Irminsul: MLA-Native Position-Independent Caching for Agentic LLM Serving

TL;DR

Irminsul introduces a content-addressed caching method for agentic LLM workloads, significantly improving cache hit rates and energy efficiency by addressing token position shifts without architectural costs.

Contribution

The paper presents Irminsul, a novel content-addressed caching system that enhances performance and efficiency for agentic LLM serving by leveraging content hashing and delta-rotation techniques.

Findings

Recovers up to 83% of prompt tokens above exact-prefix

Achieves 63% prefill energy savings per cache hit

Effective across multiple large-scale MLA-MoE deployments

Abstract

Agentic LLM workloads put bit-identical tokens at shifted positions every turn, voiding prefix caches at the first byte of divergence. Operators report cache-hit regressions ranging from moderate slowdowns to severe TTFT spikes of 10-16s on unchanged content. Prior position-independent caching systems correct RoPE on the full $d_K$-dimensional key, an architectural cost imposed by GQA, not by caching itself. Multi-Head Latent Attention, deployed at scale in DeepSeek-V2/V3/R1, Kimi-K2/Moonlight, GLM-5, and Mistral Large 3, factors each KV row into a position-free $c_{KV}$ and a 64-dim $k_r$ correctable in closed form; this structure motivates content-addressed caching as a natural fit rather than a GQA workaround. We present Irminsul, which extends SGLang's radix cache with content-hash keying over CDC-chunked segments and a $\delta$-rotation rule for $k_r$. We evaluate three native MLA-MoE deployments - DeepSeek-V2-Lite (16B/2.4B), Kimi Moonlight-16B-A3B, and JoyAI-Flash (48B/3B) - with output-consistency on all three and recovery measured on the two endpoints; Irminsul recovers up to ~83% of prompt tokens above exact-prefix on agentic traffic while delivering 63% prefill energy savings per cache hit. We argue that content-addressed caching belongs in the serving stack as a first-class primitive, not a retrofit over prefix matching.