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

arXiv: https://arxiv.org/abs/2606.01502 · HTML: https://mlsanity.com/paper/2606.01502/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in KV cache optimization, one of 12 directions in Efficient LLM inference and serving.
- Builds on 31 of its 42 references that are in the map.
- First posted on arXiv 31 May 2026.

## Placed in

- Efficient LLM inference and serving / [KV cache optimization](https://mlsanity.com/inference/kv-cache-optimization/)

## Builds on (most cited first, 30 of 31)

- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://mlsanity.com/paper/2205.14135/) (2022)
- [SGLang: Efficient Execution of Structured Language Model Programs](https://mlsanity.com/paper/2312.07104/) (2023)
- [DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model](https://mlsanity.com/paper/2405.04434/) (2024)
- [DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence](https://mlsanity.com/paper/2606.19348/) (2026)
- [DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models](https://mlsanity.com/paper/2512.02556/) (2025)
- [DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving](https://mlsanity.com/paper/2401.09670/) (2024)
- [GLM-5: from Vibe Coding to Agentic Engineering](https://mlsanity.com/paper/2602.15763/) (2026)
- [Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention](https://mlsanity.com/paper/2502.11089/) (2025)
- [Prompt Cache: Modular Attention Reuse for Low-Latency Inference](https://mlsanity.com/paper/2311.04934/) (2023)
- [FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving](https://mlsanity.com/paper/2501.01005/) (2025)
- [CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion](https://mlsanity.com/paper/2405.16444/) (2024)
- [Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving](https://mlsanity.com/paper/2407.00079/) (2024)
- [Online normalizer calculation for softmax](https://mlsanity.com/paper/1805.02867/) (2018)
- [LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference](https://mlsanity.com/paper/2510.09665/) (2025)
- [Infinite-LLM: Efficient LLM Service for Long Context with DistAttention and Distributed KVCache](https://mlsanity.com/paper/2401.02669/) (2024)
- [EPIC: Efficient Position-Independent Caching for Serving Large Language Models](https://mlsanity.com/paper/2410.15332/) (2024)
- [IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse](https://mlsanity.com/paper/2603.12201/) (2026)
- [Context Parallelism for Scalable Million-Token Inference](https://mlsanity.com/paper/2411.01783/) (2024)
- [Injecting Adrenaline into LLM Serving: Boosting Resource Utilization and Throughput via Attention Disaggregation](https://mlsanity.com/paper/2503.20552/) (2025)
- [UCCL-EP: Portable Expert-Parallel Communication](https://mlsanity.com/paper/2512.19849/) (2025)
- [MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems](https://mlsanity.com/paper/2505.11415/) (2025)
- [MEPIC: Memory Efficient Position Independent Caching for LLM Serving](https://mlsanity.com/paper/2512.16822/) (2025)
- [fabric-lib: RDMA Point-to-Point Communication for LLM Systems](https://mlsanity.com/paper/2510.27656/) (2025)
- [ContextPilot: Fast Long-Context Inference via Context Reuse](https://mlsanity.com/paper/2511.03475/) (2025)
- [Helix Parallelism: Rethinking Sharding Strategies for Interactive Multi-Million-Token LLM Decoding](https://mlsanity.com/paper/2507.07120/) (2025)
- [FlexiCache: Leveraging Temporal Stability of Attention Heads for Efficient KV Cache Management](https://mlsanity.com/paper/2511.00868/) (2025)
- [NCCL EP: Towards a Unified Expert Parallel Communication API for NCCL](https://mlsanity.com/paper/2603.13606/) (2026)
- [BanaServe: Unified KV Cache and Dynamic Module Migration for Balancing Disaggregated LLM Serving in AI Infrastructure](https://mlsanity.com/paper/2510.13223/) (2025)
- [An Efficient Hybrid Sparse Attention with CPU-GPU Parallelism for Long-Context Inference](https://mlsanity.com/paper/2605.07719/) (2026)

## Built on by (most cited first, 1 of 1)

- [Q-First: Most of Attention Needs Only the Query in Disaggregated LLM Decoding](https://mlsanity.com/paper/2608.15473/) (2026)

## Abstract (arXiv)

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.