推理分离部署：prefill 用 GPU，decode 用 Positron 专用硬件
精选理由
Positron 把推理拆成两半：prefill 交给 GPU，decode 用自家专用硬件，生成速度更快，思路挺有意思。
用户发送 prompt 后，模型先进入 prefill 阶段生成第一个 token，再进入 decode 阶段逐 token 生成完整回答。Positron 将两个阶段拆到不同硬件上：prefill 跑在 GPU，decode 跑在自家专用推理系统。这种分离部署可以加快生成速度。
原文 · Ate-a-Pi
When you send a prompt to a model, a couple of things happen:
1. The model processes the prompt and generates the first token (prefill stage).
2. Then, the model continuously processes the prompt + new tokens until it produces the complete answer (decode stage)
These guys are doing something really cool:
They are using different hardware for each stage. They run prefill on GPUs and decode on Positron's specialized inference systems.
This results in faster generation.
