KubeCon + CloudNativeCon China 2025

As AI/ML workloads continue to scale in complexity, developers and platform engineers are pushing Kubernetes beyond typical MLOps boundaries.

This talk dives into strategies for orchestrating GPU-accelerated training and inference across large-scale clusters -integrating HPC principles, operator-based scheduling, and novel debugging workflows.

Attendees will learn how to implement fine-grained GPU partitioning, harness ephemeral containers to probe and adjust multi-node training in real time, and adopt eBPF-driven instrumentation for low-overhead kernel-level performance insights. We’ll explore cutting-edge scheduling optimizations—like reinforcement-learning approaches and HPC-inspired batch-queuing orchestration on Kubernetes that dynamically respond to heterogeneous job demands.

Real-world case studies will highlight HPC integration scenarios (RDMA, GPU Direct) for data-parallel workloads and complex training frameworks such as Horovod, Ray, and Spark on Kubernetes.