This message was deleted.

Or do I install these in the VM?

what are you trying to do? harvester itself doesn’t really need graphics

I’m trying to run LLM workloads in Harvester (vllm etc)

I would use PCI passthrough and a VM :

The problem is that the vm doesn’t start at all when I enable pci passthrough for the AMD pci devices.

reboot the node. I have seen this before where I needed to disable PCI passthrough, reboot, and re-enable.

Thanks, I will try that.

you can also do a kubectl describe of the VM (pod) to see why it won’t schedule

@wooden-area-49191 make sure all devices in the group are passed. See the PCI device ID. ie some graphic cards also have a sound device, and that also needs to be passed along the GPU.

I have a AMD MI300X and the GPU:s are visible as processing accelerators. When I create a vm and enable passthrough to these I get an error saying the condition isn’t met. After some investigation I managed to add ‘modprobe amdgpu’ to the 90_custom.yaml and restarted the node- after that the vm started but stopped before reaching the shell with an error about

virtqemud

Perhaps @great-bear-19718 can comment. I don’t think you need the driver locally at all anyway since you are doing a pass through.

Hm - this is the error I receive now in the log when I’m trying again to start it up

{"component":"virt-launcher","level":"warning","msg":"MDEV_PCI_RESOURCE_AMD_COM_AQUA_VANJARAM_INSTINCT_MI300X not set for resource <http://amd.com/AQUA_VANJARAM_INSTINCT_MI300X%22,%22pos%22:%22addresspool.go:51%22,%22timestamp%22:%222025-02-10T09:56:18.435280Z%22}|amd.com/AQUA_VANJARAM_INSTINCT_MI300X","pos":"addresspool.go:51","timestamp":"2025-02-10T09:56:18.435280Z"}>
{"component":"virt-launcher","level":"warning","msg":"USB_RESOURCE_AMD_COM_AQUA_VANJARAM_INSTINCT_MI300X not set for resource <http://amd.com/AQUA_VANJARAM_INSTINCT_MI300X%22,%22pos%22:%22addresspool.go:51%22,%22timestamp%22:%222025-02-10T09:56:18.435369Z%22}|amd.com/AQUA_VANJARAM_INSTINCT_MI300X","pos":"addresspool.go:51","timestamp":"2025-02-10T09:56:18.435369Z"}>
{"component":"virt-launcher","level":"info","msg":"host-devices created: [0000:65:00.0, 0000:85:00.0]","pos":"hostdev.go:98","timestamp":"2025-02-10T09:56:18.435468Z"}
{"component":"virt-launcher","kind":"","level":"info","msg":"Synced vmi","name":"vllm","namespace":"berget","pos":"server.go:208","timestamp":"2025-02-10T09:56:18.437905Z","uid":"bb30a93c-5b72-4ca4-9abd-de74f2e693f2"}
{"component":"virt-launcher","level":"info","msg":"Process berget_vllm and pid 77 is a zombie, sending SIGCHLD to pid 1 to reap process","pos":"monitor.go:186","timestamp":"2025-02-10T09:56:19.387458Z"}
{"component":"virt-launcher","level":"info","msg":"Waiting on final notifications to be sent to virt-handler.","pos":"virt-launcher.go:281","timestamp":"2025-02-10T09:56:19.387537Z"}

is there any documentation for k8s? In the video I posted about nvidia I didn’t have to install anything on the harvester node.

Here is the AMD operator that I’m trying to install. it needs a kernel driver installed in the operating system unfortunately: https://github.com/ROCm/gpu-operator

i do not have an amd gpu.. but eventually you just need the amd drivers installed. the current addon is written to handle the nvidia driver only