[Doc]: important papers of attention-ffn disaggregation

[hsliuustc0106]

Paper list

I searched arXiv for Attention-FFN Disaggregation, attention FFN disaggregation, Attention/FFN disaggregated, AFD LLM serving,
disaggregated expert parallelism, disaggregating attention and experts,
core-attention disaggregation, chunked prefill, Dynamic CPP, near-data expert scheduling, hybrid TP/EP MoE scheduling, attention PIM runtime, and related operator-disaggregation terms.
As of 2026-09-14, the papers I found are:

Directly about Attention-FFN Disaggregation

1. Analytical Provisioning for Attention-FFN Disaggregated LLM Serving under Stochastic Workloads

   • arXiv: https://arxiv.org/abs/2601.21351
   • Published: 2026-01-29, latest version: v4, 2026-08-14
   • Focus: analytical provisioning for AFD, Attention/FFN ratio selection, stochastic workloads, barrier/synchronization overhead.

2. MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism

   • arXiv: https://arxiv.org/abs/2504.02263
   • Published: 2025-04-03
   • Focus: production MoE serving system that disaggregates Attention and FFN modules, with independent scaling, tailored parallelism, ping-pong pipeline parallelism, and an M2N communication library.

3. Revealing the Challenges of Attention-FFN Disaggregation for Modern MoE Models and Hardware Systems

   • arXiv: https://arxiv.org/abs/2602.09721
   • Published: 2026-02-10
   • Focus: AFD performance boundaries for MoE serving, communication roofline, HFU dead zone, imbalance penalty, hardware/model conditions where AFD helps.

4. How Far Can Disaggregation Go? A Design-Space Exploration of Attention-FFN Disaggregation for Efficient MoE LLM Serving

   • arXiv: https://arxiv.org/abs/2605.28302
   • Published: 2026-05-27
   • Focus: design-space exploration for AFD in MoE LLM serving, workload/SLO sensitivity, attention-vs-FFN GPU partitioning, prefix-KV reuse, and network simulation.

5. ASAP: A Disaggregated and Asynchronous Inference System for MoE Prefill

   • arXiv: https://arxiv.org/abs/2606.22541
   • Published: 2026-06-21
   • Focus: MoE prefill serving system that disaggregates attention and MoE stages, replaces global synchronization barriers with an asynchronous execution pipeline, and improves SLO-compliant prefill throughput.

6. Janus: Disaggregating Attention and Experts for Scalable MoE Inference

   • arXiv: https://arxiv.org/abs/2512.13525
   • Published: 2025-12-15, latest version: v4, 2026-08-11
   • Focus: separate Attention and MoE GPU worker pools, adaptive two-phase communication, activated-expert load balancing, expert placement, and independent SLO-aware resource scaling.

7. Efficient MoE Inference with Fine-Grained Scheduling of Disaggregated Expert Parallelism

   • arXiv: https://arxiv.org/abs/2512.21487
   • Published: 2025-12-25
   • Focus: FinDEP fine-grained task scheduling for disaggregated expert parallelism, including shared experts, variable task granularity, task ordering, and computation/communication overlap.

8. Toward Cost-Efficient Serving of Mixture-of-Experts with Asynchrony

   • arXiv: https://arxiv.org/abs/2505.08944
   • Published: 2025-05-13, latest version: v2, 2025-05-29
   • Focus: AMoE disaggregates attention from expert layers and uses asynchronous expert parallelism, per-layer micro-queuing, adaptive re-batching, and defragmentation to mitigate expert skew and synchronization overhead.

9. ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels

   • arXiv: https://arxiv.org/abs/2607.18002
   • Published: 2026-07-20, latest version: v2, 2026-07-21
   • Focus: shares large MoE experts across prefill/decode while disaggregating lightweight attention modules; uses adaptive persistent kernels, attention-initiated MoE communication, and cross-phase communication/computation overlap.

10. Energy-Efficient LLM Serving via Disaggregated Attention--FFN and Flexible Frequency Scaling

    • arXiv: https://arxiv.org/abs/2608.01891
    • Published: 2026-08-03
    • Focus: AFlex jointly optimizes A/F resource provisioning and independent GPU frequency scaling, with operator-level DVFS, an interleaved A/F pipeline, dynamic microbatch depth, and adaptive request batching.

11. AFD-Ledger: Deployment Provisioning for Attention--FFN Disaggregation

    • arXiv: https://arxiv.org/abs/2608.04502
    • Published: 2026-08-05
    • Focus: offline analytical provisioning for AFD and collocated deployments, jointly optimizing hardware assignment and deployment organization under fixed workload, SLO, hardware-budget, and runtime constraints.

12. MoE Expert Execution in Disaggregated LLM Serving with a High-Bandwidth ReRAM Near-Memory Architecture

    • arXiv: https://arxiv.org/abs/2608.13962
    • Published: 2026-08-14
    • Focus: AFD-oriented FFN-pool hardware architecture that keeps MoE expert weights in high-bandwidth ReRAM near-memory storage and addresses sparse expert unions, routing skew, occupancy, and weight-movement energy.

Related arXiv papers that include, use, simulate, or explore adjacent forms of AFD

13. Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding

    • arXiv: https://arxiv.org/abs/2507.19427
    • Published: 2025-07-25
    • Focus: Step-3 model-system co-design; introduces/uses Attention-FFN Disaggregation as part of the distributed inference system.

14. Frontier: Simulating the Next Generation of LLM Inference Systems

    • arXiv: https://arxiv.org/abs/2508.03148
    • Published: 2025-08-05
    • Focus: early Frontier simulator paper; includes Attention/FFN disaggregated serving as a simulated deployment mode.

15. Frontier: Towards Comprehensive and Accurate LLM Inference Simulation

    • arXiv: https://arxiv.org/abs/2605.21312
    • Published: 2026-05-20, latest version: v2, 2026-06-13
    • Focus: updated Frontier simulator; explicitly models co-location, Prefill-Decode Disaggregation, and Attention-FFN Disaggregation with role-specific workers.

16. AMMA: A Multi-Chiplet Memory-Centric Architecture for Low-Latency 1M Context Attention Serving

    • arXiv: https://arxiv.org/abs/2604.26103
    • Published: 2026-04-28, latest version: v2, 2026-04-30
    • Focus: attention-serving hardware architecture; mentions production-level attention-FFN disaggregation as motivation/context rather than being an AFD algorithm paper.

17. KernelFlume: Elastic Core-Attention Scaling for Agentic Long-Context Decoding

    • arXiv: https://arxiv.org/abs/2606.29207
    • Published: 2026-06-28
    • Focus: decode-centric separation of core attention from the stable projection/FFN path, using elastic weightless attention nodes and inter-layer pipelining; adjacent to AFD but uses a narrower core-attention boundary.

18. Efficient Heterogeneous Large Language Model Decoding with Model-Attention Disaggregation

    • arXiv: https://arxiv.org/abs/2405.01814
    • Published: 2024-05-03, latest version: v2, 2025-04-10
    • Focus: Lamina places the attention operator on cheaper memory-optimized devices while retaining the rest of the model on high-end accelerators; relevant to heterogeneous Attention/model disaggregation.

19. Injecting Adrenaline into LLM Serving: Boosting Resource Utilization and Throughput via Attention Disaggregation

    • arXiv: https://arxiv.org/abs/2503.20552
    • Published: 2025-03-26
    • Focus: offloads part of decode attention to prefill instances to improve utilization; related operator offloading rather than complete Attention-FFN disaggregation.

20. Beyond Prefill-Decode Disaggregation: Dissecting LLM Inference for Heterogeneous Platforms via Dynamic Operator Scheduling

    • arXiv: https://arxiv.org/abs/2607.25498
    • Published: 2026-07-28; to appear in MICRO 2026
    • Focus: DOPS dynamically places Attention, FFN, linear, and other operators across heterogeneous NPU-PIM devices based on workload and hardware state; it explores A/F and finer operator mappings rather than a fixed AFD worker-pool architecture.

21. When Does Disaggregation Pay? Simulating Prefill--Decode--Attention--FFN Specialization for Agentic LLM Inference

    • arXiv: https://arxiv.org/abs/2608.03741
    • Published: 2026-08-04
    • Focus: HeteroPanacea simulates heterogeneous four-way Prefill/Decode/Attention/FFN specialization; it includes AFD as part of a broader PDAF hardware-design space rather than implementing a fixed A/F worker-pool serving system.

22. Q-First: Attention and Feed-Forward Concurrency at the Smallest Change to the Block

    • arXiv: https://arxiv.org/abs/2608.15473
    • Published: 2026-08-16
    • Focus: changes the query/sub-layer dependency to overlap attention-side KV-cache work with projection/FFN computation on disaggregated devices; adjacent to AFD because it uses a narrower attention-sweep-versus-model-compute boundary.

23. Simthesizer: An Agent-Driven Simulation Framework for LLM Serving Systems

    • arXiv: https://arxiv.org/abs/2608.24650
    • Published: 2026-08-25, latest version: v2, 2026-08-26
    • Focus: presents the Borg composable simulator and a coding-agent-based extension workflow for rapidly evolving LLM serving mechanisms. The paper explicitly identifies Attention--FFN disaggregation as a motivating non-monolithic architecture, but its checked-in built-ins cover P/D disaggregation and expert-parallel MoE rather than native AFD; it is therefore adjacent simulator infrastructure, not an AFD implementation.

24. VPP: Virtual Pipeline Parallelism for Efficient Chunked Prefill in Long-Context LLM Inference

    • arXiv: https://arxiv.org/abs/2608.26523
    • Published: 2026-08-27
    • Focus: compares fixed-size chunked prefill pipeline parallelism with Dynamic CPP and introduces virtual stages to reduce chunk-dependent attention imbalance and pipeline bubbles. It is adjacent to AFD rather than an A/F-disaggregation design, but is directly relevant to prefill pipeline baselines and chunk-prefill studies; the evaluation includes MoE models and DeepSeek-V3.1 on Ascend NPUs.

25. DynaNDE: Dynamic Near-Data Expert Scheduling for Batched MoE Inference

    • arXiv: https://arxiv.org/abs/2609.00407
    • Published: 2026-08-31
    • Focus: dynamically schedules experts per layer across heterogeneous NPU and near-data-processing resources using an analytical model of data movement, expert concurrency, communication/computation overlap, and temporal expert reuse. It does not disaggregate Attention and FFN pools, so it is adjacent rather than direct AFD work, but its expert-side scheduling model is relevant to FFN-pool placement, batching, and imbalance studies.

26. HDA-MoE: Hybrid Parallelism and Dynamic, Adaptive Scheduling for Mixture-of-Experts with 3D Near-Memory Processing

    • arXiv: https://arxiv.org/abs/2609.08682
    • Published: 2026-09-08; accepted by IEEE TCAD
    • Focus: combines offline hybrid TP/EP mapping with online dynamic scheduling for MoE execution on heterogeneous 3D near-memory-processing resources, targeting communication cost, routing-driven load imbalance, and compute utilization. It does not separate Attention and FFN pools, so it is adjacent rather than direct AFD work, but is relevant to FFN-pool parallelism selection, hardware specialization, and expert-imbalance experiments.

27. PATTON: Enabling Commodity PIM for Production LLM Serving

    • arXiv: https://arxiv.org/abs/2609.11392
    • Published: 2026-09-10
    • Focus: integrates commodity PIM with production LLM serving engines for memory-bound decode Attention, including physical KV-block allocation, block-to-address mapping, efficient Value-cache writes, PIM command generation, prefix-cache behavior, and runtime-induced prefill recomputation. It does not implement Attention--FFN worker-pool disaggregation, so it is adjacent rather than direct AFD work, but is relevant to Attention-pool acceleration and prefix-cache experiments.

Affected area

Documentation / related work.

Suggested fix

Track the direct AFD papers separately from broader related-system papers, because some papers are about AFD itself while others only use AFD as one component, simulation mode, heterogeneous operator placement, or attention-offloading boundary.

Before submitting

• I searched existing documentation issues.
• I expanded the search beyond the literal Attention-FFN Disaggregation phrase to cover DEP, Attention/Expert disaggregation, and adjacent operator-disaggregation terminology.