Getting Started¶
This is a guide on getting started with GPT-in-a-Box 1.0 deployment on a Kubernetes Cluster. You can find the open source repository for the K8s version here.
Setup¶
Inference experiments are done on a single NKE Cluster with Kubernetes version 1.25.6-0. The NKE Cluster has 3 non-gpu worker nodes with 12 vCPUs and 16G memory and 120 GB Storage. The cluster includes at least 1 gpu worker node with 12 vCPUs and 40G memory, 120 GB Storage and 1 A100-40G GPU passthrough.
Note
Tested with python 3.10, a python virtual environment is preferred to managed dependencies.
Spec¶
Jump node:
OS: 22.04
Resources: 1 VM with 8CPUs, 16G memory and 300 GB storage
NKE:
NKE Version: 2.8
K8s version: 1.25.6-0
Resources: 3 cpu nodes with 12 vCPUs, 16G memory and 120 GB storage.
At least 1 gpu node with 12 vCPUs, 40G memory and 120 GB storage (1 A100-40G GPU passthrough)
NFS Server:
Resources: 3 FSVMs with 4 vCPUs, 12 GB memory and 1 TB storage
| Software Dependency Matrix(Installed) | |
|---|---|
| Istio | 1.17.2 |
| Knative serving | 1.10.1 |
| Cert manager(Jetstack) | 1.3.0 |
| Kserve | 0.11.1 |
Jump machine setup¶
All commands are executed inside the jump machine. Prerequisites are kubectl and helm. Both are required to orchestrate and set up necessary items in the NKE cluster.
Have a NFS mounted into your jump machine at a specific location. This mount location is required to be supplied as parameter to the execution scripts
Command to mount NFS to local folder
mount -t nfs <ip>:<share path> <NFS_LOCAL_MOUNT_LOCATION>
Follow the steps below to install the necessary prerequisites.
Download and set up KubeConfig¶
Download and set up KubeConfig by following the steps outlined in Downloading the Kubeconfig on the Nutanix Support Portal.
Configure Nvidia Driver in the cluster using helm commands¶
For NKE 2.8, run the following command as per the official documentaton:
helm repo add nvidia https://nvidia.github.io/gpu-operator && helm repo update
helm install --wait -n gpu-operator --create-namespace gpu-operator nvidia/gpu-operator --version=v23.3.1 --set toolkit.version=v1.13.1-centos7
For NKE 2.9, refer the official documentation for the validated config.
Download nutanix package and Install python libraries¶
Download the v0.1 release version from NAI-LLM-K8s Releases and untar the release. Set the working directory to the root folder containing the extracted release.
export WORK_DIR=absolute_path_to_empty_release_directory
mkdir $WORK_DIR
tar -xvf <release tar file> -C $WORK_DIR --strip-components=1
Kubeflow serving installation into the cluster¶
curl -s "https://raw.githubusercontent.com/kserve/kserve/v0.11.1/hack/quick_install.sh" | bash
Install pip3¶
sudo apt-get install python3-pip
Install required packages¶
pip install -r $WORK_DIR/llm/requirements.txt