Worked on the liguodongiot/transformers repository, focusing on stability and performance improvements rather than new feature development. Addressed a critical bug in the handling of Bad Words IDs bias during multi-processing, which previously led to incorrect bias application and performance slowdowns, particularly on MPS hardware. The solution involved optimizing the data processing path to ensure correct bias handling and improved throughput for production workloads. Utilized Python and applied expertise in machine learning and performance optimization to resolve the issue. This work laid the foundation for further multiprocessing enhancements and contributed to more reliable and efficient model inference in the repository.