About

Hardware-software co-design of a custom RTL attention engine on FPGA. Implements block-based tiling and streaming online softmax to eliminate memory bandwidth bottlenecks in LLM inference.