← All questions
LLM FoundationsMedium· MoE· Optimization· Routing

Why not solve for the globally optimal routing assignment?

Solving for a globally optimal routing assignment requires a solver step that incurs an O(N * T) cost per layer per forward pass. At the scale of modern model training, this overhead is non-trivial. Because the quality improvements over standard token-choice combined with auxiliary-loss balancing are marginal, no production system currently utilizes global assignment solvers.

Code playground (Python)