GKE AI Inference Deployment Assistant
Deploy, tune, and autoscale GPU/TPU AI inference services on GKE.
The document requires authenticated gcloud access and GPU/TPU quota, and notes that Hugging Face tokens should be stored in a Kubernetes Secret; however, it does not specify least-privilege IAM, data flows, sensitive-data handling, pre-deployment confirmation, rollback, or deletion procedures. Because kubectl apply can cause external infrastructure changes, points are deducted. No malware, credential theft, or covert exfiltration behavior is evident.
The workflow from model discovery through manifest generation, deployment, and monitoring is broadly coherent and the commands are concrete; however, it lacks version pinning, command-availability checks, diagnostic failure handling, and rollback guidance. The troubleshooting table and several tables also contain malformed Markdown. Under static calibration, no execution was performed, so reliability remains below 10.
The audience, scenarios, and non-fit boundary for generic batch or HPC queues are reasonably clear, and the skill covers GPUs, TPUs, model servers, and autoscaling. Inputs and outputs, trigger precision, detailed non-fit cases, and environment compatibility are still underspecified, with no Chinese-language or mainland-China reachability discussion.
The skill includes front matter, prerequisites, a staged workflow, examples, optimization guidance, and troubleshooting. Repository context supplies Apache-2.0 licensing, installation, contribution, and active-development information. However, the skill lacks its own version, changelog, named maintenance responsibility, update path, dependency versions, known limitations, and FAQ; information layering is basic.
For common GKE inference deployments, the commands, ComputeClass example, HPA example, and optimization tips provide useful directly adaptable material and cover the core task. Still, there are no committed representative outputs, execution results, or complete production configurations; users must fill in secrets, PVCs, monitoring, model storage, and failure handling. The score is therefore moderate and below the static ceiling.
The skill contains inspectable commands, YAML, and troubleshooting content, providing limited auditability. It lacks a test suite, CI coverage, third-party execution evidence, versioned official citations, and cross-source corroboration. The assessment relies mainly on author-provided documentation, with no execution performed.
- The deployment examples modify external GKE/Kubernetes state without requiring explicit confirmation, permission review, or rollback; production users should review manifests, IAM, Secrets, PVCs, quotas, and costs first.
- The gcloud AI Profiles commands, model/accelerator combinations, ComputeClass API, GPU metric name, and TPU/GPU availability may vary by version and region; versions and official verification links are not pinned.
- Several tables are malformed and troubleshooting is incomplete; reachability of Hugging Face, model registries, and related Google services from mainland-China networks is not established.
What does this skill do, and when should you use it?
This skill supports teams running AI/ML inference workloads on Google Kubernetes Engine. It uses Google’s Inference Quickstart and LLM-serving practices to help select models, inference servers, and GPU or TPU accelerators, then generate Kubernetes manifests. It also covers ComputeClasses, GPU-metric autoscaling, and practical inference optimization. It is explicitly not intended for generic batch jobs or HPC task queues.
It uses gcloud container ai profiles to list supported models and valid model–accelerator–server combinations, then generates an inference.yaml Kubernetes manifest. The manifest can specify a model, vLLM/TGI/Triton/TensorRT-LLM server, GPU or TPU type, and an optional normalized time-per-output-token target. The skill provides a ComputeClass example, a GPU-metric HorizontalPodAutoscaler example, and guidance on quantization, batching, tensor parallelism, and vLLM KV-cache allocation. After generation, users can deploy with kubectl apply, inspect Pods and logs, and use the listed Kubernetes MCP tools to inspect resources, rollouts, and events.
- A platform engineer needs to deploy a Llama, Gemma, or Mistral model on GKE.
- An ML engineer needs to identify a valid accelerator and inference-server combination for a model.
- A cloud infrastructure team needs a GPU-targeting ComputeClass for inference on GKE Autopilot.
- An operations team needs to scale an LLM service using GPU utilization or request queue depth.
- A GKE user is troubleshooting an invalid combination, GPU quota exhaustion, GPU OOM, or slow model cold start.
What are this skill's strengths and limitations?
- Covers model and hardware discovery, manifest generation, deployment, monitoring, and troubleshooting.
- Names concrete server options including vLLM, TGI, Triton, and TensorRT-LLM.
- Includes accelerator selection, ComputeClasses, autoscaling, and memory-optimization guidance.
- Defines its boundary clearly by excluding generic batch and HPC workloads.
- Requires an existing GKE Autopilot cluster, authenticated gcloud CLI, and sufficient GPU or TPU quota.
- Some operations are explicitly CLI-only and cannot rely solely on the listed MCP tools.
- The source provides no test suite, platform-validation scope, or actual cost estimates.
- Some models may require Hugging Face tokens, PVCs, or manifest customization.
How do you install this skill?
Install the repository’s skill collection with npx skills add google/skills. The README says the installer lets users select specific skills from the repository, but it does not document a separate command or destination directory for installing only gke-inference.
How do you use this skill?
With an authenticated gcloud environment and a GKE Autopilot cluster that supports GPU workloads, use a prompt such as: “Deploy Gemma 2 9B on GKE with vLLM and an NVIDIA L4, and generate the inference manifest.” The skill can run gcloud container ai profiles models list --quiet, query combinations with gcloud container ai profiles list --model=gemma-2-9b-it --quiet, and generate inference.yaml with the manifest command. Review placeholders such as Hugging Face tokens and PVCs, then deploy with kubectl apply -f inference.yaml and monitor the Pods.
How does this skill compare with similar options?
Compared with the repository’s gke-batch-hpc skill, this skill focuses on AI/ML inference, model serving, and LLM autoscaling on GKE; generic batch jobs and HPC task queues should use gke-batch-hpc.