ROLL 是否可以加速reward模型（非LLM）计算与LLM GRPO的高效协同训练

我有个场景是用Diffusion模型作为reward，导致reward计算相对较慢，且占用额外的显卡。请问是否有方式可以加速这类的reward的计算，因为在计算reward时，训练GRPO的显卡处于等待中。

我已经做的是对Diffusion模型本身和Diffusion去噪做了加速，也做成了一个reward server、异步请求reward计算等。

No files, tests, or entry points are named. Start by tracing the Diffusion reward calculation and the GRPO training synchronization, then review the existing reward-server and asynchronous-request approach described in the issue. Done means a documented, agreed design that reduces reward-related waiting and coordinates GPU use efficiently.

Written by the indexing model from the issue text.