Stable LatentMoE SiTU-GLU & Quantile Load Balancer

Simulate token routing across 8 experts; observe real-time quantile bias adjustments and SiTU-GLU bounding curves

Total Tokens Dispatched: 0
Expert Load & Online Quantile Routing Biases
SiTU-GLU Bounded Activation vs SwiGLU Curve

★ SiTU-GLU bounds x ∈ [-beta, beta] using tanh functions to prevent dynamic range overflow.