A Scalable Platform for Training and Inference Using Kubeflow
Talkat Kubeflow Summit Europe 2024 · on Tuesday, 2024-03-19· at 13:30· Pavilion 7 | Level 7.3 | Room S02·
summary
CERN runs a Kubeflow-based platform that manages the full ML lifecycle from GitLab-built container training images to notebooks, TensorBoard and Triton serving, integrating EOS and Ceph-like S3 storage and site auth. A gn2 Transformer model (demo uses a 1.3M-parameter build) trains on distributed GPUs sourced from a centralized pool and shows ~4× improvement over older nets; GPUs are shared with time-slicing and MIG, jobs are launched via YAML pipelines, logs and ROC plots confirm the workflow works end-to-end.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 3:58Requirements
- 8:55GPU pool and sharing
- 13:11Physics use case
- 15:22Kubeflow pipeline demo
- 19:47ROC curve evaluation
- 20:33Conclusions
- 21:14Q&A
related talks
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
14mKubeflow: Where ML Meets Kubernetes
51mMachine Learning using Kubeflow and Kubernetes
24mA Better and More Efficient ML Experience for CERN Users
29mStreamlining Competitive Data Science at CERN: Running ML...
46mDeploying scalable Machine Learning models with Kubernetes
17mDevOps for Private AI: A Scalable Platform Using Kubeflow