[ANNOUNCEMENT] Kubeflow Trainer v2.1.0 is released!

[Andrey Velichkevich]

Hi Folks,

We’re excited to announce that Kubeflow Trainer v2.1.0 is officially released!

$ helm install kubeflow-trainer oci://ghcr.io/kubeflow/charts/kubeflow-trainer --version 2.1.0

$ pip install -U kubeflow

Major Highlights:

🚀 Distributed Data Cache

Load massive datasets efficiently in-memory, maximize GPU utilization with zero-copy transfer, and minimize I/O for large-scale pre- or post-training distributed AI workloads.

Learn more in the Kubeflow Trainer guide.

⏱️ Enhanced Kueue Integration

Topology Aware Scheduling for optimized Pod placement and reduced inter-node bandwidth – crucial for large-scale training on advanced GPUs like the GB200

Check out the Kueue documentation.

🔥 MLX Runtime with CUDA Support

Fine-tune and evaluate LLMs across multiple GPUs using MLX and mlx-lm with Kubernetes + OpenMPI.

🧩 Official Volcano Scheduler Support

Network topology awareness and advanced scheduling for improved TrainJob orchestration.

🧠 LLM Post-Training Enhancements

Supports LoRA, QLoRA, and DoRA for parameter-efficient fine-tuning with BuiltinTrainers.

Adds new post-training runtime for Qwen2.5.

📣 Release announcement: https://bit.ly/4qO8iM1

📝 Release notes: https://github.com/kubeflow/trainer/releases/tag/v2.1.0

Massive thanks to all contributors who made this possible!

Regards,

Andrey