Skip to content

v1.2.0

Latest

Choose a tag to compare

@github-actions github-actions released this 02 Nov 22:03
· 12 commits to main since this release
Immutable release. Only release title and notes can be modified.
935e074

This release adds NVKE GB200 recipes with CNCF AI conformance evidence and CKS training previews, updates GPU stack components, introduces the AICR Operator preview and core and ops bundles, and begins the NKTP telemetry pipeline, alongside opt-in NVSentinel features and new bundle output options. It also welcomes 7 new contributors.

Highlights

New and Updated Recipes

  • CKS H100 and RTX PRO 6000 - training previews on CoreWeave Kubernetes Service (NVIDIA/aicr#3527, NVIDIA/aicr#3526)
  • NVKE GB200 - training and Kubeflow training, details below
  • SLES on RKE2 - nodes are now tuned with the nvidia-tuned generic profile (NVIDIA/aicr#3336)
  • Rocky Linux - recognized as an OS criteria value (--os rocky) (NVIDIA/aicr#3332)
  • nvidia-setup moves to 0.9.0 and the EKS setup chain is reordered, moving EKS nodes to the 7.0 kernel (NVIDIA/aicr#3519)
  • nvidia-tuned 0.11.0, nvidia-dra-driver-gpu 0.5.1, nfd 0.20.1, and topograph 1.1.0
  • k8s-aibom moves to v1.6.2 and ships in the GKE Dynamo recipes (NVIDIA/aicr#3373, NVIDIA/aicr#3375)
  • gpu-operator operands tolerate NVSentinel's release taint (NVIDIA/aicr#3360)

NVKE GB200 and CNCF AI Conformance - A new NVKE GB200 Ubuntu training recipe uses SR-IOV InfiniBand and deploys Multus and the CNI plugins (NVIDIA/aicr#3295, NVIDIA/aicr#3468). A Kubeflow variant is included. AICR's NVKE v1.36 CNCF AI conformance submission is published with its evidence (NVIDIA/aicr#3455).

AICR Operator (Preview) - The opt-in operator runs a pinned AICR image in-cluster, validates each bundle's recipe, fingerprints inventory specs, and gates bundle rollout on approval (NVIDIA/aicr#3475, NVIDIA/aicr#3490). A default in-cluster registry removes the need for an external registry; operator images ship in the release train (NVIDIA/aicr#3533, NVIDIA/aicr#3491). The v1.2.0 operator defaults to AICR v1.1.2, the release with which it was tested. See the operator support policy to run another release.

Core and Ops Bundles - Per ADR-018, every registry component now declares a class (NVIDIA/aicr#3350):

  • core covers what a plain GPU pod needs to run, with its fabric and install prerequisites.
  • ops covers everything else, including workload platforms.

aicr bundle --split renders a class-partitioned bundle set, so the core GPU stack can be deployed and owned separately from the ops layer (NVIDIA/aicr#3420).

NKTP Telemetry (Initial Implementation) - ADR-026 adopts the NVKE Telemetry Pipeline (NKTP) as AICR's telemetry pipeline (NVIDIA/aicr#3390). This release lays the groundwork:

Switching the stock recipes to NKTP comes in a later release.

NVSentinel - New opt-in mixins add new-node validation with a DCGM diagnostic (NVIDIA/aicr#3464), an event exporter (NVIDIA/aicr#3371), and a log collector (NVIDIA/aicr#3354). AICR now owns node-drainer drain scope and timing (NVIDIA/aicr#3415).

Other Improvements

Security - The Go toolchain moves to 1.27.3 for the crypto/x509 CA distrust, sigstore/rekor to v1.5.5 for a go-chi CVE, and the Percona MongoDB and aiperf-bench images are refreshed (NVIDIA/aicr#3377, NVIDIA/aicr#3516, NVIDIA/aicr#3524).

Deprecations

  • Argo CD options as deployer: value overrides - --set deployer:<key>=<value>, the matching POST /v1/bundle set value, and deployer: entries in spec.bundle.deployment.set are deprecated. Use the dedicated --argocd-* flags, spec.bundle.deployment.argocd config fields, or REST parameters instead. Removal is planned for v2.0.0 (deprecations).

New Contributors - Seven contributors landed their first change in this release, from new recipes and OS support to validator and attestation fixes:

Thanks to @ArangoGutierrez, @atif1996, @ayuskauskas, @csy20, @depohmel, @ErikJiang, @git-jxj, @gusraggio, @La1itchauhan, @lockwobr, @mikecook, @mlsorensen, @mohityadav8, @njhensley, @ravisoundar, @robbert229, @rorajani, @srao-nv, @varmesh, @vineeth-bandi, @vjsairam, @yuanchen8911, and @mchmarny.


Changelog

New Features

Bug Fixes

Other Tasks