RSS Google 开发者博客 Supercharging LLM inference on Google TPUs: Achieving 3X speedups with diffusion-style speculative decoding developers.googleblog.com
