# HAMi

> HAMi (Heterogeneous AI Computing Virtualization Middleware) is a CNCF Incubating, Apache-2.0
> project that brings device sharing, hard memory isolation, enforced compute quotas and
> topology-aware scheduling to heterogeneous AI accelerators on Kubernetes. One physical GPU or NPU
> can be shared by several containers, across NVIDIA, Huawei Ascend, Cambricon MLU, Hygon DCU,
> Enflame GCU, Iluvatar, Moore Threads, MetaX, Kunlunxin and Vastai devices.

Provenance: GENERATED by API Evangelist on 2026-09-12 from this repository's apis.yml and artifacts.
HAMi does not publish an llms.txt of its own — https://project-hami.io/llms.txt returned 404 on
2026-09-12 — so this file is our summary of HAMi's public surface, not a document HAMi authored.
HAMi DOES publish machine-readable discovery of its own: an RFC 9727 API catalog, an OpenAPI 3.1
service description and an Agent Skills discovery index, all linked below.

What HAMi is not: a hosted service. There is no HAMi account, no API key and no pricing page. You run
it in your own Kubernetes cluster. Read this before assuming there is an endpoint to call.

## APIs

- [HAMi WebUI API](https://project-hami.io/docs/installation/webui-installation): Four gRPC services —
  Card, Node, Container and Monitor — over cluster accelerator inventory, node inventory,
  GPU-consuming workloads, and PromQL range/instant/summary queries. Eleven RPCs, all read-only. Each
  RPC carries a google.api.http annotation, so the same surface is served as REST through
  grpc-gateway under /v1 (POST /v1/gpus, GET /v1/node, POST /v1/monitor/query/range-vector, …). The
  contract is the .proto set, mirrored in this repository under grpc/. Self-hosted: the backend
  listens on :8000 (HTTP) and :9000 (gRPC) in-cluster and the Helm service publishes :3000; the docs
  say the WebUI is exposed via localhost only, reached with kubectl port-forward. No inbound
  authentication — access control is Kubernetes RBAC and network placement.
- [HAMi Website Discovery API](https://project-hami.io/docs/): Four anonymous GET endpoints on
  https://project-hami.io described by a first-party OpenAPI 3.1.0 — /robots.txt,
  /.well-known/api-catalog, /.well-known/agent-skills/index.json and /.well-known/site-status.json.

## Machine-readable discovery published by HAMi

- [API catalog (RFC 9727)](https://project-hami.io/.well-known/api-catalog): linkset declaring the
  service-desc, service-doc and status documents.
- [OpenAPI 3.1.0](https://project-hami.io/.well-known/openapi.json): the service description.
- [Agent Skills index (0.2.0)](https://project-hami.io/.well-known/agent-skills/index.json): the
  hami-docs skill, with a sha256 digest.
- [robots.txt](https://project-hami.io/robots.txt): carries Content-Signal
  (ai-train=no, search=yes, ai-input=yes) for ten named AI crawlers. Honour it.
- [OAuth protected-resource metadata](https://project-hami.io/.well-known/oauth-protected-resource):
  declares the resource public — empty authorization_servers and scopes_supported.
- No security.txt, no apis.json, no ai-plugin.json, no agent card on any HAMi host (checked
  2026-09-12).

## Agent Skills published by HAMi

- [hami-docs](https://project-hami.io/.well-known/agent-skills/hami-docs/SKILL.md): discover, search
  and cite HAMi documentation and release resources.
- [hami_vgpu_metrics_summarizer](https://github.com/Project-HAMi/HAMi/blob/master/skill/hami-vgpu-metrics-summary/SKILL.md):
  summarize HAMi vGPU metrics from Prometheus-style /metrics output by node, device, pod and namespace.
- [k8s-debug-pending-pod](https://github.com/Project-HAMi/HAMi/blob/master/skill/k8s-debug-gpu-pod/SKILL.md):
  triage pods stuck in Pending, CrashLoopBackOff or ImagePullBackOff.

## Documentation

- [Documentation](https://project-hami.io/docs/)
- [Get started — deploy with Helm](https://project-hami.io/docs/get-started/deploy-with-helm)
- [Developer guide](https://project-hami.io/docs/category/developer-guide)
- [Protocol design](https://project-hami.io/docs/developers/protocol): device registration handshake
  and the Pod allocation annotations the scheduler passes to the device plugin.
- [GPU utilization metrics](https://project-hami.io/docs/developers/gpu-utilization-metrics)
- [WebUI installation](https://project-hami.io/docs/installation/webui-installation)
- [Hardware support roadmap](https://project-hami.io/docs/contributor/roadmap)
- [Changelog](https://project-hami.io/changelog) — v2.10.0 released 2026-08-21.
- [Blog](https://project-hami.io/blog)
- [Community](https://project-hami.io/community)

## Distribution

- Helm chart `hami` 2.10.0 (2026-08-21) — https://project-hami.github.io/HAMi
- Helm chart `hami-webui` 1.3.0 (2026-08-29) — https://project-hami.github.io/HAMi-WebUI
- Go modules under github.com/Project-HAMi — HAMi-WebUI v1.3.0, HAMi-DRA v0.2.2,
  ascend-device-plugin v1.4.1, volcano-vgpu-device-plugin v1.12.0
- Container images under the `projecthami` Docker Hub organization
- The npm package named "hami" is an unrelated Vue.js plugin. It is not a HAMi SDK.

## Security

- [Security policy](https://github.com/Project-HAMi/HAMi/blob/master/SECURITY.md) — report privately
  via GitHub Security Advisories; no bug bounty. Security fixes for 2.8.x, 2.9.x and 2.10.x.
- [OpenSSF Security Insights 2.0.0](https://github.com/Project-HAMi/HAMi/blob/master/SECURITY-INSIGHTS.yml)
- Threat model, stated by the project: in-container enforcement limits GPU memory and compute for
  cooperative multi-tenant sharing on a trusted cluster, and is NOT a hard security boundary against a
  workload privileged enough to bypass its own hook.

## Optional

- [GitHub organization](https://github.com/Project-HAMi)
- [Terms of use](https://lfprojects.org/policies/terms-of-use/) and
  [privacy policy](https://lfprojects.org/policies/privacy-policy/) — LF Projects, LLC
- License: Apache-2.0