CS Paper Radar

Efficient Mixture-of-Experts with Speculative Decoding via Expert Coactivation

Kumari Nishu, Han-Byul Kim, Santosh Chilkunda, Maxwell Horton, Arnav Kundu, Mohammad Samragh, Lauren Hannah, Mohammad Sekhavat, Nikhil Bhendawade, Manuel Ciosici, Iman Mirzadeh, Keivan Alizadeh Vahid, David Harrison, Irina Belousova, Mehrdad Farajtabar, Minsik Cho

cs.LG cs.CL ·submitted 2026-09-18 18:30 UTC · v1

Abstract

Mixture-of-Experts (MoE) models are increasingly deployed alongside Speculative Decoding (SD) to accelerate inference, but combining the two is challenging. SD improves the inference speed of dense models by verifying groups of tokens in parallel. However, the inference speedup for SD with MoEs depends heavily on the number of tokens being verified. Using more verification tokens results in more experts being transferred from DRAM to the Neural Processing Unit (NPU), which increases the memory transfer cost. This negatively impacts model runtime, as memory transfer is typically the bottleneck in inference. In this work, we investigate the impact of MoE router design during training on the speed of MoEs with SD. We find that routers with high degrees of expert coactivation result in much faster runtimes, mitigating the impact of using more verification tokens. Motivated by this observation, we assess the impact of various router design choices on expert coactivation and runtime using billion-parameter transformer models. We find that combining a global load-balancing loss, shared experts, a consistency loss, and an autoregressive expert selection mechanism during training results in significantly stronger expert coactivation. This increased coactivation translates into higher overall runtime throughput: our exploration yields a model that improves throughput by 21% over MoE baselines, while maintaining on-par accuracy with the baseline MoE.

Closest papers

Nearest neighbours by meaning across the whole index.

Why Speculative Decoding Works Better Than Predicted on Sparse MoE Models

cs.LG cs.AI ·2026-10-28 ·Ekagra Ranjan, Komal K. Teru, Bharat Venkitesh, Acyr Locatelli
top venue similarity 0.79
Abstract · Accepted as at NeurIPS 2026 main track conference

Speculative decoding (SD) accelerates autoregressive inference by leveraging the memory bound nature of decoding, verifying multiple draft tokens in a single forward pass. For mixture-of-experts (MoE) models, prior work predicts a concave speedup-batch size relationship, with gains peaking at intermediate batch sizes due to expert saturation. We evaluate this across 10 sparse MoE models (30B--1T parameters, 8--512 experts) and find recent fine-grained models break this trend, showing higher speedup at low batch size. We find verification cost is governed not only by activated-expert bytes, but also by temporal overlap across draft tokens and GPU execution dynamics. At low batch size, MoE layers in fine-grained models account for a much smaller share of runtime than of weights; on H100, kernel profiling attributes this mainly to higher bandwidth utilization of MoE kernels relative to attention, with L2 absorbing repeated reads of co-routed experts. Across draft tokens, routing overlap further reduces the unique experts activated during verification relative to independence assumptions. We model this effect using a maximum-entropy Iterative Proportional Fitting (IPF) framework that predicts unique expert counts from routing-overlap statistics with less than 4\% error, and substantially narrows the gap to real-model curves and captures the qualitative low-batch trend from synthetic routing.

0.32

BASE: Batch-Aware Selection of Experts Using Predicted Removal Error for Efficient MoE Decoding

cs.LG ·2026-09-28 ·Ali Abbasi, Justin Shi, Soheil Kolouri
similarity 0.77
Abstract · 19 pages, 4 figures, 8 tables

Large language models are increasingly expensive to serve. In large-scale serving systems, autoregressive decoding is often bottlenecked by transferring model weights from accelerator high-bandwidth memory into on-chip SRAM. Mixture-of-experts (MoE) models reduce computation by activating only a small subset of experts per token, but this sparsity does not translate directly to batched decoding. Different requests select different experts; therefore, the combined active set across many concurrent requests can span a substantial fraction of the expert pool and require significantly more expert weights to be transferred. Most expert-reduction techniques make retention decisions independently for each token and therefore do not address this batch-level expansion. More recently, batch-aware methods have attempted to coordinate expert use across concurrent requests and reuse experts already fetched for the batch. Yet their selection criteria are based primarily on router rankings or expert statistics collected during calibration. Consequently, these criteria are not directly tied to the output error caused by dropping an expert, nor do they capture how an expert's contribution changes across tokens at inference time. We instead rank experts according to how much their removal would change the MoE-layer output. To apply this criterion during serving, we train a lightweight linear predictor during calibration that estimates the expert removal cost for each incoming token, and develop custom GPU kernels for cost prediction and expert selection. Across three MoE architectures, BASE improves the quality-efficiency tradeoff without retraining. On Qwen3-30B-A3B, it improves average accuracy by 29.5 points over the strongest baseline at comparable throughput under a tight expert budget. At a higher expert budget, it is 60% faster than dense inference while remaining within 0.4 accuracy points.

0.39

Communicate the Change: Efficient Expert Parallelism via Activation Compression

cs.LG cs.AIcs.DC ·2026-11-17 ·Fengyuan Liu, Jay Gala, Siva Reddy, Dzmitry Bahdanau
similarity 0.73
Abstract

Scaling has driven the success of large language models (LLM), with sparse mixture-of-experts (MoE) architectures enabling greater parameter capacity without a proportional increase in computation. At large scales, MoE models rely on expert parallelism to distribute experts across devices. Each MoE layer requires two synchronous collective communication steps during inference: dispatching hidden states to the activated experts and combining the resulting expert outputs. As compute throughput has outpaced interconnect bandwidth, these synchronous collectives have become a major bottleneck to MoE scaling and efficiency. We address this challenge with EchoMoE, a lightweight architecture modification that retrofits existing MoE models, requiring approximately 8,500x fewer training tokens than full pretraining. Motivated by the observation that transformer blocks update the residual stream incrementally, we use a hidden state from an earlier layer as a reference and communicate only the compressed difference between it and the current state, rather than communicating each hidden state in full. EchoMoE periodically refreshes these reference states and transmits only the compressed differences synchronously, reducing the traffic on latency-critical paths without significant computational overhead. EchoMoE reduces inter-node communication volume by 58% and improves throughput by 22.6%, while preserving average performance within 3.8 percentage points across 18 benchmarks. EchoMoE further improves inference throughput by up to 2.38x over bandwidth-constrained Ethernet, paving the way for wide-area distributed inference. EchoMoE also tolerates quantization of the communicated representations without measurable quality loss. We release an efficient implementation for applying EchoMoE to pretrained MoE models, together with an inference integration for vLLM.

0.17

MoRE: Mixture of Reused Experts

cs.LG cs.AI ·2026-09-16 ·Eric S. Qiu, Utku Umur Acikalin, Justin Lovelace, Christian Belardi, Arjun B. Mulchandani +2 more
published similarity 0.73
Abstract · Accepted to the Conference on Language Modeling (COLM 2026)

Mixture-of-Experts (MoE) architectures decouple model capacity from computational cost, yet incur high memory footprints as parameters grow linearly with the number of experts. Recurrent Transformers achieve parameter efficiency by reusing layer weights, but typically lack the capacity for competitive language modeling. We propose Mixture of Reused Experts (MoRE), a hybrid that shares expert pools across groups of adjacent layers. Each layer retains its own router but selects from a larger shared pool, expanding the diversity of routing combinations without additional parameters. To enable shared experts to distinguish between layers, we introduce lightweight learnable depth embeddings that condition each layer's input before routing. Experiments across three model scales (114M-1.15B parameters) show that MoRE consistently achieves lower perplexity and stronger downstream performance than standard MoEs and state-of-the-art weight-sharing architectures at matched compute and parameter budgets, with only minimal modifications to existing MoE implementations.

0.50

S2-MoE: Enabling Efficient Self-Speculative Decoding for Mixture-of-Experts on Edge Devices

cs.AI ·2026-08-15 ·Haochen Huang, Shengxuan Qiu, Meng Li
code/data similarity 0.73
Abstract · 13 pages, 10 figures

Deploying large language models (LLMs) for inference on edge devices is challenging due to severe memory and bandwidth constraints. While speculative decoding and Mixture-of-Experts (MoE) have been proposed to improve inference efficiency, naively combining them often incurs excessive verification overhead and poor expert reuse, limiting their effectiveness in memory-bound edge settings. In this work, we propose S2-MoE, an efficient self-speculative decoding framework for MoE inference on edge devices. S2-MoE reduces redundant verification through routing-aware adaptive speculative expansion, improves verification efficiency with reuse-aware expert gating, and aligns draft and target execution via shared context. Implemented in llama$.$cpp, S2-MoE achieves up to $5.3\times$ speedup (about $2.0\times$ on average) over standard autoregressive decoding across diverse MoE models and datasets on edge devices. Code is available at https://github.com/angerybob/S2-MoE.

0.67

3D-MoE: Complementarity-Guided Hardware-Software Co-Design for MoE Decode on 3D-Stacked NMP

cs.AR cs.AIcs.DC ·2026-10-19 ·Yudong Pan, Lian Liu, Shixin Zhao, Yintao He, Kaiyan Chang +6 more
similarity 0.72
Abstract

Mixture-of-Experts (MoE) have become a mainstream way to scale Large Language Models (LLMs), because each token activates only a small subset of experts, reducing inference cost. However, MoE decode remains highly memory-bound and exhibits strongly skewed, time-varying expert loads. 3D-stacked DRAM near-memory processing (NMP) provides the bandwidth that decode demands, yet its tiled organization makes load balancing expensive: expert parallelism (EP) preserves locality but suffers from hotspot-induced imbalance, while more distributed strategies weaken locality and increase NoC traffic. To address this locality-balance conflict, we observe that domain-specialized token routing creates offsetting expert-load fluctuations across decode steps, a property we call expert complementarity. Using this insight, we propose 3D-MoE, a hardware--software co-designed system for MoE decode on 3D-stacked NMP. 3D-MoE uses offline, one-time complementarity-guided grouping and locality-aware mapping to establish coarse-grained inter-group balance, handles residual intra-group hotspots through a lightweight Collaborative Memory-Network (CMN) that streams each hot expert's weight tiles once from its host DRAM channel to helper cores over a reconfigured local broadcast path, and uses cost-aware online remapping to remain robust to workload drift. Across MoE models, 3D-MoE achieves an average 4.63$\times$ MoE-layer speedup over EP and improves end-to-end throughput by 1.58$\times$--1.86$\times$ over the strongest NMP baselines, with 2.11\% additional logic-die area.

0.14

Fuse-All MoE Inference Kernel with Load Balancing on TPUs

cs.LG ·2026-11-13 ·Rupeng Liu, Bangsheng Tang, Zongwei Zhou, Chengji Yao, Yejin Lee
similarity 0.72
Abstract

Mixture-of-Experts (MoE) architectures activate only a subset of experts for each token, allowing large language models (LLMs) to scale parameter capacity without a proportional increase in computation per token. However, intermediate data movement and uneven expert loads limit the efficiency of MoE inference on TPUs. To address these bottlenecks, we introduce Fuse-All MoE, a TPU kernel design that jointly optimizes expert computation, output communication, and load balancing. The kernel fuses token collection, expert projections, gated activation, and output writes, retaining expert computation intermediates on chip. Fixed token-slot destinations allow scheduling and expert placement to change while preserving routing semantics. In prefill-decode (PD) disaggregated serving, Fuse-All MoE balances EP16 prefill workloads through scheduling and expert redistribution, while symmetric EP8TP2 shares active expert computation between the two TensorCore devices within each chip during decode. Our experiments show that Fuse-All MoE reduces kernel runtime for a measured skewed EP16 prefill workload from 2.3x to 1.09x the same fixed reference time for a balanced workload. Across 32-512 decode tokens, symmetric EP8TP2 reduces kernel latency by up to 21.3% relative to fused EP16. With accompanying serving stack optimizations, the integrated system improves serving throughput by up to 34.5% over the Megablox EP16 baseline.

0.16

MoE-CORE: Coordinated Expert Offloading and Residency for Memory-Constrained MoE Inference

cs.DC ·2026-10-01 ·Ke Yang, Yongji Gao, Xushi Li, Kui Luo, Sicheng Zhang +14 more
similarity 0.72
Abstract

Sparse expert activation reduces MoE models' computation, yet expert weights can exceed limited device memory. Offloading makes inference feasible on a compact AI appliance but exposes host-to-device transfers to the inference path. We present MoE-CORE, a system that coordinates expert offloading and residency for memory-constrained MoE inference. It stages complete expert layers in alternating buffers during prefill. During decode, it combines nonuniform layer-wise cache capacity, domain-informed initialization, routing-history-aware replacement, and cross-layer prefetching. The main configuration executes router-selected experts exactly; an optional score-based substitution path handles eligible low-score misses. The main comparison uses 1K- and 128-token output caps for MoE-CORE and vLLM Prefetch, respectively. Across five workloads per model, MoE-CORE records a mean time per output token (TPOT) of 38.0-44.8 ms versus 1268.9-1269.1 ms for the evaluated vLLM Prefetch configuration on DeepSeek-V4-Flash-W4A8; the corresponding values on GLM-5.2-W4A8C8 are 206.6-220.5 and 5941.5-5941.8 ms. Under an 84-GB NPU-memory cap, the best measured DeepSeek GSM8K configuration achieves a TPOT of 21.5 ms with approximate expert substitution and multi-token prediction (MTP) at depth 2. These results support coordinated expert residency and transfer scheduling under a device-memory constraint. The code is here.

0.39