DFLASH 2: KEEP DRAFTING PARALLEL

Summary from the TLDR Data newsletter

DFlash 2 makes speculative decoding more efficient, delivering roughly 3x faster LLM inference by accepting more tokens per pass while keeping output unchanged.