Devashish Lal developed quantized RMSNorm and fused normalization-quantization kernels for FP8 inference in the flashinfer-ai/flashinfer repository. Leveraging CUDA, PyTorch, and deep learning quantization techniques, he engineered a faster, more memory-efficient FP8 path by fusing normalization and quantization, reducing kernel launches and runtime overhead. His implementation supported both FP16 and FP8 with configurable scaling, and included comprehensive tests across data types and scaling modes to ensure correctness and regression safety. The work enabled seamless deployment of FP8 models through Torch compile passes, benefiting downstream consumers and laying a foundation for future FP8 enhancements and centralized numeric handling.