hami

v2.9.0AI & ML

2026年5月19日リンクをコピー

v2.9.0 adds HAMi-DRA for NVIDIA (now production-ready), Ascend HAMi-core mode, VastAI support, and patches a scheduler DoS vulnerability. Prometheus metric renames require dashboard updates before upgrading.

• securityPatch scheduler DoS vector and Go security upgrades

  Two scheduler-level security fixes land in this release: an io.LimitReader guard on scheduler HTTP routes to prevent DoS (issue #554), and a Go runtime upgrade to 1.26.2 for upstream security fixes. If you run HAMi scheduler exposed to any untrusted network path, upgrade promptly.

• enhancementHAMi-DRA for NVIDIA is production-ready — start evaluating

  HAMi-DRA (Dynamic Resource Allocation) for NVIDIA is now marked ready for use. If you're on Kubernetes 1.26+ and want finer-grained GPU resource management without relying solely on device plugins, this is the release to start evaluating DRA. Test in a non-prod cluster first — DRA changes how the scheduler sees GPU resources.

• breakingPrometheus metric names changed — update dashboards before upgrading

  Prometheus metric and label names have been realigned to follow best practices (renamed fields). If you have dashboards or alerts built against HAMi vGPU metrics, audit your metric names after upgrading. The existing dashboard.md has been updated — cross-reference it. The new ServiceMonitor Helm chart options also make scrape config cleaner if you're on the Prometheus Operator stack.