Streamline LLM Fine-Tuning on Kubernetes with Kubeflow LLM Trainer
summary
Kubeflow LLM Trainer v2 claims to let data scientists fine-tune LLMs on Kubernetes with a simple Python API (about ten lines of code or even one-line bootstrap), hiding k8s configs by separating train job from cluster training runtime. It builds on Kubeflow Trainer v2 + Torch 2, runs a demo bootstrapping three pods: dataset initializer, model initializer, trainer to fine-tune Llama 3.2 1B on an Alpaca-like dataset, writes checkpoints to a PVC (PVC creation and cloud exporter are still partially manual/not yet implemented).
AI-written summary · about
chapters (8)
- 0:00Kubeflow LLM Trainer intro
- 3:22LLM fine-tuning pain points
- 8:05Kubeflow Trainer V2 simplifies
- 10:22Kubeflow M Trainer V2 lifecycle
- 12:10V2 vs V1 comparison
- 16:54Key improvements in V2
- 17:40Summary and roadmap
- 20:11Q&A
related talks
16mStreamline LLM Fine-tuning on Kubernetes With Kubeflow LLM Trainer
20mEmpowering ML Workloads With Kubeflow: JAX Distributed Training and LL...
36mStreamlining ML Workflows With the Unified Kubeflow SDK
51mMachine Learning using Kubeflow and Kubernetes
32mRunning and Fine-tuning Open Source LLMs on Google Kubernetes Engine
54mScaling Machine Learning Pipelines in Cloud
25mFine-Tuning Large Language Models With Declarative ML Orchestration
18mKubeflow Ecosystem: Past, Present, Futu...