Implement generative AI and agentic solutions 2.1 Build generative applications by using Foundry

A chat model whose runtime is chat-completions-compatible is deployed on managed compute…

Study guide skill: Deploy and consume LLMs, small models, code models, and multimodal models

Question B6-013

A chat model whose runtime is chat-completions-compatible is deployed on managed compute (preview) in a Foundry project. An existing app already calls other deployments with the OpenAI SDK on the project's /openai/v1/ route. Which two statements are correct? (Choose two.)

Select 2 answers.

  • A The app must call a separate virtual machine endpoint that the team provisions and patches
  • B The app can call the managed compute deployment with the OpenAI SDK on the standard /openai/v1/ route ✓ Correct answer
  • C The team must pick a virtual machine SKU before the deployment can serve requests
  • D The deployment uses the same project endpoint, authentication options, and private networking as other deployment types ✓ Correct answer
  • E The deployment can't use Microsoft Entra ID, so the app must switch to API keys

Answer: B, D

Managed compute deployments are served on the unified Foundry project endpoint with the same API keys, Microsoft Entra ID, and private networking as other deployment types, and chat-completions-compatible runtimes also work on the standard /openai/v1/ route with the OpenAI SDK. There are no virtual machines, clusters, or serving runtimes to own. Deployments are sized in model-centric terms, so you don't pick virtual machine SKUs.

Deep dive

How it works

Managed compute (preview) is built to look like every other Foundry deployment from the caller's side:

  • Same endpoint and auth. It uses the unified Foundry project endpoint, with the same API keys, Microsoft Entra ID and private networking as pay-per-token and provisioned throughput deployments.
  • Same SDK route. Chat-completions-compatible runtimes also work on the standard /openai/v1/ route with the OpenAI SDK.
  • No infrastructure. There are no virtual machines, clusters or serving runtimes to own.
  • No VM SKUs. Deployments are sized in model-centric terms. Foundry picks GPUs per instance based on model size, architecture, context length and whether the workload is optimized for latency or throughput.

For keyless authentication, Microsoft recommends Entra ID, with key-based authentication also supported.

The other options, and when you'd use them

  • Separate VM endpoint the team patches: Managed compute is reached through the same project endpoint, with no virtual machines to own.
  • Pick a VM SKU first: You don't pick VM SKUs. Foundry chooses GPUs per instance.
  • No Entra ID, switch to keys: Managed compute supports the same Entra ID and API key options as other deployment types.

Exam tip

Managed compute = same project endpoint, same auth, same private networking, /openai/v1/ for chat-completions runtimes, no VMs or SKUs to pick.

References

Last reviewed:

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →

All Generative AI & agents questions · AI-103 study guide

Thanks for your feedback!