Objective

在相同 base model、prompt data、rollout backend、reward 和 sequence limits 下,比较 PPO 与 GRPO 的 runtime roles、batch fields、显存、step timing 和 metric behavior。

Controlled Variables

  • model checkpoint;
  • prompt dataset and split;
  • rollout temperature and response length;
  • reward implementation;
  • actor optimizer and learning rate;
  • hardware and training backend。

算法本身要求不同的 rollout count、critic 和 advantage estimator,这些差异作为实验变量明确记录。

Measurements

enabled roles
peak memory by role
rollout / reward / advantage / update timing
advantage mean and std
response length
reward distribution
policy KL and clip fraction
critic loss and explained variance, if present

Acceptance Criteria

  • 两份 resolved config 可直接 diff;
  • 调用链差异与测量结果一致;
  • 不用单次 reward 高低直接判断算法优劣;
  • 结果回填 09 PPO and GRPO Codepath。