Searcharxiv⌕ Search

arXiv · 2611.22280

XM-Sparse: High-Throughput Sparse-Attention LLM Serving over Scale-up Memory Pools

Abstract

Sparse attention has been increasingly adopted by production LLMs for long-context inference, yet serving systems must still retain the full KV history, making GPU HBM capacity a key limit on serving concurrency. Existing systems offload the KV cache to host memory, but suffer from the limited bandwidth of PCIe and the limited capacity of host memory. Emerging scale-up memory pools (e.g., X-Mem) relieve both limits and open new opportunities for sparse-attention offloading. In this paper, we first conduct an in-depth analysis and reveal that, even with scale-up memory pools, decode throughput remains bounded by HBM capacity, as hot KV buffers and the fully resident indexer key (K-index) cache occupy substantial HBM. Moreover, the two caches exhibit distinct access patterns. Based on these insights, we present XM-Sparse, the first sparse-attention offloading system over scale-up memory expansion. The core idea of XM-Sparse is to offload both caches aggressively while matching each with an I/O mechanism tailored to its access pattern. Specifically, for dynamically selected KV entries, XM-Sparse fuses TMA-based retrieval with attention in a tile-level pipeline to hide on-demand retrieval latency. For predictable K-index accesses, it prefetches pages via SM-efficient GPU-initiated DMA and coordinates prefetching with latency-critical operations to avoid interfering with decoding. We have implemented XM-Sparse atop SGLang, and our experiments on DeepSeek-V3.2 and GLM-5.3 with an emulated scale-up memory pool show that XM-Sparse delivers up to 4.98$\times$ and 4.52$\times$ higher decode throughput than PCIe-based ECHO and HiSparse, respectively, and up to 1.34$\times$ over a naive integration on the same platform, across 16K-1M contexts.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yijun Sun, Xudong Liao, Zhuobin Huang, Kai Chen, Nikita Lazarev, Zhizhen Zhong. 2026-11-17. XM-Sparse: High-Throughput Sparse-Attention LLM Serving over Scale-up Memory Pools. https://searcharxiv.com/abs/2611.22280

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

JigSLO: Solving the Resource Puzzle for Efficient Multi-LLM Serving

GPU spatial sharing co-locates instances with complementary demands, such as compute-bound prefill and memory-bound decode, promising cost-efficient multi-LLM serving. Prior systems partition compute and coordinate shared memory capacity, but do not model bandwidth contention among co-located instances of heterogeneous LLMs. This contention makes each instance's latency depend on the memory-access patterns of its co-runners, even when its own resource allocation remains unchanged. Drawing on extensive measurements, we find that the latency impact of bandwidth contention can be captured by a time-shared memory-access model with contended and uncontended periods, so the overall latency can be analytically estimated as a time-weighted average of the latency estimates for these periods. We propose JigSLO, a contention-aware multi-LLM serving system that manages memory resources and schedules each model's prefill and decode phases as separate instances, treating their co-location as a placement decision. Its contention-aware predictor estimates per-instance latency under any co-location from standalone profiles alone. Guided by these predictions, its placement algorithm jointly selects each instance's tensor parallelism, compute share, and batch configuration, pairing complementary memory-access patterns to maximize throughput within each model's latency targets. JigSLO's shared resource manager complements this on the capacity side, sharing weights, pooling KV cache, and transferring it between prefill and decode instances. Across three datasets covering chat, code generation, and long-context summarization, JigSLO outperforms state-of-the-art baselines by $2.3\times$ in SLO attainment and $1.2\times$ in throughput on average, matching their best attainment with half the GPUs on long-context workloads.

cs.DC↗

Angelfish: Leader, DAG, or Anywhere in Between

To maximize performance, many modern blockchain systems rely on eventually synchronous, Byzantine fault-tolerant (BFT) consensus protocols. Two protocol designs have emerged in this space: protocols that minimize latency using a leader that drives both data dissemination and consensus, and protocols that maximize throughput using a separate, asynchronous data dissemination layer. Recent protocols such as Partially Synchronous Bullshark and Sailfish combine elements of both approaches by using a DAG to enable parallel data dissemination and a leader that paces DAG formation. This improves latency while achieving state-of-the-art throughput. Yet leader-based protocols still achieve better latency under moderate loads, which are common in practice. We present Angelfish, a hybrid protocol that adapts smoothly across this design space, from leader-based to Sailfish-like DAG-based consensus. Angelfish lets a dynamically adjusted subset of parties use best-effort broadcast to issue lightweight votes instead of reliably broadcasting costlier DAG vertices. This reduces communication, helps lagging nodes catch up, and lowers latency in practice compared to prior DAG-based protocols. Our empirical evaluation shows that Angelfish attains state-of-the-art peak throughput while significantly lowering latency under moderate load, delivering the best of both worlds.

cs.DC↗

AInfer-PD: Communication-Safe In-Place Prefill-Decode Multiplexing for Distributed MoE Rollouts

In agentic reinforcement learning (RL), asynchronous multi-turn trajectories repeatedly introduce continuation prefills (P) while other trajectories remain in decode (D). Efficient rollout completion therefore requires D to advance during long P executions. In-place multiplexing shares devices, model weights, and KV state, but phase-specific communication in distributed MoE introduces two constraints: cross-rank collective ordering and ownership of mutable communication state. Whole-prefill coordination can ensure safe ordering but delays D. DeepEP's normal and low-latency paths suit large P and small D batches, respectively, but safely overlapping them requires separating their mutable communication state. We present AInfer-PD, which extends in-place P/D multiplexing to distributed MoE rollouts. AInfer-PD establishes rank-consistent collective ordering at safe communication boundaries, allowing successive D iterations to advance without waiting for an entire prefill execution to be enqueued or completed. It gives P and D separate communication buffers and protocol state within one DeepEP runtime, allowing P to use the normal path while D retains the low-latency path. Together, these mechanisms preserve useful P/D overlap while sharing model weights and KV storage on the same devices. Across repeated single-node prefill-intensive workloads, AInfer-PD reduces fixed-workload rollout completion time by 7.1-22.5% relative to the same AInfer engine with P/D multiplexing disabled and by 24.8-32.9% relative to SGLang. On two nodes, the reductions are 18.0-35.3% and 18.3-31.8%, respectively. In a same-engine ablation, fine-grained boundaries reduce completion time by a further 8.6-19.8% over whole-prefill asynchronous enqueue.

cs.DC↗