ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving

framework 2607.00466 — Cross-paper Synthesis

ELDR (2607.00466) — L3 Per-paper Synthesis #

Mode A · target category = framework. ELDR adds expert locality as a second decode-routing axis on top of a PD-disaggregated MoE stack, losslessly cutting median TPOT 5.9–13.9%.

1. 相关论文 #

The 8 related entities do not form one tight cluster; ELDR sits at the

intersection of three lineages, and its novelty is precisely that it borrows

mechanisms from all three without belonging to any.

A — PD-disaggregation + KV-centric serving (direct ancestors).

B — decode-phase memory-bandwidth optimization (mechanism cousins).

C — MoE substrate & training-side placement (the workload & a distant echo).


2. 本篇 vs 相关论文的 delta #

New (no peer has it):

Incremental / borrowed:

Contradictory / tension:


3. 可攻击面 #


4. 生态位 #


5. 未探索方向 #

Hybrid / adaptive directions surfaced by triangulating the cluster:

  1. ELDR × CPU-offload (NEO/FastDecode) co-routing. ELDR shrinks the per-step
  2. expert union; NEO offloads the attention half to CPU

    [2411.01142]. A decoder whose expert union is small has

    spare HBM for a larger batch — exactly the batch NEO wants to feed its CPU

    pipeline. A joint scheduler that routes by expert-locality and sizes the

    CPU-offload batch per-decoder is unexplored, and directly reconciles the §2

    batch-size contradiction.

  3. Signature-aware prefill routing. ELDR reads prefill activations only at
  4. handoff for decode routing. The same signature could steer prefill worker

    choice too, unifying Mooncake's KV-affinity prefill routing

    [2407.00079] with expert-affinity into a single

    prefill+decode co-placement objective.

  5. Online / drift-adaptive centroids. Replace the static offline fit with a
  6. streaming balanced-K-means that tracks distribution shift — mirroring the

    move HybridFlow/FlexRLHF have not made either (both stay static)

    [2409.19256] [2312.11819],

    so it is an open direction for the whole placement-framework family, not just

    ELDR.

  7. Signature as a training-side load-balancer input. DeepSeek-V3's
  8. auxiliary-loss-free bias balances experts at train time

    [2412.19437]; ELDR reveals a serving-time domain

    structure. Feeding observed inference-time expert-locality clusters back into

    redundant-expert placement decisions closes a train↔serve loop neither side

    currently exploits.

  9. Batch-size regime characterization. No paper in scope charts the crossover
  10. where MoE decode transitions from union-bound (ELDR's regime) to batch-bound

    (FastDecode/NEO's regime). Mapping that curve for fine-grained MoE would tell a

    deployer when to reach for locality routing vs. offloading — the missing

    decision boundary between the two sub-clusters.