Optimizing MoE Inference: Selecting Experts per Document, Not Token

nbvehrfr · reddit · 2026-08-25

The author challenges the current MoE approach where every token independently selects from all experts, leading to excessive VRAM usage during inference. They propose an alternative: use a small draft model to read the first 100 tokens of a document, predict a relevant expert pool, load that pool into VRAM, and keep it resident. Token-level routing would then occur only within that specific pool, arguing that semantic domains in long-context workloads don't teleport every few milliseconds.

Original post →

More from Infra

Infra channel →