Lightning Talk: Dynamic GPU Fraction and Sharing With Cloud Native Principle
Lightningat KubeCon + CloudNativeCon China 2025 · on Tuesday, 2025-06-10· at 16:19· Level 16 | Grand Ballroom I·
summary
Claims a working PoC called GPU.x that uses a Python interposer to intercept ML-framework calls (PyTorch, TensorFlow) and enforce dynamic fractional GPU memory allocation without changing application code. It integrates with Kubernetes to read GPU policies, implements live isolation between workloads, and demos partitioning one GPU between two jobs; exceeding an allocation triggers OOM. Lightweight, framework-level approach avoids low-level CUDA runtime builds and can extend to other accelerators.
AI-written summary · about
related talks
09mLightning Talk: Bringing Multi-Tenancy to GPU-Based Workloads
05mLightning Talk: Sharing a GPU Among Multiple Containers
10mLightning Talk: K8s Issue #52757: Sharing GPUs Among Multiple Containers
34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments
07mProject Lightning Talk: From Static Slices To Elastic GPUs: Dynamic MIG With HAMI
29mSharing is Caring – Fractional GPU Allocations With MetaGPU Device Plugin
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
38mMinimizing GPU Cost for Your Deep Learning on Kubernetes