🛰️ Daily AI Frontier
‹ back to 2026-09-12

Building py-kvcache: A Performance Characterization of External KV Caching for vLLM with NVMe SSDs

Research Efficiency & Systems

Ranking

Overall 78
Content 95
Popularity 37

Observed public metrics from 1 member.

Representative image for Building py-kvcache: A Performance Characterization of External KV Caching for vLLM with NVMe SSDs

Merged summary

TL;DR - This paper characterizes when external KV caching improves vLLM inference and introduces py-kvcache, an NVMe-backed connector that overlaps asynchronous disk reads with computation. It shows that caching should be enabled selectively based on workload, hardware, and prefix length rather than raw storage bandwidth alone.

  • At 80K tokens, py-kvcache loads KV states from disk 2.0× faster than LMCache; scheduler-aware preloading contributes a 1.34× speedup.
  • With GPU, CPU, and disk caching enabled, it is 1.23× faster than LMCache and within roughly 4% of native vLLM KV Offload.
  • Performance depends on transfer granularity, staging-memory use, and scheduling, with benefits extending to irregular prefix chains and multi-turn workloads.
  • Trace replays improved time to first token on a weaker GPU, but not on an H100 where requests were often below the caching break-even point and GPU memory retained enough prefixes.

Sources (1)

Building py-kvcache: A Performance Characterization of External KV Caching for vLLM with NVMe SSDs

arXiv cs.DC Joseph Kanichai, Tiziano De Matteis, Animesh Trivedi 2026-09-10 arXiv:2609.11744
Public signals Semantic Scholar citations 0 · Semantic Scholar influential citations 0
Providers: Hugging Face · N/A OpenAlex · N/A Publisher · N/A Semantic Scholar · Citations 0 · Influential citations 0 X · N/A Fetched 2026-09-26 14:16:11.207148 UTC

TL;DR - This paper characterizes when external KV caching improves vLLM inference and introduces py-kvcache, an NVMe-backed connector that overlaps asynchronous disk reads with computation. It shows that caching should be enabled selectively based on workload, hardware, and prefix length rather than raw storage bandwidth alone.

  • At 80K tokens, py-kvcache loads KV states from disk 2.0× faster than LMCache; scheduler-aware preloading contributes a 1.34× speedup.
  • With GPU, CPU, and disk caching enabled, it is 1.23× faster than LMCache and within roughly 4% of native vLLM KV Offload.
  • Performance depends on transfer granularity, staging-memory use, and scheduling, with benefits extending to irregular prefix chains and multi-turn workloads.
  • Trace replays improved time to first token on a weaker GPU, but not on an H100 where requests were often below the caching break-even point and GPU memory retained enough prefixes.
item →