partition_gpu

Partition GPUs

Simple command line tool to partition the GPUs as specified in a GPU configuration file. The GPU configuration file specifies the desired partition size, and this tool will use nvidia-smi to create the maximum number of partitions on the desired size on the node.

To build GPU partitoner image

From root of the repository, run: docker build -f partition_gpu/Dockerfile .

To deploy GPU partitioner on all GPU nodes in GKE cluster

kubectl apply -f partition_gpu.yaml