DeepOps 26.07 is the second release of the modern CalVer train, following 26.05 on a six-week cadence.
General
- New DGX software stack documentation and refreshed
nvidia-dgxrole covering DGX OS 6 (Ubuntu 22.04), DGX OS 7 (Ubuntu 24.04), DGX Software for RHEL 8, and Enterprise Linux 9, while preserving legacy Ubuntu 18.04/20.04 and RHEL 7 paths (#1346). - Clarified current release validation targets versus legacy/community-supported OS paths across the README and docs (#1354).
- Script debug output now reports a meaningful DeepOps version derived from the checkout's git tag description, with a
config/env.shoverride (#1357).
Container runtime and OS support
- New
nvidia_container_toolkitrole: Ubuntu 24.04+ hosts install the NVIDIA Container Toolkit from the current libnvidia-container repository and configure Docker withnvidia-ctk(#1347). - Red Hat family 8+ hosts now use the same role via the official NVIDIA Container Toolkit RPM repository; RPM airgap docs refreshed for EL8/EL9 (#1350).
- Container Toolkit airgap docs moved from the legacy
nvidia-dockerwrapper to mirrored package repositories plusnvidia-ctkconfiguration (#1352).
Provisioning
- Retired DeepOps-owned DGXIE, minimal PXE/Pixiecore, and Ubuntu 16.04/18.04 preseed assets. MAAS owns machine lifecycle, DHCP/PXE, images, and deployment; DeepOps consumes MAAS tags and dynamic inventory (#1351).
Kubernetes
- GPU Operator chart default updated to v26.3.3; Kubernetes GPU device plugin and GPU feature discovery charts updated to 0.19.3; MIG manager packages updated to mig-parted v0.14.2; Network Operator default updated to 26.4.0 (#1357).
- Kubernetes storage helper support boundaries clarified: the NFS subdir provisioner remains the simple default path (chart 4.0.18), NetApp Trident defaults/docs refreshed to Trident 26.02.1 with opt-in backend creation, and the deprecated Rook/Ceph helper now requires explicit opt-in (#1358).
Slurm
- Slurm default version updated to 26.05.1, the current upstream stable line (#1357).
- DCGM exporter refreshed to 4.5.3-4.8.2 with current upstream metrics CSVs and a distroless-image-compatible DaemonSet (#1348).
Monitoring
- Prometheus v3.13.0, Alertmanager v0.33.0, Grafana 13.1.0, node exporter v1.11.1, and kube-prometheus-stack chart 87.5.1 (#1348, #1357).
Examples and documentation
- CUDA/NGC example images refreshed to current tags across airgap docs, Kubernetes workload examples, and smoke scripts (#1349).
- RAPIDS/Dask, rootless Docker, RoCE/OFED, NGC-Ready workload, Slurm validation, and registry cache examples refreshed to current NGC and Kubernetes references; new Kubernetes RDMA deployments are pointed at the NVIDIA Network Operator with the DeepOps RoCE role retained as a legacy direct-host path (#1356).
- Jenkins-era CI harness and Vagrant virtual lab documented as legacy/community-supported reference material (#1353).
Agent operability
- New
AGENTS.mdoperating guide: repository map, first-time setup, Slurm and Kubernetes golden paths, operating rules, and known gotchas that read as failures but are not (#1360). - New
scripts/validation/machine-readable validation tools for automation and AI agents:deepops_doctor.pyprovisioning preflight,validate_slurm.py, andvalidate_k8s.py, each with--jsonflat output, explicit failure lists, and strict exit codes (#1359). - New
skills/directory with four reusable procedures (deploy-slurm-cluster, deploy-k8s-gpu-cluster, validate-gpu-cluster, diagnose-driver-install) in the cross-tool agent-skills format (#1360).
Fixes
parse_manifest.pyfirmware version parsing no longer aborts on a non-dict parseable manifest line; adds a regression test (#1355).
Known issues and validation notes
- Component support boundaries recorded for this release: registry/cache and the refreshed workload examples are supported paths; the RoCE backend role is a documented legacy direct-host path; Open OnDemand, the OpenShift helper role, and the Singularity wrapper are community-supported, with ownership calls planned next release.
- Release validation ran on single-node GPU hardware: fresh Slurm deployment with a GPU job, fresh Kubernetes deployment with GPU Operator and a CUDA pod, and a MAAS provisioning smoke. Multi-node scheduling and Network Operator install behavior on RDMA-capable networking hardware were not covered in this cycle.
- Direct in-place Kubernetes upgrades from 23.08 remain blocked by Kubespray's Calico minimum-version prechecks; use the documented staged upgrade or redeploy path for much older clusters.