Prerequisites for using Multi-Instance GPU (MIG)
Review the prerequisites for using the Multi-Instance GPU on your Cloudera Embedded Container Service clusters.
The Cloudera Embedded Container Service users must:
- Install the NVIDIA driver on every MIG-capable node. Use a
version verified by NVIDIA and listed in the GPU Operator
Component Matrix under GPU operator v26.3.2,
which
cdp-privateCloudera Data Services on premises 1.5.5-SP4 installs: - Add the MIG capable nodes to the cluster.
Prerequisites to review for the Cloudera Embedded Container Service clusters
Before upgrading to Cloudera Data Services on premises 1.5.5 SP4 on
Cloudera Embedded Container Service, you must:
- uninstall the NVIDIA GPU Operator
- uninstall the NVIDIA NicClusterPolicy and Network
OperatorAdditionally, on every MIG capable node clean up the following directories:
rm -rf /etc/cdi/* rm -f /etc/nvidia-container-runtime/config.toml rm -rf /var/run/cdi/* rm -rf /usr/local/nvidia/toolkit
The cdp-private installer installs the GPU operator to Helm release
gpu-operator and Kubernetes namespace gpu-operator.
