Vector Institute Demystifies MoE: Slashes Logit Memory from 23.3GB to 0.3GB

VectorInst · x · 2026-07-30

The Vector Institute detailed the internal mechanics and practical deployment optimizations for Mixture-of-Experts (MoE) models, as part of the AIXPERT initiative for explainable AI.

To reduce massive memory footprints, the team stacked three crucial optimizations: keeping each MoE layer intact without sharding, applying 4-bit quantization, and utilizing Flash Attention 2. A forward hook was also introduced to cut logit memory drastically from 23.3 GB down to 0.3 GB. Removing any one of these three causes the pipeline to fail.

With this setup, the team audited the model's learned routing behavior, revealing how capacity is allocated and driven across its layers.

Original post →

More from Infra

Infra channel →