This message was deleted.
# harvester
a
This message was deleted.
s
Afaik not possible, at least via terraform per https://github.com/rancher/terraform-provider-rancher2/issues/1030. If you do find a way, I'd love to know!
b
most likely we can automate this by updating each VM manifest (which is of type kubevirt.io.virtualmachine) following https://kubevirt.io/user-guide/virtual_machines/host-devices/#preparation-of-mediated-devices-such-as-vgpu
to fully automate this is it possible to create a Harvester Addon that will run a CronJob to assign the right PCI device to each VM
w
afaik full vGPU's are not available yet in harvester right? We currently need to fully assign a PCI GPU device to an individual VM.
Would be neat if we could have an admission controller modify the vm spec before it gets created. This way the VM would boot up immediatly with the correct gpu attached to it
👀 1
m
https://github.com/inaccel/cloud-init/blob/master/internal/mutate.go Our workaround is to mutate VMIs based on the cloud config in user data. For example, you add the following cloud config in the user data of a node pool and the webhook adds the respective host devices in the VMI spec:
Copy code
inaccel:
  <http://nvidia.com/GP102GL_Tesla_P40|nvidia.com/GP102GL_Tesla_P40>: 1
There's also a controller that detects user data changes in a VM spec.
We might consider packaging this functionality as a Harvester addon. Currently, that's how we deploy it in Harvester: https://github.com/inaccel/helm/tree/master/charts/fpga-cloud
🙌 2
g
we do have work in progress to allow vgpu to be added a resource during VM creation via rancher
❤️ 1
s
💯 awesomeness!