Deploy and Manage Weights & Biases on Azure

Azure.ResourceManager.WeightsAndBiases is the Azure Resource Manager SDK for deploying and managing Weights & Biases ML experiment-tracking instances via the Azure Marketplace, currently a preview package (v1.0.0-beta.1, API version 2024-09-18-preview). It authenticates through ArmClient with DefaultAzureCredential. Resources nest as subscription > resource group > WeightsAndBiasesInstance, whose properties hold marketplace offer/plan details, the admin user, W&B-specific partner properties (region and subdomain), and optional Entra ID single sign-on configuration, plus an optional system-assigned managed identity.

Creating an instance sets the marketplace subscription and offer (publisher wandb, offer wandb-pay-as-you-go, a monthly wandb-payg plan), the admin user's name/email, and W&B-specific settings - a supported region (East US, Central US, West US, West Europe, Japan East, or Korea Central) and a subdomain for the deployed instance. Instances can be read and listed both at the resource-group and subscription level. Single sign-on is configured separately by updating the instance with SAML SSO details - an enable/disable state, the Entra enterprise application ID, and the allowed AAD domains. Updates otherwise go through a patch object (for example, adding tags), and instances can be deleted outright. The SDK has no first-class name-availability check; the documented workaround is to attempt a Get and treat a 404 as "name available."

Use it when provisioning or managing the W&B Azure Marketplace deployment itself - the instance, its SSO configuration, and its lifecycle - as infrastructure-as-code, not for the actual ML experiment tracking, which happens through the separate W&B Python SDK once the instance exists. Being a preview package, expect the API surface to still change; check the provisioning state reaches Succeeded before treating an instance as ready, per the skill's own best practices.

Input is the instance configuration (marketplace offer details, admin user, region/subdomain, optional SSO and managed identity settings). Output is the created/updated WeightsAndBiasesInstanceResource with its provisioning state and connection details. Install it with:

dotnet add package Azure.ResourceManager.WeightsAndBiases --prerelease
dotnet add package Azure.Identity

Once the instance exists, experiment tracking itself goes through the W&B Python SDK: wandb.login(host="https://<subdomain>.wandb.ai"), wandb.init(project=...), wandb.log({...}), and run.finish(). Its own best practices: use DefaultAzureCredential for flexible auth, enable managed identity for secure access to other Azure resources, configure SSO for enterprise security, tag resources for cost tracking, wait for Succeeded provisioning state before use, pick a region close to your compute, and monitor the resource with Azure Monitor. Errors surface as RequestFailedException - 409 for a name conflict, 400 for invalid configuration.

It provisions against the Azure Marketplace wandb/wandb-pay-as-you-go offer and Entra ID for SSO, and hands off to the separate W&B Python SDK for actual experiment tracking; a related package, Azure.ResourceManager.MachineLearning, manages Azure ML workspaces instead.

.NET platform engineers provisioning and managing Weights & Biases Marketplace deployments - instance lifecycle, SSO, and tagging - as infrastructure automation, distinct from ML engineers using the W&B SDK for day-to-day experiment tracking.