AI Workload Migration to GKE Inference
Move existing AI inference workloads to self-hosted inference on Google Kubernetes Engine.
The skill requires permission before read-only discovery, gates mutations, and includes controls for Workload Identity, token handling, external exposure, and internal load balancing. It lacks a complete rollback plan, digest-pinned images, dependency-security verification, and full data-flow disclosure, so points are deducted.
The four-phase workflow, checklist, and routing rules are fairly clear, but the supplied file is truncated in the VRAM formula section and the required assets are not provided. Static review cannot reproduce command execution, template generation, or validation; testing, edge-case coverage, and failure diagnostics therefore remain thin.
Invocation conditions, migration-versus-new-deployment boundaries, MCP off-ramp, and major inputs are explicit. However, Chinese-language support and mainland-China reachability are not established, and the workflow depends on Google Cloud, Hugging Face, and external documentation, so points are deducted.
The name, description, phased information architecture, installation entry point, Apache-2.0 license, and repository support channel are clear. The skill itself lacks a version, changelog, FAQ, named maintenance owner, and explicit update path; its asset dependencies are also not fully evidenced here.
The skill covers discovery, design, implementation, and validation with concrete GKE, vLLM, storage, gateway, and secret-handling guidance. However, the key templates and generated-file contents are absent, the formula section is truncated, and static evidence cannot establish directly usable migration output, so only limited points are awarded.
Concrete commands, template names, checklists, and reference links provide some auditability. There is no skill-specific test suite, CI coverage, or third-party execution evidence in the supplied material, and key paths cannot be reproduced statically, so points are deducted.
- The supplied skill evidence ends mid-way through the VRAM sizing formula, and the required assets are not included; do not assume the remaining templates or formula details are complete.
- Before execution, verify gcloud/kubectl permission scope, Workload Identity, image tags, accelerator quota, external-document reachability, and a rollback plan.
- Before external exposure of the vLLM endpoint, configure authentication or strict access controls; never place Hugging Face or other tokens in prompts, manifests, or logs.
What does this skill do, and when should you use it?
This skill is for AI inference workloads already running on Cloud Run, the Gemini API, Gemini Enterprise Agent Platform, or a custom VM. It uses a four-phase process covering discovery, solution design, implementation, and validation with traffic cutover. It inspects resources with gcloud and kubectl, designs compute, storage, networking, and serving components, and produces Kubernetes manifests. It is not intended for brand-new GKE inference deployments and does not perform automated migration through the Gemini Cloud Assist MCP server.
After receiving permission, it uses filtered gcloud list and describe commands to inspect Cloud Run services, GKE clusters, Vertex AI endpoints, and Cloud Storage; gathers model, hardware, storage, traffic, and autoscaling requirements; estimates VRAM from parameter count, KV-cache overhead, and a safety margin; plans vLLM, Custom Compute Classes, Workload Identity, Cloud Storage FUSE or Managed Lustre, Gateway API, and Prometheus; generates YAML for a vLLM Deployment, ComputeClass, Gateway, storage configuration, and model-staging Job; and provides kubectl commands for application, waiting, logging, rollout checks, port forwarding, inference testing, traffic cutover, and rollback.
- A team running an AI inference service on Cloud Run wants to move it to self-hosted GKE for more control over GPUs and the serving stack.
- A Gemini API or Gemini Enterprise Agent Platform user wants to evaluate replacing the managed source workload with an open-weight model on GKE.
- A platform engineer with a custom VM inference workload needs a GKE design covering accelerators, model storage, and traffic migration.
- An organization needs a concrete vLLM, GPU, model-staging, and internal Gateway plan for an existing large-model inference service.
What are this skill's strengths and limitations?
- Covers the migration lifecycle from discovery through validation and traffic cutover.
- Clearly separates read-only inspection from commands that change cloud resources.
- Provides concrete guidance for VRAM sizing, model staging, Workload Identity, Gateway configuration, and rollback.
- Includes security guardrails for Hugging Face tokens, external exposure, and unauthenticated inference endpoints.
- Focused on migrating existing workloads and unsuitable for fresh GKE inference deployments.
- Requires gcloud, kubectl, access to a Google Cloud environment, and local filesystem writes.
- The source provides no test suite, automated validation results, or evidence of platform-specific compatibility testing.
- The workflow still requires user approval of discovery findings, the design, and the final migration state.
How do you install this skill?
Run npx skills add google/skills, then select skills/cloud/google-cloud-solution-guided-gke-ai-migration from the repository during installation. The README does not document a more specific client installation directory or version requirement.
How do you use this skill?
In an Agent Skills-compatible client with the skill installed, use a prompt such as “Migrate my existing Cloud Run AI inference service to GKE, starting with environment discovery.” Before discovery, explicitly authorize read-only gcloud inspection. Later phases normally generate files and commands for the user to run unless the user explicitly asks the agent to execute mutations. The source requires using the current directory for migration-state.md and generated YAML manifests, but does not provide a complete trigger syntax or the referenced template files.
How does this skill compare with similar options?
Unlike the repository's gke-inference skill, this skill focuses on migrating existing workloads from Cloud Run, the Gemini API, Agent Platform, or other platforms to GKE; gke-inference is intended for new GKE AI model-server deployments. For automated infrastructure analysis or resource mutation, the source identifies the Gemini Cloud Assist MCP server as an alternative.