Module expert_pool 

MoE expert residency pool (TIDE-style predictive offload).

Mirrors the policy in ims-kdks/TIDE LLaDA2MoeSparseMoeBlock: rank experts by token hits, refresh placement every τ steps, paired promote/demote to limit PCIe churn.

Router logits and expert indices are unchanged — placement only.

Structs

ExpertPool

Tracks which logical experts are GPU-resident and applies TIDE placement updates.

ExpertPoolConfig

Configuration for ExpertPool.

ExpertPoolStats

Cumulative counters (TIDE offload_stats).

ExpertRefreshResult

Result of one placement refresh.

Enums

ExpertRefreshPolicy

When to re-run hit counting and expert placement.

MoEExecMode

Per-forward hint from the runner (maps to TIDE refresh_experts).

Functions

gpu_expert_budget_from_vram

merged_resident_mask

Union of GPU-resident experts across per-layer pools (legacy single graph mask).

per_layer_resident_masks

Per-layer resident bitmasks (TIDE placement; one row per MoE FFN in forward order).