Sample script to deploy VPAIF-N DL workloads on a Deep Learning VM

Products

• VCF Private AI Services
• Private AI Foundation

Issue/Introduction

The attached zip bundle contains a collection of cloud-init scripts that deploying DL workloads on Deep Learning VM through the vSphere Client UI and VM Service.
The bundle includes a variety of sample scripts tailored for different DL frameworks and use cases.

Environment

 

• VMware Private AI Foundation 9.0

• VMware Private AI Foundation 9.1

 

Resolution

Each script is organised in its respective directory and contains two files: a cloud-init script and a config.json file.

• For each specific use case, the associated cloud-init script is provided as a cloud-config.yaml file located in its corresponding directory.

• To deploy the workload, encode the cloud-config.yaml cloud-init script in base64 format and assign the resulting value to the user-data OVF parameter of the deep learning VM image.

• Workload-specific configurable parameters are defined in the respective config.json files. After applying any required modifications, encode the file content in Base64 and assign it to the config-json OVF parameter.

Environment

• VMware Private AI Foundation 9.0
• VMware Private AI Foundation 9.1

VMware Private AI Foundation 9.0

DL Workload	Folder
NVIDIA RAG	./nvidia-rag/
CUDA Sample	./cuda-sample/
DCGM Exporter	./dcgm-exporter/
PyTorch	./pytorch/
TensorFlow	./tensorflow/
Triton Inference Server	./triton-inference-server/

 

For comprehensive guidance on deploying and configuring the Deep Learning VM and DL workloads, please refer to the VMware Private AI Foundation 9.0 documentation.

VMware Private AI Foundation 9.1

What's Changed in 9.1

• Deprecated workloads: The NVIDIA RAG (./nvidia-rag/) and TensorFlow (./tensorflow/) workloads have been removed and are no longer supported. 

• Container image selection: The cloud-init scripts now automatically select the appropriate container image based on whether NVIDIA AI Enterprise (NVAIE) registry credentials are provided:

  • If valid nvcr.io registry credentials (username + API key) are supplied, the scripts pull production branch images (e.g., nvidia/pytorch-pb25h1:25.03.03-py3).

  • If no credentials are provided, the scripts fall back to feature branch images (e.g., nvidia/pytorch:25.09-py3).

Supported DL Workloads (9.1)

DL Workload	Folder
CUDA Sample	./cuda-sample/
DCGM Exporter	./dcgm-exporter/
PyTorch	./pytorch/
Triton Inference Server	./triton-inference-server/

 

For comprehensive guidance on deploying and configuring the Deep Learning VM and DL workloads, please refer to the VMware Private AI Foundation 9.1 documentation