Move the Query, Not the Cache: Characterizing Cross-Instance Latent Attention Redistribution Across GPU Fabrics

• URL: http://arxiv.org/abs/2606.01502v1
• Date: Sun, 31 May 2026 23:53:24 GMT
• Title: Move the Query, Not the Cache: Characterizing Cross-Instance Latent Attention Redistribution Across GPU Fabrics
• Abstract summary: Multi-head Latent Attention inverts arithmetic, compressing each token's key and value into one narrow vector.<n>We characterize cross-instance MLA attention on a real multi-node H100 cluster.
• Score: 1.081571058570587
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Frontier LLMs increasingly decide what a query attends to with a sparse-attention indexer that picks a few KV-cache blocks per query: attention's unit is now a small, reusable chunk. Agentic workloads hammer it: many sub-agents query one large codebase, reusing the same blocks. When that corpus outgrows one GPU it is partitioned across instances, so a query and the blocks it selects often sit on different GPUs: answering it means attention across instances. The reflex of prior cross-instance KV systems is to move the cache: pull the selected blocks to the requester. Multi-head Latent Attention inverts the arithmetic, compressing each token's key and value into one narrow vector, so a routed query row is only ~1 KB, smaller than the chunk it attends; routing the query is then often cheaper than moving the cache. Which primitive wins, over which fabric and request shape, is uncharted, least of all on device-initiated RDMA that makes per-request cross-node transfers cheap. We characterize cross-instance MLA attention on a real multi-node H100 cluster, distilling two reusable artifacts: a topology-aware cost model (probe / transfer / compute / return / merge) and a closed-form route/fetch/local predicate, whose constants we measure on real IBGDA, where the model tracks batched round-trips to within ~7%. At decode it routes the query, trading the cost of moving the cache (a ~3 ms re-adaptation splice for a contiguous chunk, or a scattered gather under selection) for a tens-of-microsecond round trip, and picks the fabric by probe latency, not peak bandwidth. We instantiate the cost model and predicate for MLA, but neither is MLA-specific: they apply wherever compression or sparse selection shrinks attention to small chunks (DeepSeek-V3.2, V4, and GLM-5.1 today). Extending them to a new architecture requires measuring just two coefficients: the routed payload and fetch's move-the-cache cost.

Related papers

• CacheBridge: Efficient Cross-Model KV Cache Transfer [12.058564378845373]
  Full-Head Mapping maps each target KV head from every source KV head in selected layers, making transfer quality sensitive to architectural differences.<n>We introduce CacheBridge, which co-designs architecture-indexed mapper support, attention-aligned calibration, and bounded mapper construction.<n>On Qwen3 $14mathrmBto32mathrmB$, it reduces mapper storage by $8times$, accelerates application by up to $3.0times$, matches fullhead with one tenth of the calibration data, and reduces 500-sequence construction from 92.63 to 8.
  arXiv  Detail & Related papers  (2026-09-01T08:20:35Z)
• Back from the Future: Key-Value Cache Management by Counter-Causal Surprise [67.1056509495879]
  Key-value (KV) cache management through compression and eviction strategies has emerged as an important research direction in recent years.<n>We propose a simple yet effective KV eviction scheme motivated by the insight that past tokens which can be well-predicted from more recent tokens are redundant.<n>We evaluate our strategy on various open-source LLMs and benchmark datasets showing competitive or improved performance over other state-of-the-art methods.
  arXiv  Detail & Related papers  (2026-07-30T02:42:51Z)
• What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents [1.0742675209112622]
  We show how four separate research communities have each learned to compact large language memory.<n>We argue that these are instances of one problem: a rate--distortion decision about what context-derived information to retain versus discard.<n>We make this lens precise with a single compaction objective and a layer-agnostic lower bound, use it to build a seven-axis taxonomy that classifies methods from across the stack uniformly.
  arXiv  Detail & Related papers  (2026-07-09T01:15:03Z)
• You Only Index Once: Cross-Layer Sparse Attention with Shared Routing [61.29627714699688]
  Cross-layer sparse attention (A) is built on top of KV-sharing architectures such as YOCO.<n>A single indexer computes token-level top-k selection once and reuses the resulting index across layers.<n>Results show CLSA is both accurate and efficient, achieving up to 7.6x decoding speedup and 17.1x overall throughput improvement at 128K context.
  arXiv  Detail & Related papers  (2026-06-04T17:54:04Z)
• Nectar: Neural Estimation of Cached-Token Attention via Regression [36.709992864731866]
  We show how a compact neural network can be used to evaluate softmax attention over a fixed long context.<n>Nectar fits two networks per layer and KV-head: a target network that predicts the attention output and a score network that predicts the log-normalizer.<n>We report experiments on models from 1.7B to 8B parameters across five long-context datasets.
  arXiv  Detail & Related papers  (2026-05-10T21:51:36Z)
• Sparse Prefix Caching for Hybrid and Recurrent LLM Serving [0.7284556903703034]
  We formalize sparse prefix caching as checkpoint placement under a distribution over overlap depths.<n>For use cases where requests share a non-trivial prefix, we show that our method consistently improves the frontier traced by standard asymmetry on real-world data.<n>It preserves exact outputs, does not change the recurrent computation itself or require new recurrent update kernels.
  arXiv  Detail & Related papers  (2026-04-17T09:24:58Z)
• Latent-Condensed Transformer for Efficient Long Context Modeling [60.72493959155964]
  We propose Latent-Condensed Attention (LCA) for large language models.<n>LCA condenses context within MLA's latent space, where representation is disentangled into semantic latent vectors and positional keys.<n>LCA achieves up to 2.5$times$ prefilling speedup and 90% KV cache reduction at 128K context.
  arXiv  Detail & Related papers  (2026-04-14T08:40:31Z)
• Joint Encoding of KV-Cache Blocks for Scalable LLM Serving [3.3230675313521716]
  Existing KV-cache compression methods rely on rigids, disrupt tensor layouts, or require specialized compute.<n>We propose joint encoding of KV-cache blocks, which fuses similar blocks across requests and input chunks into shared representations.<n>This alleviates the KV-cache memory bottleneck, supporting high-concurrency serving without specialized hardware.
  arXiv  Detail & Related papers  (2026-01-06T14:50:58Z)
• vCache: Verified Semantic Prompt Caching [95.16654660556975]
  This paper proposes vCache, the first verified semantic cache with user-defined error rate guarantees.<n>It employs an online learning algorithm to estimate an optimal threshold for each cached prompt, enabling reliable cache responses without additional training.<n>Our experiments show that vCache consistently meets the specified error bounds while outperforming state-of-the-art static-threshold and fine-tuned embedding baselines.
  arXiv  Detail & Related papers  (2025-02-06T04:16:20Z)
• Cache-Craft: Managing Chunk-Caches for Efficient Retrieval-Augmented Generation [14.842469293627271]
  CacheCraft is a system for managing reusing precomputed KVs corresponding to the text chunks.<n>We show how to identify chunk-caches that are reusable, how to efficiently perform a small fraction of recomputation to fix the cache, and how to efficiently store and evict chunk-caches in the hardware.
  arXiv  Detail & Related papers  (2025-02-05T14:12:33Z)
• Compute Or Load KV Cache? Why Not Both? [6.982874528357836]
  Cake is a novel KV cache loading system that optimally utilizes both computational and I/O resources in parallel.<n> Cake achieves on average 2.6x reduction in Time to First Token (TTFT) compared to compute-only and I/O-only methods.
  arXiv  Detail & Related papers  (2024-10-04T01:11:09Z)
• ThinK: Thinner Key Cache by Query-Driven Pruning [63.13363917871414]
  Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications.<n>This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference.<n>We propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels.
  arXiv  Detail & Related papers  (2024-07-30T17:59:08Z)
• Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching [56.286064975443026]
  We make an interesting and somehow surprising observation: the computation of a large proportion of layers in the diffusion transformer, through a caching mechanism, can be readily removed even without updating the model parameters. We introduce a novel scheme, named Learningto-Cache (L2C), that learns to conduct caching in a dynamic manner for diffusion transformers. Experimental results show that L2C largely outperforms samplers such as DDIM and DPM-r, alongside prior cache-based methods at the same inference speed.
  arXiv  Detail & Related papers  (2024-06-03T18:49:57Z)
• CORM: Cache Optimization with Recent Message for Large Language Model Inference [57.109354287786154]
  We introduce an innovative method for optimizing the KV cache, which considerably minimizes its memory footprint. CORM, a KV cache eviction policy, dynamically retains essential key-value pairs for inference without the need for model fine-tuning. Our validation shows that CORM reduces the inference memory usage of KV cache by up to 70% with negligible performance degradation across six tasks in LongBench.
  arXiv  Detail & Related papers  (2024-04-24T16:11:54Z)
• KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache [67.9776980972508]
  We develop a tuning-free 2bit KV cache quantization algorithm named KIVI. KIVI can enable Llama, Falcon, and Mistral models to maintain almost the same quality while using $mathbf2.6times$ less peak memory.
  arXiv  Detail & Related papers  (2024-02-05T06:06:47Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.