#7043 fix(propeller): use configured GPU resource name instead of hardcoded nvidia.com/gpu Pull request opened by themavik Summary • Read

GpuResourceName

from the K8s plugin config instead of hardcoding

"<http://nvidia.com/gpu|nvidia.com/gpu>"

in

convertTaskResourcesToRequirements()

Problem

convertTaskResourcesToRequirements()

in

taskexec_context.go

uses the hardcoded constant

utils.ResourceNvidiaGPU

(

"<http://nvidia.com/gpu|nvidia.com/gpu>"

) for GPU resource requests and limits. The K8s plugin config already provides a configurable

GpuResourceName

field (via

gpu-resource-name

in the config), but this function ignores it. This means operators using alternative GPU resource types (e.g.

"<http://amd.com/gpu|amd.com/gpu>"

,

"<http://nvidia.com/gpu.shared|nvidia.com/gpu.shared>"

,

"<http://habana.ai/gaudi|habana.ai/gaudi>"

) cannot use the existing configuration to change the GPU resource name — it's always hardcoded to

<http://nvidia.com/gpu|nvidia.com/gpu>

. The config field is already used correctly in

flyteplugins/go/tasks/pluginmachinery/flytek8s/

(e.g.

container_helper.go

), but

flytepropeller

bypasses it. Fix Replace

utils.ResourceNvidiaGPU

with

k8sConfig.GetK8sPluginConfig().GpuResourceName

in

convertTaskResourcesToRequirements()

. This reads the configured GPU resource name at runtime, defaulting to

"<http://nvidia.com/gpu|nvidia.com/gpu>"

when not explicitly set. Testing • Default behavior unchanged:

GpuResourceName

defaults to

"<http://nvidia.com/gpu|nvidia.com/gpu>"

• Custom GPU resource names (e.g.

"<http://amd.com/gpu|amd.com/gpu>"

) are now respected in task resource requirements Closes #6746 flyteorg/flyte