Worked on performance optimization for the safe_softmax function in the FlagOpen/FlagGems repository, focusing on large tensor workloads in deep learning applications. Leveraged Python and GPU programming to introduce dynamic block and warp scaling, enabling efficient processing and online chunked normalization for infinite sequence support. Isolated FP32 Flash softmax optimizations specifically to the Turing backend, maintaining a generic baseline for compatibility while achieving targeted performance gains. Refactored optimization paths and improved code maintainability by aligning with pre-commit and flake8 standards. The work resulted in improved throughput, reduced latency, and a more maintainable codebase for high-performance deep learning operations.