Developed a high-performance softmax computation path for the FlagOpen/FlagGems repository, focusing on both speed and numerical stability in deep learning inference. The work centered on implementing a fused Triton kernel with fp32 accumulation, supporting FP16, BF16, and FP32 formats to enhance throughput and reliability across various floating-point precisions. Leveraging expertise in GPU programming, machine learning, and numerical optimization, the developer improved inference latency and laid the foundation for future kernel enhancements. The project was executed in Python, with attention to code quality through consistent linting and formatting practices, resulting in more maintainable and robust production code for softmax operations.