What problem does it solve? Running distributed machine learning training on Kubernetes requires coordinating framework-specific CRDs, gang scheduling, GPU allocation, and job orchestration, which is error-prone when done manually. This Skill provides the operational knowledge to define and manage Kubeflow Trainer v2.2 workloads using the unified TrainJob and TrainingRuntime CRDs. ## Core Features & Use Cases - Framework-agnostic training jobs: Define TrainJobs for PyTorch, TensorFlow, JAX, XGBoost, MPI, and Flux using a single CRD with TrainingRuntime execution templates. - Scheduling integration: Configure gang scheduling via coscheduling or Volcano, and queue-based admission with Kueue multikueue for shared GPU clusters. - Ecosystem integration: Orchestrate TrainJobs inside Kubeflow Pipelines v2 components and use them as trial templates for Katib hyperparameter tuning. - Use Case: A platform engineer needs to run a 4-node PyTorch training job with 2 GPUs per node on a shared cluster. The Skill provides the TrainJob manifest, a TrainingRuntime with torch mlPolicy and coscheduling, plus guidance on Kueue queue labels and common pitfalls like GPU count mismatches. ## Quick Start Ask the agent to create a Kubeflow Trainer TrainJob and TrainingRuntime for a multi-node PyTorch training workload with GPU limits and gang scheduling.
