Running async RL on my DGX Spark cluster
A practical walkthrough of turning a DGX Spark cluster into a small async RL engine, and optimizing weight sync 131x and RL step time 6.4x.
Recipes
Reproducible recipes for inference, training, and frontier evals.
Optimizing throughput for async RL
What I learned serving DeepSeek-V4-Flash as an RL rollout engine on 2 DGX Spark machines.