mixture-of-experts-adaptive-capacity

【效率与协作】Optimize Mixture-of-Experts efficiency by decoupling token-level expert activation from layer architecture. Use dynamic threshold routing where expert count per token varies by input complexity, and apply layer-wise capacity scheduling to match representational diversity patterns.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准