for: Engineering
What happened
Release v5.4.0 introduces major breaking changes, including the refactoring of `PreTrainedConfig` into dataclasses, which now strictly require keyword arguments. The image processing backend has also been unified, removing the `BaseImageProcessorFast` and `image_processing_utils_fast` modules in favor of a single unified module.
Why it matters
Existing code using positional arguments for model configurations will break upon upgrading. However, the release offers significant performance gains, including up to 30x faster FP8 matmuls for quantized models and initial support for Flash Attention 4 (FA4), alongside a requirement to upgrade to FA2 version 2.3.3 or newer.
What this enables
- If you use vision models, the unified image processor backend simplifies your import paths and reduces maintenance overhead between fast and slow implementations.
- If you run large-scale inference, the 30x faster FP8 grouped and batched matmuls significantly reduce latency for quantized workloads.
- If you deploy MoE models, the new static FP8 expert support improves multi-GPU distribution and memory efficiency.
Get personalized AI briefings for your role at Changecast →