Haolin Yan (闫 浩霖)

[recipe] feat: asynchronous reward agent with mini-batch pipeline and one-step off-policy training

10

### What does this PR do? This PR introduces the **asynchronous reward agent** to schedule and mitigate communication bottlenecks in RL training scenarios that rely on remote reward services (e.g.,...

Performance issue: allreduce_benchmark slower than ncclAllReduce

1

First of all, I'd like to express my sincere gratitude to all the contributors of this repository! I'm able to run the `allreduce_benchmark` smoothly, but unfortunately, its performance is significantly...