17x is a wild improvement regardless of the baseline. Impressive results.

does this only work for vLLM or is generally applicable?

The algorithm works for MoE load balancing in general