MoE Router
moe_router stresses sparse Top-K-style routing and gated accumulation, which
are local execution components of mixture-of-experts inference.
| Area | What it tests |
|---|---|
| Routing | Deterministic sparse expert-selection pressure. |
| Compute | Gated projection-style accumulation. |
| Memory | Synthetic expert-state working-set reads. |
| Result | Synthetic routed tokens per second. |
This is a single-GPU routing diagnostic. It does not include multi-GPU all-to-all communication or a framework-specific MoE implementation.