bootc-pytorch's Introduction

Bootc Image with PyTorch and GPU support

Resources

Prerequisites

RHSM subscription - org ID and activation key
A non-root user with sudo permissions
Optional for running a VM:
- Install KVM for running a VM
```
$ sudo dnf install -y qemu-kvm libvirt virt-install virt-viewer
$ sudo systemctl enable libvirtd --now
```
- Enable PCI pass-through to allow attaching a GPU to a VM - add intel_iommu=on to the host's boot arguments.
- Update KVM/libvirt permission.

Attaching a host PCI device to a VM

Attaching a PIC device to a running VM can be done either via the VM manager GUI (Add Hardware > PCI Host Device), or using the CLI:

Create an XML file, e.g. pci-device.xml:

<hostdev mode='subsystem' type='pci' managed='yes'>
<driver name='vfio'/>
  <source>
    <address domain='0x<domain>'
             bus='0x<bus>'
             slot='0x<slot>'
             function='0x<function>'/>
  </source>
</hostdev>

Attach the device to the VM:

$ sudo virsh attach-device <domain> pci-device.xml

Note: A host PCI device can be attached when creating a VM, but I never tried that:

Verifying

One of the ways to verify that the GPU is being used, after booting into a bootable image:

$ curl -O https://raw.githubusercontent.com/waggle-sensor/gpu-stress-test/main/stress.py
$ python stress.py

and observe nvidia-smi.

Triggering utilization of multiple GPUs (e.g. 0 through 2):

$ curl -O https://raw.githubusercontent.com/waggle-sensor/gpu-stress-test/main/stress.py
$ for i in $(seq 0 2); do (CUDA_VISIBLE_DEVICES=$i python stress.py &) ; done

Recommend Projects