Transformers 5.20.0 Released, Removes Legacy Cache Interfaces and Warns of Expert Parallel Inference Performance Regression

Hugging Face released Transformers 5.20.0 on October 15. The package was made available on PyPI the same day, and the commit listed in the release attestation matches the GitHub version tag. For engineering teams maintaining existing models, the most immediate effects of this update concern compatibility and distributed execution behavior, so the actual inference path should be checked before upgrading. Package and release attestation

In the breaking changes section, the project explicitly warns of a performance regression in expert parallel (EP) inference and says the team is working on it. This adds a validation task for deployments using mixture-of-experts models: even though the new version fixes some expert dispatch logic, that does not establish that overall throughput will improve. The release notes do not specify which models or hardware are affected, or the size of the regression, so the actual risk must be assessed with your own workloads. Release notes

Another migration risk comes from cache_utils.py: the new version removes deprecated checks and methods, so custom models or wrappers that still call these interfaces need to switch to the current cache API. The release also introduces SubConfigSpec for centralized initialization of sub-configurations; existing nested configurations may need adjustment. These changes can affect loading and generation workflows, so confirming that the package imports successfully is not enough to complete upgrade validation. Breaking changes

For deployment, the new version says quantization can be applied directly in export workflows for Dynamo, ONNX, OpenVINO, and ExecuTorch. It integrates PT2E, ONNX Runtime, and NNCF, covering static and dynamic quantization as well as INT8/INT4 weight compression. Distributed checkpoint support is also expanded for saving and loading models and optimizers with FSDP and tensor parallelism, including conversion between different parallel configurations. Export and parallelism updates

The next engineering step is to hold the model, input lengths, and hardware constant, then compare generation outputs, time to first token, decoding throughput, and memory usage before and after upgrading. Quantized exports also require task quality checks. These validation steps are recommendations based on the scope of the changes; the official release page does not provide comparative data sufficient to characterize performance across all backends.