Project Lightning Talk: Extend Large Language Model Training Beyond Single Kubernetes Cluster
Lightningat KubeCon + CloudNativeCon Europe 2025 · on Tuesday, 2025-04-01· at 15:04· Platinum Suite | Level 3·
summary
Demonstrates running LLM training across clusters by combining Volcano and Karmada to overcome single-cluster limits (default ~5,000 nodes) and huge GPU + network needs, implementing a unified API for federation, a meta-framework that exposes workload networking/storage to the infra, and cross-cluster job queuing with policies to prevent busy waiting and deadlock. Shows network-aware scheduling and a scalable Volcano Global component that improves multi-cluster utilization and training performance.
AI-written summary · about
related talks
12mLightning Talk: Benchmarking Your Distributed ML Training on the K8s Platform
22mIntroduction to Large Language Models (LLM) on Kubernetes
37mImproving Performance of Deep Learning Workloads With Volcano
08mKeynote: Plug in and Scale: Serving LLM Models on Kubernetes Ma...
39mBeyond Model Sharding: Atomic Scheduling and Disaggregated LLM Serving With ...
16mStreamline LLM Fine-tuning on Kubernetes With Kubeflow LLM Trainer
46mMachine Learning with Kubernetes
30mVolcano: A Unified Scheduling Platform for Cloud Native AI