A member of the development team told us: "We've optimized a series of key operators, such as GeMM and MoE, at the instruction level to achieve native processing capabilities for FP8 data."

→ View original post on X — @jiqizhixin