AutoML & Training Working Group Updates
Talkat Kubeflow Summit Europe 2024 · on Tuesday, 2024-03-19· at 16:25· Pavilion 7 | Level 7.3 | Room S02·
summary
Roadmap items: Katib targets V1 API graduation, advanced parameter distributions like log-uniform, and native LLM hyperparameter tuning; Training Operator adds an LLM Python SDK, advanced suspend semantics and indexed jobs, plus plans for JAX support. Demo shows fine‑tuning a 108M‑parameter BERT on a Yelp subset across three GPUs from a Jupyter notebook using Hugging Face and create_job API, outputting a ~400MB model to S3 with no Docker or YAML. MPI Operator adds MPS support, launcher policies and multi‑cluster job dispatch.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:33Katib and training overview
- 9:25LLM fine-tuning demonstration
- 14:46New train API for LLMs
- 18:32Roadmap and orchestration
- 21:00Training and MPI operator updates
- 28:14Community and contributions
- 31:04Q&A session
related talks
20mEmpowering ML Workloads With Kubeflow: JAX Distributed Training and LL...
16mStreamline LLM Fine-tuning on Kubernetes With Kubeflow LLM Trainer
13mPreview: ML Training
26mAirbnb's LLM Evolution: Fine-Tuning with Ray
35mCreating Your Own LLM Tuning Platform with Open Source Technologies
21mOperationalizing ML Training Infra at Meta Scale
36mStreamlining ML Workflows With the Unified Kubeflow SDK
21mStreamline LLM Fine-Tuning on Kubernetes with Kubeflow LLM Trainer