kueue

Kueue extension for Tekton

Seven hands-on labs for running AI workloads on Kubernetes: GPU scheduling, distributed training, model serving, and agents. Runs on a local kind cluster, no GPU required. Companion to a KubeCon EU 2026 talk.

GPU Labs

Manage heterogeneous accelerators, vGPU slicing, shared KV cache, model delivery, and inference workloads.

Make Kueue the single admission controller for mixed Slurm, Ray, and Kubernetes workloads on one shared pool. Experimental, live-validated prototype. Apache-2.0.

Checkpoint readiness, recovery, and mobility control plane for AI training clusters.

Multi-backend LLM serving and training platform — vLLM/Triton/Ray Serve/KServe/BentoML behind one contract, Kueue/Karpenter GPU orchestration, Ray Train/FSDP/DeepSpeed with LoRA/PEFT and DVC, MLflow/W&B tracking, and a tool-grounded LangGraph advisor — CI-validated without real GPU cost.

Ground-side mission plan compiler for ORCHIDE-class cloud-native satellite platforms. Validates plans, applies OPA policy guardrails, and renders Argo/Kueue workflow artifacts.

Job scheduling api for finetuning ML models on clusters

AdmiRL: Learning-Augmented Admission Control for Gang and Elastic GPU Workloads in Kueue

A deterministic simulator for GPU quota borrowing, preemption, and requeue.

Atomic N-or-0 EKS node provisioning for gang jobs through a Kueue AdmissionCheck

Alethia starter template — RAG, a vector database, CPU model serving and Kueue batch queueing, split across both trust levels. CPU-only by default; no GPU is provisioned.

Production-style Kubernetes MLOps control plane with MLflow 3, Airflow 3.3, KServe/Kueue architecture labs, a tested KubeOps console, failure drills, and a study guide.

A OpenShift Console Plugin to create Queues for Reserved GPUs + kubevirt Resources for VMs on OpenShift Virtualization. It also provides an Controller for dealing with Kueue artifacts

A local Kubernetes lab to learn GPU scheduling without owning GPU Hardware

Production-style model reliability platform with drift checks, durable incidents, transactional notifications, RCA and lineage, a tested SignalOps console, and a study guide.

Benchmark Kubernetes GPU schedulers (default, Kueue, Volcano, KAI) on identical AI workload traces: utilisation, fragmentation, queue wait, fairness and cost. No GPUs needed.

MultiKueue setup walkthrough for multi-cluster Kubernetes batch scheduling

Kubernetes job queue system for research computing