Overview

At a glance

Simple resource declarations and bash UX

# my_task.yaml
name: tune
num_nodes: 2 # scale up your workload

resources:
  cpus: 15
  memory: 90
  accelerators: H100:8
  image_id: gcr.io/k8s-staging-jobset/pytorch-mnist:latest
  labels:
    kueue.x-k8s.io/queue-name: user-queue
    maxRunDurationSeconds: "3200"

run: |
  set -e
  NCCL_DEBUG=INFO torchrun --rdzv_id=123 --nnodes=$NUM_NODES --nproc_per_node=1 --master_addr=$MASTER_ADDR --master_port=1234 --node_rank=$RANK /workspace/mnist.py

Run with:

$ konduktor launch my_task.yaml

Out of the box observability: 📊 Trainy provides telemetry (prometheus metrics and logs) for tracking cluster performance, utilization, and health. Getting Started