Skip to main content

Learn Learning paths Courses

GPU Scheduling and AI Batch Workloads on Kubernetes

Run real queues, quotas and gang scheduling on simulated GPUs

고급 · Lessons 25 · Lab 7

Curriculum

What a device plugin writes on the node

DRA: which device, not how many

Kueue: the queue admits, the scheduler places

Queue instead of reject reading
Set GPU quota and queue jobs with Kueue lab
Quiz: Kueue admission and quota quiz

Kueue: borrow, reclaim, preempt

What you borrow can always be reclaimed reading
Borrow, reclaim and preempt with Kueue lab
Quiz: cohorts and preemption quiz

JobSet: distributed training as one unit

When one dies, all restart reading
Group distributed training jobs with JobSet lab
Quiz: JobSet quiz

LeaderWorkerSet: one inference engine across nodes

KAI-Scheduler: deciding at placement time

How Dynamo runs on Kubernetes

Seeing it in layers, then fixing it upstream

Name the layer before the tool reading
Quiz: seeing it in layers quiz

Reference docs