Kubeflow Trainer

Kubeflow Trainer is the Kubeflow project’s component for distributed AI model training and LLM fine-tuning on Kubernetes. With around 2,300 stars on GitHub and active releases continuing into 2026, it gives platform teams a Kubernetes-native way to run large training jobs across GPU clusters with a consistent API.

Overview

Trainer v2 centers on the TrainJob API: you declare a training workload and a runtime — for example PyTorch with TorchTitan, DeepSpeed, JAX, XGBoost, or MPI — and Trainer handles pod orchestration, distributed coordination, and resource scheduling. Runtime templates can be reused and customized, so data scientists submit fine-tuning jobs without hand-writing Kubernetes manifests for every experiment.

Key Features

• TrainJob API for declarative distributed training on Kubernetes
• Built-in runtimes for PyTorch/TorchTitan, DeepSpeed, JAX, XGBoost, and MPI
• Purpose-built support for LLM fine-tuning at cluster scale
• Reusable, customizable runtime templates with a decoupled lifecycle
• GPU scheduling and multi-node coordination handled by the controller
• Part of the Kubeflow ecosystem, integrating with its pipelines and dashboards

The project is under active development: the Kubeflow Trainer v2.3.0 release added a runtime snapshot mechanism for a decoupled runtime lifecycle, enhanced MPI support, and Megatron-LM examples for very large model training. Trainer is governed as part of the Kubeflow community alongside components for pipelines, notebooks, and model serving, so jobs submitted through TrainJob can plug into a broader end-to-end ML platform when teams need one.

Trainer abstracts away the differences between training frameworks: the same TrainJob pattern can launch a PyTorch fine-tuning run today and an MPI-based HPC-style job tomorrow, while cluster administrators define the runtimes once and enforce resource policies centrally. Integration with Kubernetes scheduling features helps large organizations share scarce GPU capacity across teams without everyone hand-managing namespaces and quotas.

💳 Pricing

Kubeflow Trainer is free and open source under the Apache-2.0 license. Costs are limited to your own Kubernetes infrastructure.

Who Is It For

ML platform engineers and MLOps teams running distributed training or LLM fine-tuning on shared Kubernetes GPU clusters.