acoustic-carpenter-78188
04/18/2026, 7:47 PM"<http://nvidia.com/gpu|nvidia.com/gpu>" when building Kubernetes ResourceList entries from Flyte GPU resource requirements. This hardcoding took precedence over the GpuResourceName field already configurable in K8sPluginConfig (via the gpu-resource-name Helm/propeller config value), making it impossible to use alternative GPU resource types such as <http://amd.com/gpu|amd.com/gpu> or <http://intel.com/gpu|intel.com/gpu>.
The affected paths were:
• flyteplugins/go/tasks/pluginmachinery/flytek8s/utils.go — ToK8sResourceList used the package-level ResourceNvidiaGPU constant.
• flytepropeller/pkg/controller/nodes/task/taskexec_context.go — convertTaskResourcesToRequirements used utils.ResourceNvidiaGPU.
Note that SanitizeGPUResourceRequirements and ApplyResourceOverrides in container_helper.go already correctly read from config.GetK8sPluginConfig().GpuResourceName. This PR brings the two remaining call sites into alignment.
What changes were proposed in this pull request?
• flyteplugins/.../flytek8s/utils.go: Import flytek8s/config and replace ResourceNvidiaGPU with config.GetK8sPluginConfig().GpuResourceName in ToK8sResourceList.
• flytepropeller/.../taskexec_context.go: Import flytek8s/config and replace utils.ResourceNvidiaGPU with flytek8sConfig.GetK8sPluginConfig().GpuResourceName in convertTaskResourcesToRequirements. Remove now-unused flytepropeller/pkg/utils import.
• Tests: Update both test files to assert against config.GetK8sPluginConfig().GpuResourceName rather than the hardcoded constant, confirming the behavior is config-driven.
How was this patch tested?
• go test ./go/tasks/pluginmachinery/flytek8s/... in flyteplugins/ — passes.
• go test ./pkg/controller/nodes/task/... in flytepropeller/ — passes.
Labels
• fixed
• I updated the documentation accordingly.
• All new and existing tests passed.
• All commits are signed-off.
flyteorg/flyteacoustic-carpenter-78188
05/16/2026, 1:29 AM