# vllm-project/production-stack

vLLM’s reference system for K8S-native cluster-wide deployment with community-driven performance optimization

Repository: https://github.com/vllm-project/production-stack
Canonical: https://ross.abutalabs.com/products/production-stack
Homepage: https://docs.vllm.ai/projects/production-stack
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-08-18T19:39:12+00:00

## Health v2 (maintenance only)
Score: 85/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 98, release rhythm 94, longevity 42
- inputs: {"age_days": 589, "days_push": 15, "days_rel": 40, "gap_med": 10.5, "n_releases_24m": 23}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2526, forks 472 (observed 2026-08-28T04:06:58.610582+00:00)

## What it is
vLLM Production Stack is a reference implementation for deploying vLLM LLM inference at scale on Kubernetes, distributed via Helm charts. It adds a request router with KV-cache-aware, prefix-aware, and load-aware routing, plus KV cache offloading via LMCache and Prometheus-based monitoring.

## Use cases
- deploy vllm on kubernetes
- scale llm inference across multiple gpu nodes
- route llm requests with kv cache aware load balancing
- share kv cache across vllm instances with lmcach
- deploy openai-compatible llm api on aws or gcp
- monitor vllm inference metrics with prometheus
- set up disaggregated prefill for llm serving
- enable tool calling for llama models in production

## When to choose
- you already use vLLM and need cluster-wide, multi-replica deployment on Kubernetes
- you want smart request routing (prefix/KV-cache/load aware) across vLLM instances
- you need KV cache offloading and sharing to improve throughput and fault tolerance
- you want an OpenAI-compatible LLM endpoint with built-in monitoring and autoscaling

## When to avoid
- you only need a single vLLM instance on one machine without Kubernetes
- you use a different inference engine like TGI, SGLang, or TensorRT-LLM
- you cannot run a Kubernetes cluster or Helm
- you need a fully managed LLM API rather than self-hosted serving

## Facets
- artifact type: framework
- maturity: active
- function: llm-inference, deployment, monitoring, caching, container-orchestration, infrastructure-as-code, api-framework, benchmarking, tracing
- domain: large-language-models, artificial-intelligence, cloud-computing, self-hosted, microservices
- platform: python, cloud, self-hosted
- tags: vllm, helm-chart, kv-cache, lmcache, request-routing, openai-compatible-api, llm-serving, inference-serving, load-balancing, prefix-caching, devops, containers, kubernetes, docker, gpu, linux

## Member repositories
- vllm-project/production-stack (main) score 85

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:58.610582+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 2026-08-30T02:25:29.701754+00:00, confidence not recorded.
  - readme: https://github.com/vllm-project/production-stack (fetched 2026-08-28T04:06:58.610582+00:00, sha b514e3a9f77f)
  - homepage: https://docs.vllm.ai/projects/production-stack (fetched 2026-08-29T10:08:05.338696+00:00, sha c69e86ed0631)
  - site_page: https://docs.vllm.ai/projects/production-stack/en/latest/getting_started/quickstart.html (fetched 2026-08-29T10:08:05.347739+00:00, sha 32af1506795c)
  - site_page: https://docs.vllm.ai/projects/production-stack/en/latest/getting_started/faq.html (fetched 2026-08-29T10:08:05.349965+00:00, sha 7826f84e37d9)
  - site_page: https://docs.vllm.ai/projects/production-stack/en/latest/use_cases/tool-enabled-installation.html (fetched 2026-08-29T10:08:05.351617+00:00, sha b2bc9d5e3520)
  - site_page: https://docs.vllm.ai/projects/production-stack/en/latest/use_cases/sharing-kv-cache.html (fetched 2026-08-29T10:08:05.353378+00:00, sha 7c75603fd497)
- Data as of 2026-08-30T08:39:29.467469+00:00.