Cover Image

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

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Wu, Hanjiang, Bambhaniya, Abhimanyu Rajeshkumar, Banerjee, Sarbartha, Khare, Tuhin, Srinivasan, Sudarshan, Subramanian, Suvinay, Kundu, Souvik, Kumar, Madhu, Elavazhagan, Midhilesh, Won, William, Yazdanbakhsh, Amir, Krishna, Tushar
Format:	Preprint
Published:	2026
Subjects:	Machine Learning Artificial Intelligence Distributed, Parallel, and Cluster Computing
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

MIST: A Co-Design Framework for Heterogeneous, Multi-Stage LLM Inference
by: Bambhaniya, Abhimanyu Rajeshkumar, et al.
Published: (2025)

Demystifying AI Platform Design for Distributed Inference of Next-Generation LLM models
by: Bambhaniya, Abhimanyu, et al.
Published: (2024)

Progressive Gradient Flow for Robust N:M Sparsity Training in Transformers
by: Bambhaniya, Abhimanyu Rajeshkumar, et al.
Published: (2024)

TACOS: Topology-Aware Collective Algorithm Synthesizer for Distributed Machine Learning
by: Won, William, et al.
Published: (2023)

Analytical Provisioning for Attention-FFN Disaggregated LLM Serving under Stochastic Workloads
by: Song, Chendong, et al.
Published: (2026)

Revealing the Challenges of Attention-FFN Disaggregation for Modern MoE Models and Hardware Systems
by: Liu, Guowei, et al.
Published: (2026)

ASTRA-sim2.0: Modeling Hierarchical Networks and Disaggregated Systems for Large-model Training at Scale
by: Won, William, et al.
Published: (2023)

RAGO: Systematic Performance Optimization for Retrieval-Augmented Generation Serving
by: Jiang, Wenqi, et al.
Published: (2025)

Trinity: Disaggregating Vector Search from Prefill-Decode Disaggregation in LLM Serving
by: Liu, Yi, et al.
Published: (2025)

ModServe: Modality- and Stage-Aware Resource Disaggregation for Scalable Multimodal Model Serving
by: Qiu, Haoran, et al.
Published: (2025)

COMET: A Comprehensive Cluster Design Methodology for Distributed Deep Learning Training
by: Kadiyala, Divya Kiran, et al.
Published: (2022)

BestServe: Serving Strategies with Optimal Goodput in Collocation and Disaggregation Architectures
by: Hu, Xiannan, et al.
Published: (2025)

DynaServe: Unified and Elastic Execution for Dynamic Disaggregated LLM Serving
by: Ruan, Chaoyi, et al.
Published: (2025)

P/D-Serve: Serving Disaggregated Large Language Model at Scale
by: Jin, Yibo, et al.
Published: (2024)

Janus: Disaggregating Attention and Experts for Scalable MoE Inference
by: Zhang, Zhexiang, et al.
Published: (2025)

Lotus: Optimizing Disaggregated Transactions with Disaggregated Locks
by: Hu, Zhisheng, et al.
Published: (2025)

MemServe: Context Caching for Disaggregated LLM Serving with Elastic Memory Pool
by: Hu, Cunchen, et al.
Published: (2024)

StreamServe: Adaptive Speculative Flows for Low-Latency Disaggregated LLM Serving
by: Kumar, Satyam, et al.
Published: (2026)

Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving
by: Qin, Ruoyu, et al.
Published: (2024)

Observation, Not Prediction: Conversation-Level Disaggregated Scheduling for Agentic Serving
by: Ding, Jianru, et al.
Published: (2026)

Efficient Multi-round LLM Inference over Disaggregated Serving
by: He, Wenhao, et al.
Published: (2026)

Efficiently Serving Large Multimodal Models Using EPD Disaggregation
by: Singh, Gursimran, et al.
Published: (2024)

DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving
by: Zhong, Yinmin, et al.
Published: (2024)

EPD-Serve: A Flexible Multimodal EPD Disaggregation Inference Serving System On Ascend
by: Bai, Fan, et al.
Published: (2026)

Efficient MoE Inference with Fine-Grained Scheduling of Disaggregated Expert Parallelism
by: Pan, Xinglin, et al.
Published: (2025)

Revisiting Disaggregated Large Language Model Serving for Performance and Energy Implications
by: Li, Jiaxi, et al.
Published: (2025)

Not All Prefills Are Equal: PPD Disaggregation for Multi-turn LLM Serving
by: Li, Zongze, et al.
Published: (2026)

Nexus:Proactive Intra-GPU Disaggregation of Prefill and Decode in LLM Serving
by: Shi, Xiaoxiang, et al.
Published: (2025)

Disaggregating Time-Series with Many Indicators: An Overview of the DisaggregateTS Package
by: Mosley, Luke, et al.
Published: (2023)

TokenScale: Timely and Accurate Autoscaling for Disaggregated LLM Serving with Token Velocity
by: Lai, Ruiqi, et al.
Published: (2025)

MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism
by: Zhu, Ruidong, et al.
Published: (2025)

Prefill-Decode Aggregation or Disaggregation? Unifying Both for Goodput-Optimized LLM Serving
by: Wang, Chao, et al.
Published: (2025)

HydraInfer: Hybrid Disaggregated Scheduling for Multimodal Large Language Model Serving
by: Dong, Xianzhe, et al.
Published: (2025)

RAPID-Serve: Resource-efficient and Accelerated P/D Intra-GPU Disaggregation
by: Masood, Amna, et al.
Published: (2026)

SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM Serving
by: Guo, Yipin, et al.
Published: (2026)

vLLM-Omni: Fully Disaggregated Serving for Any-to-Any Multimodal Models
by: Yin, Peiqi, et al.
Published: (2026)

BanaServe: Unified KV Cache and Dynamic Module Migration for Balancing Disaggregated LLM Serving in AI Infrastructure
by: He, Yiyuan, et al.
Published: (2025)

BanaServe: Unified KV Cache and Dynamic Module Migration for Balancing Disaggregated LLM Serving in AI Infrastructure
by: Yiyuan He, et al.
Published: (2026)

DOPD: A Dynamic PD-Disaggregation Architecture for Maximizing Goodput in LLM Inference Serving
by: Liao, Junhan, et al.
Published: (2025)

SUN: Shared Use of Next-token Prediction for Efficient Multi-LLM Disaggregated Serving
by: Woo, Sunghyeon, et al.
Published: (2026)