Accelerating MoE Training through Expert Placement and Replica Coordination

Zhiqiang Tan; Maoxin Wang; Wenxiang Lin; Sijie Wang; Qiang Wang; Shaohuai Shi; Bo Li

Home/Archive/arXiv

arXiv · 2610.14798

Accelerating MoE Training through Expert Placement and Replica Coordination

Zhiqiang Tan·Maoxin Wang·Wenxiang Lin·Sijie Wang·Qiang Wang·Shaohuai Shi·Bo Li

View original source ↗Read source document ↗Export metadata ↓

Abstract

Distributed training of sparsely activated mixture-of-experts (MoE) models is often bottlenecked by workload imbalance and all-to-all (A2A) communication induced by dynamic token routing. Existing token deduplication, expert placement, and replication techniques alleviate these bottlenecks but do not jointly capture their distinct load semantics. Under hierarchical token deduplication, communication counts each token once per destination group at each hierarchy level, whereas expert computation processes every token-expert assignment. With replication, both costs depend jointly on physical layout and token-to-copy mapping, making separate optimization suboptimal. We present PlaceMoE to address this coupling. First, we formulate the joint optimization of expert layout and source-aware token-to-copy mapping for E experts distributed across P devices, accounting for both hierarchical communication traffic and expert-compute load. Second, we develop a solver that combines co-selection-aware expert grouping, bounded dynamic programming for replica allocation, and alternating refinement of layout and mapping. Third, PlaceMoE overlaps replica-gradient synchronization with attention backward. We evaluate PlaceMoE on a 32-GPU cluster and a 64-NPU cluster using Qwen3-VL and DeepSeek-V3. Experimental results show that PlaceMoE achieves up to a 6.94x speedup in A2A communication and 1.74-2.33x speedups in end-to-end training over the state-of-the-art training framework VeOmni.

Explore related subjects

cs.DC

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Zhiqiang Tan, Maoxin Wang, Wenxiang Lin, Sijie Wang, Qiang Wang, Shaohuai Shi, Bo Li. 2026-10-10. Accelerating MoE Training through Expert Placement and Replica Coordination. https://arxiv.org/abs/2610.14798

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

PerCache: Predictive Hierarchical Cache for RAG Applications on Mobile Devices

Retrieval-augmented generation (RAG) has become a de facto paradigm in various large language model (LLM)-driven applications on mobile devices, such as mobile assistants leveraging personal emails or meeting records. However, due to lengthy prompts and resource constraints, mobile RAG systems exhibit significantly high response latency. To address this issue, one promising approach is to reuse intermediate computational results across different queries to eliminate redundant computation. But most existing approaches, such as key-value (KV) cache reuse and semantic cache reuse, are designed for cloud settings and perform poorly on mobile devices as they overlook the distinctive characteristics of mobile RAG. We propose PerCache, a novel hierarchical cache framework designed for reducing end-to-end latency of personalized RAG applications on mobile platforms. PerCache adopts a hierarchical architecture that progressively matches similar queries and query-key-value (QKV) cache to maximize the reuse of intermediate results at different computing stages. To improve cache hit rate, PerCache applies a predictive method to populate the cache with queries that are likely to be raised in the future. In addition, PerCache can adapt its configurations to dynamic system loads, aiming to maximize caching utility with constrained resource consumption. We implement PerCache on top of an existing mobile LLM inference engine and deploy it on commodity mobile phones. Extensive evaluations show that PerCache achieves the lowest average end-to-end latency across five datasets and two LLM architectures, reducing latency by up to 63.5% over the strongest baseline while maintaining comparable answer quality.

cs.DC↗

arXiv2026

Dynamic Resource Management in Production HPC Clusters

Many large-scale scientific applications exhibit time-varying behavior, yet production HPC clusters still rely on rigid, fixed-size allocations, and most dynamic techniques remain confined to laboratory prototypes. This work presents a practical MPI malleability methodology that integrates with state-of-the-art high-performance computing (HPC) software stacks and operational practices. The methodology is implemented in the Dynamic Management of Resources (DMR) framework and is designed to ease adoption by existing applications without requiring intrusive code changes or scheduler modifications. We evaluate our approach by integrating the DMR API into two large-scale scientific applications and deploying them on three TOP500 supercomputers under realistic production configurations. Our non-invasive malleability solution achieves performance comparable to static baselines in controlled environments while substantially reducing node-hour consumption for identical workloads. These results show that malleability can be effectively exploited on production systems using vanilla resource managers, lowering the barrier to adoption of dynamic resource management in HPC.

cs.DC↗

arXiv2026

Purlin: Separating Orchestration from the Datapath of Collectives

Distributed inference depends on GPU collective communication that must keep pace with evolving hardware and specialized workloads. However, existing works often couple the semantics, orchestration (where and when data moves), and datapath (how data moves) of collective communication. This coupling makes it costly to adopt new hardware mechanisms and customize communication for applications. We present Purlin, a scale-up communication framework that separates these concerns. At the top of Purlin, we specify collectives as a naming of an input and output layout and a copy or reduction operation. In the middle, we introduce a shared orchestration protocol, Stage, Notify, And Consume (SNAC), which derives coordination from these specifications. Below SNAC sits a hardware-specific datapath we call Atom, which implements two key data movement primitives for collectives: copy and reduce. This separation lets us customize collectives and adopt new hardware mechanisms while reusing orchestration via SNAC. We evaluate Purlin on A100, H200, and B200 GPUs. Across seven collectives, Purlin achieves latency speedups of up to 5.14x and bandwidth improvements of up to 4.50x over baselines. Integrated into SGLang, Purlin improves offline LLM serving throughput and interactivity by 1.13x on average and up to 1.37x over baselines. For online LLM inference, Purlin improves interactivity by 1.26x on average and up to 2.85x, with the largest gain occurring under overload. For diffusion image generation, Purlin reduces end-to-end latency by up to 1.13x.

cs.DC↗