Just a little over one week ahead of KubeCon + CloudNativeCon EU 2026, the Kubeflow team is excited to ship Trainer v2.2. The v2.2 release reinforces our commitment to expanding the Kubeflow Trainer ecosystem – meeting developers where they are by adding native support for JAX, XGBoost, and Flux, while also delivering deeper observability into training jobs.

Kubeflow

Kubeflow Trainer v2.2 ships with several major additions: native JAX and XGBoost distributed training runtimes on Kubernetes, Flux Framework integration for HPC and MPI bootstrapping, real-time TrainJob progress and metrics observability (with Hugging Face Transformers integration), a new activeDeadlineSeconds API for job timeouts, and a RuntimePatches API replacing PodTemplateOverrides for multi-controller environments. Breaking changes include removal of PodTemplateOverrides, ElasticPolicy API, and numProcPerNode from Torch MLPolicy. The roadmap includes Multi-Node NVLink support, automatic GPU resource configuration, and Workload-Aware Scheduling with native Kubernetes gang-scheduling.

Kubeflow Trainer v2.2: JAX & XGBoost Runtimes, Flux for HPC Support, and TrainJob progress and metrics observability

Bringing XGBoost to Kubernetes with Trainer

Track TrainJob Progress and Expose Metrics

Bringing Flux Framework for HPC and MPI Bootstrapping

RuntimePatches API to override TrainJob defaults