A container image packages the software environment used to create a container. In GPU training setup, choosing an image is a separate decision from choosing the VM machine type: the image supplies the operating-system, Python, framework, and accelerator foundations. section 5.1.1
Use an image to:
- Start from a prebuilt deep-learning environment instead of configuring every dependency yourself. section 5.1.1
- Choose among Base, TensorFlow, and PyTorch images with compatible CUDA and Python versions. section 5.1.1
- Customize a prebuilt image when the standard environment is only a starting point. section 5.1.1
A GPU-ready starting point
The image common-gpu-v20230925-debian-11-py310 provides a Debian 11 and Python 3.10 environment prepared for GPU use. It can be selected while creating a VM with a compatible machine type such as g2-standard-24. section 5.1.1
$ gcloud compute images list --project deeplearning-platform-release
--format=”value(NAME)” --no-standard-images
List available deep-learning images before choosing a framework, CUDA, and Python combination. section 5.1.1
$ INSTANCE_NAME="g2-llminprod-example"
$ gcloud compute instances create ${INSTANCE_NAME} --zone=us-west1-a
--machine-type=g2-standard-24 --image-project=deeplearning-platform-release
--image=common-gpu-v20230925-debian-11-py310 --boot-disk-size=200GB --scopes
cloud-platform --metadata=install-unattended-upgrades=False,install-nvidia-
driver=True --maintenance-policy TERMINATE --restart-on-failure
See the image selected directly in a gcloud VM-creation command. section 5.1.1
