How to set GPU memory limit via accelerator API for vGPU resources like nvidia.com/gpumem

[HisaYing]

Currently KFP SDK v2 provides official API to configure GPU card count via accelerator interface:

task.set_accelerator_type("nvidia.com/gpu")
task.set_accelerator_limit("1")

This compiles to Kubernetes resource limits: {nvidia.com/gpu: "1"} and works well for standard nvidia-device-plugin to schedule whole physical GPUs.

However, in shared GPU clusters with vGPU solutions (HAMi, vGPU-device-plugin, 4Paradigm vGPU, MIG extension), we need to declare GPU video memory quota as extended resource, e.g. nvidia.com/gpumem: "2048" (unit MB) to enforce hard VRAM isolation for each pipeline task.

There is no native accelerator API in KFP to attach extra GPU extended resources like nvidia.com/gpumem.