Trainer

Kubeflow Trainer is a Kubernetes native platform for distributed AI model training and large language model fine tuning. It is a CNCF project that gives machine learning teams a single TrainJob resource and a unified Python SDK, so the same job definition can run on a laptop or across a multi node GPU cluster. It supports PyTorch, JAX, DeepSpeed, MLX, HuggingFace, Megatron, XGBoost and Flux, and it handles the setup of distributed strategies such as DDP and FSDP along with gang scheduling across nodes. Fine tuning is supported through TorchTune, including LoRA, QLoRA and full fine tuning with your own HuggingFace models and datasets. Developers can test jobs locally with Docker or Podman before moving them to a cluster, and a distributed data cache built on Apache Arrow and Apache DataFusion streams tensors to GPU nodes. Job scheduling integrations include Kueue, Volcano, KAI Scheduler and Slurm Bridge, and runtimes are extensible through documented patches and templates. It is aimed at platform engineers and ML practitioners who already run Kubernetes and want a common way to launch training workloads across frameworks. The project is open source, and the documentation lists no paid plans or prices. The older Training Operator (v1) with its framework specific job types remains documented, with a migration guide to the current version.