According to @_avichawla, DFlash speeds LLM inference 8.5x via parallel draft tokens, maintaining accuracy and integrating with vLLM, SGLang, and Transformers.