Upgrade NVIDIA NIM Microservices#

That page details how to upgrade you NVIDIA NIM models. Refer to the Upgrading NVIDIA NIM Operator for details on upgrading the NIM Operator.

The NIM Operator performs Kubernetes rolling upgrades on NIM microservices. NIM availability is not impacted while updating NIMService and NIMPipelines to use new NIM cache instance if the number of replicas is greater than one.

To upgrade to new a NIM microservice version you must upgrade both your NIMCache and NIMService or NIMPipeline instance. The NIM Operator does not support updating cached models with new profile versions once cached. To upgrade your NIM microservices to use newer NIM model versions, you must create a new NIMCache that references the newer model, then update your corresponding NIMService or NIMPipeline instances to use the new NIM cache and NIM model version.

Before upgrading#

Refer to the NIM microservice release notes for details on what is different in your desired version of the NIM microservices Links to all supported NIMs is available in the NVIDIA NIM documentation.

Upgrade NIM models#

To upgrade your NIM model,

  1. Create a new NIMCache instance that is configured to use the updated NIM images.

    Refer to the Caching Models for details on configuring a NIMCache. You can re-use an existing PVC with the new NIMCache, by setting the spec.storage.pvc.create to false and the spec.storage.pvc.name field in the NIMCache custom resource to the existing PVC name.

  2. Update your corresponding NIMService or NIMPipeline instances to use the new NIMCache once it is in a Ready state.

For examples of the latest NIM Operator custom resource, refer to the NIM Operator repository samples folder.

Next Steps#

  1. Upgrading NVIDIA NIM Operator

  2. Upgrading NeMo Microservices

  3. Caching Models