The paper presents py‑kvcache, a new KV offload connector for vLLM that uses asynchronous direct I/O, bounded shared staging, and scheduler‑aware preloading to improve external KV caching performance on NVMe SSDs. Experiments across synthetic workloads, long‑context benchmarks, and production traces show that py‑kvcache can load 80k‑token prefixes 2.0× faster than LMCache, with preloading contributing an additional 1.34× speedup, and achieves overall performance within 4% of native vLLM KV Offload. The study highlights that cache effectiveness depends on transfer granularity, intermediate memory use, and scheduling timing rather than just device bandwidth, indicating that external KV caching should be considered a setup‑specific admission decision.
By Joseph Kanichai, Tiziano De Matteis, Animesh Trivedi
arXiv:2606. 24506v1 Announce Type: cross Abstract: Emerging LLM services increasingly host many sparse MoE models, yet most models receive sparse requests and remain cold.
By Zhuoren Ye, Tianyu Wo, Dinghao Xue, Mingming Zhang, Yuchen Teng, Chunming Hu, Renyu Yang
arXiv:2609.16215v1 Announce Type: new
Abstract: GPU high bandwidth memory is scarce and expensive, and KV caches consume much of it as chats, agent loops, and document question answering accumulate s...
By Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly
arXiv:2607. 26491v1 Announce Type: cross Abstract: The energy consumption of Large Language Model (LLM) serving is becoming a major system challenge as deployment scales, driven by hardware power and thermal constraints and rising electricity costs.
By Ming-Yen Lee, Hanchen Yang, Faaiq Waqar, Harsono Simka, Tushar Krishna, Muhammed Ahosan Ul Karim, Shimeng Yu
The paper argues that for large‑context autoregressive language‑model inference, memory bandwidth—specifically the Key‑Value (KV) cache—becomes the limiting resource rather than arithmetic throughput. It analytically derives how arithmetic intensity decays with context length for NVIDIA H100, NVIDIA B200, and AMD MI300X, identifies crossover points where KV traffic overtakes weight traffic, and evaluates representative techniques across five compression domains. The study finds a three‑regime behavior: below the crossover, weight traffic dominates and KV compression offers little benefit; beyond it, KV traffic dominates and compression methods trade quality for bandwidth, with paging and prefix sharing being lossless but capacity‑limited, while quantization and eviction directly reduce bandwidth at the cost of accuracy.
whyItMatters":"The work provides a unified analytical framework and a standardized protocol that enable consistent comparison of KV‑compression techniques across hardware and workloads, guiding practitioners in selecting appropriate methods for long‑context inference."
By Tejinder Singh
arXiv:2605. 09735v2 Announce Type: replace-cross Abstract: Static-graph LLM decoders provide predictable launches, fixed tensor shapes, and low submission overhead, but online decoding exposes highly irregular KV-cache behavior: request lengths differ, EOS events arrive asynchronously, and logical histories fragment over time.
By Zhiqing Zhong, Zhijing Ye, Jian Zhang, Weijian Zheng, Bolun Sun, Xiaodong Yu
arXiv:2607. 27090v1 Announce Type: cross Abstract: Large language models are increasingly deployed with persistent personalized context, such as accumulated memory profiles or long conversation histories, that is shared across a user's many requests.
By Peter Li, Prashant Pandey
arXiv:2606. 06302v1 Announce Type: new Abstract: Multi-turn Large Language Model (LLM) serving is critical for consistent user experiences, yet the linear growth of the Key-Value (KV) cache imposes significant pressure on GPU memory and bandwidth.
By Hyungmin Kim, Minsoo Kim, Hongseok Kim, Jungwook Choi
arXiv:2606. 21633v2 Announce Type: replace Abstract: The KV cache dominates GPU memory in long-context LLM serving, crowding out batch capacity and leaving GPU compute idle.
By Omin Kwon, Doyeon Kim, Jongseok Park, Seung Yul Lee, Ion Stoica, Jae W. Lee
DPS (Dual-Mode Precision LLM Serving) is a system that treats model‑weight memory as elastic by using a multi‑precision representation. Under normal load it serves the full‑accuracy model, but when KV‑cache pressure spikes it switches to a lower‑precision variant and reallocates unused weight memory for KV cache blocks. Built on Semi‑Unified Memory and implemented on top of vLLM, DPS boosts sustained throughput by 2.1–3.3× and effective pass@1 by up to +41 pp over static FP16 while maintaining FP16‑class accuracy.
The paper introduces a Kubernetes Dynamic Resource Allocation driver that treats composable CXL memory as a schedulable cluster resource, enabling cross-node shared memory for large language model (LLM) serving. By composing CXL regions on demand, materializing them as DAX devices, and exposing them via a single Container Device Interface name, pods on different nodes can access the same physical memory region. A shared‑memory connector for vLLM/llm‑d uses this region as a KV‑cache tier, eliminating external metadata services and achieving significant reductions in time‑to‑first‑token (TTFT) with minimal additional latency compared to same‑node reuse.
By Hongjian Fan, Kevin Zhang, David Habinsky, Sean Dykstra
arXiv:2609.14507v1 Announce Type: cross
Abstract: Single-GPU long-context inference with Mixture-of-Experts (MoE) models requires spilling the key-value cache (KVCache) to CPU memory. The spilled KV...
By Enda Yu, Dezun Dong, Xiangke Liao