Adam Tetelman

Assuming you are on NVIDIA GPUs, if you upgrade your Kubernetes cluster to use the latest version of the GPU Operator ([v1.11.0](https://github.com/NVIDIA/gpu-operator/releases/tag/v1.11.0)) we have added a feature that allows you...

I am encountering this same issue when I launch my custom defined `ServingRuntime` and `InferenceService` (https://github.com/supertetelman/nim-kserve/blob/main/nim-models/llama-2-7b_1-a100_24.01.yaml) running on `build version Kubeflow v1.8.0` When I deploy this InferenceService I need to...

> > In addition, current KServe using [emptyDir](https://github.com/kserve/kserve/blob/master/pkg/webhook/admission/pod/storage_initializer_injector.go#L318) as share volume between Storage initializer and main container, if we have option to configure a PVC to it (and storage initializer...

Having an annotation like that in the `InferenceService` would be perfect for me.

Do you have any error messages that you can fail from the failed install?

That is the opposite of `VolumeMount`, but it retains backwards compatibility and enforces a more secured default option (`ReadOnly`). I'd be very happy with that design.

In my case, there will be basic logic inside my inference application that creates a lockfile to prevent multiple pod replicas from writing at the same time. There will also...

That makes sense, but wouldn't that logic be covered in a different feature request specific to the initializer or by the person writing the custom storage initializer?