Contributed to the FlagOpen/FlagGems repository by developing a dynamic block-size padding optimization for the AIPU backend, replacing hardcoded values with a runtime-tuned configuration to improve performance for small tensor workloads. Addressed edge cases by introducing logic for small-tensor size handling and ensured code quality through formatting updates. Additionally, stabilized the test suite by resolving issues with remainder and floor division test handling and fixing a KeyError in test reporting logic, which enhanced test reliability and CI feedback. Work was primarily implemented in Python, leveraging Triton, YAML, and pytest to improve backend flexibility and repository stability over two months.