执行摘要
修正分离异步模式吞吐量 GPU 计数
PR body 指出:In separate-async mode, the standalone rollout runs on dedicated GPUs that are not part of the trainer resource pool. However, compute_throughout_metrics used resource_pool_manager.get_n_gpus() as the denominator, which only counts trainer-side GPUs. This caused perf/throughput to be inflated in separate-async mode, making it incomparable with sync and colocate-async modes.
建议合入。逻辑清晰,修改量小,且修复了明确的数据错误。值得关注的是通过模板方法模式(基类默认 + 子类重写)分离关注点,是平滑扩展的良好实践。未来其他异步模式如有类似需求,也应遵循此模式。
无人工审核评论。审核者 wuxibin89 已批准。仅 gemini-code-assist[bot] 自动评论总结代码变更,未提出具体问题。
参与讨论