AKS GPU Inference

AKS GPU Inference Day-2

Quick Reference

When to Use This Skill

Use for scheduling, VRAM/OOM, KAITO readiness, and workload scaling. Route exclusions as described above.

MCP Tools

Use a fitting host-advertised Azure read or the references’ read-only queries.

Host Capability Gate

Before executing any referenced pipeline, verify that the host authorizes the required shell and kubectl/az commands against the bound target. File-backed collection also requires approved artifact storage and access to any bundled files it uses. A governed Azure CLI tool alone does not establish these capabilities. Equivalent host reads may replace commands only where their advertised schemas provide the required evidence.

If execution is unavailable or prohibited, say so and analyze supplied or redacted events, status, logs, and metrics, or give the operator a scoped collection plan. State which reads did not run and leave conclusions requiring missing evidence unconfirmed. Never route kubectl through Azure MCP or bypass host policy. The authorization requirements below still apply.

Workflow

1. Bind subscription, cluster, kube context, pool, and affected resource.
2. Capture exact events/status and observed gpuProfile; state which reads ran.
3. Route to scheduling (opens in a new tab), observability (opens in a new tab), KAITO (opens in a new tab), or scaling (opens in a new tab).
4. Separate container/host OOM from device-allocation failures. For OOMKilled, correlate container memory limits/usage and node conditions. For Managed+Install device-memory evidence, use the exporter on port 19400 and incident-window FB_USED/FB_FREE; sizing tables do not establish a cause.
5. For KAITO not-ready, warn that Workspace deletion leaves its GPU pools; cleanup is separate and needs explicit authorization.
6. Report evidence, confidence, missing evidence, and owner handoff.

Require explicit authorization before scaling, cordon/drain, deletion, add-on enablement, or monitoring mutation.

Error Handling