skills.rest
Envision-Construction
Envision-ConstructionOFFICIAL
Explore all skills by this author →
GitHub

GKE AI Platform

Provision GKE clusters, schedule GPU batches with Kueue, and serve models with vLLM or TGI.

1|Updated May 17, 2026
Search Tags:#kubernetes#terraform#gpu#gke#vllm#model-serving#kueue
One-click install
npx skills add https://github.com/Envision-Construction/Envision-Skill-Repo --skill gke-ai-platform-envision-construction
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: GKE AI Platform
Source: https://github.com/Envision-Construction/Envision-Skill-Repo/tree/main/plugins/infra/gke/skills/ai-platform
Command: npx skills add https://github.com/Envision-Construction/Envision-Skill-Repo --skill gke-ai-platform-envision-construction

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Deploying production AI infrastructure on Google Kubernetes Engine requires coordinating Terraform cluster provisioning, GPU quota management, batch job scheduling, and model serving, which is error-prone when done ad hoc. ## Core Features & Use Cases - Cluster Provisioning: Terraform modules for private Standard/Autopilot GKE clusters with GPU (L4, H100) and TPU node pools, GCS Fuse CSI, and Ray operator support. - GPU Batch Scheduling: Kueue ResourceFlavors, ClusterQueues, and priority classes implementing a Reserved → On-demand → Spot cost ladder with cohort borrowing and preemption. - Model Inference Serving: vLLM and TGI deployment manifests with GCS Fuse model loading, image streaming, and GPU-metric HPA for cold-start optimization. - Parallel Task Dispatch: Lossless wave executor that dispatches agentic framework tasks (GSD, Ralph Loop) as Kubernetes Indexed Jobs with idempotent replay and crash recovery. - Use Case: A platform engineer needs to run a multi-team GPU training platform — apply the Terraform configs, install the Kueue queues, then dispatch parallel analysis waves to the cluster. ## Quick Start Ask the assistant to provision a private GKE cluster with an L4 GPU node pool in project claude-mcp-457317 using the Terraform patterns in this skill.

Frequently Asked Questions about GKE AI Platform

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I provision a GKE cluster with GPU node pools using Terraform?▼

Use the terraform-google-modules kubernetes-engine private-cluster module with node_pools combining CPU and GPU pools. Set accelerator_type to nvidia-l4 with machine_type g2-standard-24, and always pin node_locations to avoid quota failures from random zone selection.

How does Kueue schedule GPU batch jobs across spot and reserved capacity?▼

Kueue tries ResourceFlavors in the order listed in the ClusterQueue, so listing reserved first, then on-demand, then spot creates a cost-optimized fallback ladder. Cohort borrowing lets teams share unused quota, capped by borrowingLimit.

vLLM vs TGI for model inference on GKE?▼

Both are supported with near-identical manifests requesting nvidia.com/gpu resources. vLLM uses --tensor-parallel-size for multi-GPU sharding while TGI uses NUM_SHARD; either can load models via an init container or GCS Fuse CSI volumes.

How do I reduce cold-start latency for GPU inference pods on GKE?▼

Enable image streaming with gcloud container clusters update --enable-image-streaming, cutting pull time from roughly 190s to 30s. Add local SSDs to node pools, use zstd-compressed images, or run a preloader DaemonSet on GPU nodes.

Can I dispatch parallel agent tasks to GKE from frameworks like GSD?▼

Yes, the dispatch workflow normalizes a wave manifest of tasks, uploads inputs to GCS, and runs them as a Kubernetes Indexed Job or pre-warmed pod pool. Completed tasks are skipped idempotently, and re-running with the same wave_id resumes from checkpoint.

Why do GPU workloads fail to schedule on GKE Autopilot?▼

GPU pods on Autopilot can take over 10 minutes to provision nodes, so set progressDeadlineSeconds to 1800 on deployments. Also verify GPU quota in the selected zones and use nodeSelector with cloud.google.com/gke-accelerator labels.

Related Skills

Explore More Software Engineering Skills→
lloydchanglloydchang
2

manage-container-registry

Automate container registry operations across AWS, Azure, GCP, and on-prem environments.

Community
Advanced
coreweavecoreweave
7

setup-sunk-gpu-monitoring

Set up GPU and Slurm metrics collection on a SUNK cluster.

Official
Advanced
arbazkhan971arbazkhan971
26

infra

Validate and manage Terraform, CloudFormation, Pulumi, and CDK deployments.

Community
Advanced
skypilot-orgskypilot-org
10.5k

skypilot

Launch and manage AI workloads across clouds, Kubernetes, and Slurm clusters.

Official
Advanced