Author here. We were hitting tail latency and low GPU utilization issues serving SLMs via Triton.

I built a scrappy client-side router using Redis and Lua to track real-time GPU load. It boosted utilization by ~40% and improved latencies.

Happy to hear feedback on the implementation or thoughts on better ways to do this!