Kubeflow Trainer is a Kubernetes-native distributed AI platform for multi-framework training and LLM fine-tuning, integrating MPI, Kueue, and data caching to stream data to GPUs.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Kubeflow Trainer is a Kubernetes-native distributed AI platform for scalable large language model (LLM) fine-tuning and training across multiple frameworks (including PyTorch, MLX, HuggingFace, DeepSpeed, JAX, XGBoost, and more). It provides MPI-based distributed execution, integrating with Kubernetes tools like Kueue for topology-aware scheduling and with JobSet and LeaderWorkerSet for orchestration. It includes a distributed data cache designed to stream large-scale data with zero-copy transfer directly to GPU nodes.
How it works
It orchestrates multi-node, multi-GPU distributed jobs across HPC clusters, using MPI for fast inter-process communication. It integrates with the Kubeflow Python SDK to enable development and fine-tuning of LLMs via TrainJob and Runtimes APIs. It emphasizes GPU utilization and memory efficiency through its data cache design and streaming mechanisms.
Getting started
Please check the official Kubeflow Trainer documentation to install and get started with Kubeflow Trainer.
Recent releases
- v2.3.0-rc.3 (2026-07-24)
- v2.3.0-rc.2 (2026-07-23)
- v2.3.0-rc.1 (2026-07-23)
- v2.3.0-rc.0 (2026-07-23)
- v2.2.1 (2026-06-18): This is the Kubeflow Trainer v2.2.1 release 🚀
Traction
stargazers: 2167 forks: 1011 open_issues: 141
Behind the repo
The project partners with the Kubeflow ecosystem and provides integration points for Kueue, JobSet, and LeaderWorkerSet as part of its orchestration capabilities. It states integration with Kubeflow Python SDK and multiple AI frameworks.
Caveats
license: Apache-2.0 created: 2017-06-28 last_push: 2026-08-04 language: Go