[Feature Request] Add Aliases in Device spec

[yeahdongcn]

I'm using Nvidia Container Toolkit to provision GPU-capable containers for AI training. After I shifted to CDI, I could only use the index of the GPU devices rather than UUID. So I think it would be good if Aliases (or something else) could be introduced into specs.Device. As an end user, then I can specify both index and UUID. From Nvidia Container Toolkit perspective, the user experiences will become more consistent.

@@ -17,6 +17,7 @@ type Spec struct {
 // Device is a "Device" a container runtime can add to a container
 type Device struct {
        Name           string         `json:"name"`
+       Aliases        []string       `json:"aliases"`
        ContainerEdits ContainerEdits `json:"containerEdits"`
 }

[elezar]

@yeahdongcn aliases were part of the original proposal but removed to simplify the API once we started actively developing this. It would defintely be worth including again.

As a matter of interest, are you working in an environment where you need both simultaneously, or would more flexibility in generating the spec be sufficient to cover your use cases?

[yeahdongcn]

As a matter of interest, are you working in an environment where you need both simultaneously, or would more flexibility in generating the spec be sufficient to cover your use cases?

I'm still in the stage of migrating from the previous version of Nvidia Container Toolkit to the latest one. nvidia-ctk info generate-cdi is configured to be invoked at system startup as the physical GPU card/slot may get changed during my test.

For validation training, I prefer to use the index, while for performance tuning, I choose to use UUID (looks more accurate to me). This can be done using --gpu previously, but now I need 2 versions of nvidia-ctk and regenerate the CDI config file if I switch between the 2 kinds of test.

[elezar]

@yeahdongcn I was just thinking about this and realized that if you generate two specs with nvidia-ctk cdi generate (the final CLI as of the v1.12.0 relase) then both device names would be available.

sudo nvidia-ctk cdi generate --device-name-strategy=uuid --output /etc/cdi/nvidia-uuid.yaml
sudo nvidia-ctk cdi generate --device-name-strategy=index --output /etc/cdi/nvidia-index.yaml

(the latter being the default)

When checking the specifications you would apear to have double the number of devices, but since CDI is not intended to restrict or count resources in this way, this should not matter. There will also be duplicate nvidia.com/gpu=all devices but this should be handled by the CDI package's conflict resolution.

[yeahdongcn]

@yeahdongcn I was just thinking about this and realized that if you generate two specs with nvidia-ctk cdi generate (the final CLI as of the v1.12.0 relase) then both device names would be available.

sudo nvidia-ctk cdi generate --device-name-strategy=uuid --output /etc/cdi/nvidia-uuid.yaml
sudo nvidia-ctk cdi generate --device-name-strategy=index --output /etc/cdi/nvidia-index.yaml

(the latter being the default)

When checking the specifications you would apear to have double the number of devices, but since CDI is not intended to restrict or count resources in this way, this should not matter. There will also be duplicate nvidia.com/gpu=all devices but this should be handled by the CDI package's conflict resolution.

@elezar Thank you for letting me know about the new usage. Do you mean after the CDI spec generation, /etc/cdi/nvidia-uuid.yaml and /etc/cdi/nvidia-index.yaml will both be picked by Nvidia Container Toolkit? I'll give it a try later.

[elezar]

Any CDI client (consumer) such as podman, crio, containerd, or the nvidia-container-runtime in CDI mode loads all spec files to determine what valid CDI devices exist. Any of these will then have loaded both of the specs and see both device names as valid.

[elezar]

@yeahdongcn I have just done a quick test myself, and the duplicate all devices in the two specs generated by the commands above will cause issues when injecting devices. This means that this should be removed from at least one of them.

We will update our tooling to make this easier.

[github-actions]

This issue is stale because it has been open 90 days with no activity. This issue will be closed in 30 days unless new comments are made or the stale label is removed.

[github-actions]

This issue was automatically closed due to inactivity.