4x faster LLM inference (Flash Attention guy’s company)