# LMCache/LMCache

LMCache: Supercharge Your LLM with the Fastest KV Cache Layer

Repository: https://github.com/LMCache/LMCache
Canonical: https://ross.abutalabs.com/products/lmcache
Homepage: https://lmcache.ai/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: amd, cuda, inference, kv-cache, llm, pytorch, rocm, vllm, fast, speed
Last push: 2026-08-27T00:06:49+00:00

## Health v2 (maintenance only)
Score: 86/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 99, release rhythm 86, longevity 59
- inputs: {"age_days": 827, "days_push": 7, "days_rel": 13, "gap_med": 3, "n_releases_24m": 74}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 11460, forks 1774 (observed 2026-08-28T04:10:47.016478+00:00)

## What it is
LMCache is a KV cache management layer for LLM inference that stores, compresses, and reuses KV caches across requests, sessions, and serving engine instances. It reduces time-to-first-token and improves throughput for long-context, agentic, multi-turn, and RAG workloads, integrating with engines like vLLM on NVIDIA and AMD hardware.

## Use cases
- reduce time-to-first-token for LLM serving
- reuse KV cache across multi-turn conversations
- offload KV cache from GPU memory to tiered storage
- share KV cache across multiple vLLM instances
- speed up RAG and long-context inference
- monitor KV cache hit rates in production LLM serving

## When to choose
- you serve LLMs with vLLM or compatible engines and want lower TTFT and higher throughput
- your workloads are multi-turn, agentic, or RAG-heavy with repeated prompt prefixes
- you need persistent, engine-independent KV cache storage across CPU, local disk, and remote backends
- you run on NVIDIA or AMD GPUs and want vendor-neutral cache infrastructure

## When to avoid
- you need a general-purpose key-value database rather than LLM KV cache management
- your inference engine has no LMCache connector or KV cache integration support
- you run tiny single-user deployments where prefill cost is negligible

## Facets
- artifact type: library
- maturity: active
- function: caching, llm-inference, monitoring, gpu-computing
- domain: large-language-models, machine-learning, databases, performance, developer-tools
- platform: python, cloud
- tags: kv-cache, vllm, prefill-reuse, ttft-optimization, tiered-storage, cuda, rocm, inference-optimization, gpu, linux, docker

## Member repositories
- LMCache/LMCache (main) score 86

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:47.016478+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:15:34.702221+00:00, confidence not recorded.
  - readme: https://github.com/LMCache/LMCache (fetched 2026-08-28T04:10:47.016478+00:00, sha 93e820e2f457)
  - homepage: https://lmcache.ai/ (fetched 2026-08-29T08:14:20.726910+00:00, sha 7c5b0846ac38)
  - site_page: https://docs.lmcache.ai/ (fetched 2026-08-29T08:14:20.736127+00:00, sha 76c7f9abf8ee)
  - site_page: https://docs.lmcache.ai/getting_started/quickstart.html (fetched 2026-08-29T08:14:20.738859+00:00, sha 59c96ed058bf)
  - registry_pypi: https://pypi.org/pypi/lmcache/json (fetched 2026-08-29T08:14:20.741508+00:00, sha 9d1c52305ac2)
- Data as of 2026-08-30T08:39:29.467469+00:00.