Worked on stabilizing bucket-based processing in the vllm-gaudi repository by addressing a critical bug in bucket setup logic. Developed and integrated the Bucket Decoding Corrector and Bucket Generation Corrections, which enforce context length truncation to the maximum allowed and apply targeted fixes within the generate_buckets function. This approach reduced edge-case errors and improved inference stability for downstream applications. Updated unit tests to ensure they accurately reflect the corrected bucket behavior, enhancing test coverage and reliability. The work leveraged Python for backend development, with a focus on performance optimization and robust unit testing to minimize production risk and improve traceability.