Kv-Cache

Mooncake: Turning KV Cache into a Distributed LLM Serving Layer

Mooncake reframes KV cache as a first-class distributed resource. This post walks through its architecture, the KV store design, and why prefilling and decoding benefit so differently from cache disaggregation.

Diagram of pages, GQA heads, and a prefix-sharing trie routing memory blocks.

Building a Paged KV Cache with GQA Sharing and Prefix-Reuse Eviction

TL;DR — A paged KV cache with grouped-query attention (GQA) sharing and a prefix-reuse eviction policy is one of the most credible LLM-systems projects you can ship to a CV in a weekend. It mirrors the design choices in vLLM, SGLang, and the RadixAttention paper, demonstrates that you understand memory hierarchies, attention math, and eviction, and produces a runnable artifact you can benchmark on real prompts. Why This Project Stands Out on a CV Hiring managers for inference, ML platform, and applied-LLM roles scan portfolios for evidence that a candidate can translate a paper into a working system. A paged KV cache hits that target harder than almost any other weekend project because the design space is genuinely contested in production: vLLM ships PagedAttention, SGLang ships RadixAttention, TensorRT-LLM ships in-flight batching with custom paged kernels, and HuggingFace TGI ships its own block manager. Building one from scratch proves you have read those systems and understood why they look the way they do. ...

Stylized illustration of stacked HBM dies next to a GPU die on an interposer.

GPU Memory in 2026: HBM, KV-Cache, and the Real Bottlenecks Behind LLM Inference

A working engineer’s guide to GPU memory: HBM bandwidth, KV-cache math, fragmentation, and the production patterns that keep large model inference honest.

Diagram showing token positions feeding a transformer decoder with KV cache blocks reused across steps.

KV Caching in Production: How Transformers Trade Memory for Latency

A practitioner’s guide to KV caching: what gets cached, why decode steps are cheap, and how production systems like vLLM and TensorRT-LLM squeeze out more throughput.

A diagram showing the stages of an LLM inference request moving from the API edge through the scheduler, KV cache, and GPU.

The LLM Inference Roadmap I Wish More Engineers Followed

Most teams treat LLM inference like a REST call. This roadmap walks through the layers that actually decide cost and latency in production — from continuous batching to KV cache layout to speculative decoding.