Streamlining ML Workflows With the Unified Kubeflow SDK
summary
Unified Kubeflow SDK (v0.2) packages Trainer, Katib and Pipelines APIs into one pip install kubeflow workflow and exposes a consistent Python surface: use pip install kubeflow, call train to launch a Job/JobSet that references a preconfigured training runtime, and call optimize to spin up Katib HPO. Demo trains a PyTorch CNN on FashionMNIST with distributed training across 2 nodes (1 GPU each), runs 10 trials with 2 parallel trials and improves loss from ~0.41 to ~0.30. Local execution supports local process runtime (virtualenv, single-node, PyTorch only) and a container backend (Docker/Podman, multi-node, NVIDIA toolkit required).
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:54AI workloads at scale
- 4:46Kubeflow Unified SDK
- 6:37Training and hyperparameter tuning demo
- 15:37Kubernetes backend details
- 19:27Local execution in SDK 0.2
- 24:25Roadmap and future
- 25:51Community contributions
- 28:03Q&A
related talks
54mScaling Machine Learning Pipelines in Cloud
31mKubeflow 2026 State of the Union: Orchestrating the Full ML Lifecycle
51mMachine Learning using Kubeflow and Kubernetes
14mKubeflow: Where ML Meets Kubernetes
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
16mBringing Kubeflow Training Local: SDK-Driven “Local-exec”...
36mMachine learning pipeline Kubeflow, let's start
23mEfficient Integration of Kubeflow Tools: An ISS Data Case Study on Anomaly Detection