Research

1 min read

We let the harness redesign our MoE dispatch. It removed the all-to-all.

Decode latency cut 34% at EP=32. We didn't suggest it, and for two days we didn't believe it.

Vladimir Zeljkovic

The all-to-all in expert-parallel MoE dispatch is one of those things everyone treats as load-bearing. The harness didn’t, and it was right.

The proposal

Given the routing statistics of the actual workload, most tokens reach a small, stable set of experts. The rewrite replaced the global exchange with a locality-aware two-stage dispatch and a fallback path for the tail.

Why we didn’t believe it

A 34% decode latency cut at EP=32 is the kind of number that usually means a broken benchmark. It took two days of re-running with different routing seeds, different batch shapes and a numerics audit before we accepted it.

Be there when
the next world opens.

Book a call