Kubeflow Ecosystem: What’s Next for Cloud Native AI/ML and LLMOps
Talkat KubeCon + CloudNativeCon Europe 2025 · on Thursday, 2025-04-03· at 11:00· Level 3 | ICC Capital Suite 14-16·
summary
Kubeflow announces the 1.0 release and stitches concrete LLMOps pieces: a donated Spark Operator now in the ecosystem with a benchmark of 60,000 Spark apps across 36,000 pods, a new Kubeflow Trainer B2 (supports PyTorch, DeepSpeed, MLX, TensorFlow; Hugging Face coming), model registry UI and model caching via PV/PVC, KServe + Envoy AI gateway for multi‑RM routing, and a demo fine‑tuning Gemma‑3‑1B with GRPO (Unsloth) via KFP that completes training and shows improving rewards.
AI-written summary · about
chapters (12)
- 0:00Introduction and panel
- 1:44Kubeflow ecosystem overview
- 3:40Kubeflow notebooks update
- 4:13Spark operator improvements
- 5:40Hyperparameter tuning with Katib
- 6:51Distributed training with Trainer
- 10:51Distributed data caching
- 11:52SDK and developer experience
- 14:29Model registry and management
- 15:53Model serving with KServe
- 17:46End-to-end LLMOps demo
- 27:20Roadmap and community call
related talks
18mKubeflow Ecosystem: Past, Present, Futu...
31mKubeflow 2026 State of the Union: Orchestrating the Full ML Lifecycle
44mEngaging the Kubeflow Community: Building an Enterprise-Ready AI/ML Platform
32mKubeflow Deep Dive
14mKubeflow: Where ML Meets Kubernetes
08mWelcome + Opening Remarks
26mKubernetes and MLOps for Scalable and Reproducible Generative AI
22mKeynote: Kubeflow ML on Kubernetes