Snowflake for Gen AI · Topic 6 of 13

Bringing Your Own Models: Model Registry and Snowpark Container Services

Bring your own models into Snowflake with the Model Registry and Snowpark Container Services: when to choose each, how models run and the privileges.

  • 4 min
In a nutshell
  • Model Registry: log_model, versions, aliases, schema level object, CREATE MODEL.
  • Custom model: you supply all dependencies.
  • target_platforms: WAREHOUSE, SNOWPARK_CONTAINER_SERVICES or both.

What it is

Cortex gives you hosted models, but sometimes the model you need is your own: a custom trained model, an open source model or a specialized vision model. Snowflake lets you bring it in two main ways. The Snowflake Model Registry stores and versions the model as a schema level object. Snowpark Container Services (SPCS) runs it inside containers on compute you control, often with GPUs.

The two work together. You log a model in the registry, then deploy it to a service on SPCS when it needs a GPU or a custom runtime.

Your modeltrained or open sourcelog_modelModel Registryversioned objectdeployWarehouse or SPCSwhere it runsserveSQL and apps

How it works

Snowflake Model Registry

Adding a model to the registry is called logging it, done with log_model from Snowpark ML. Each call creates a version, and the name and version pair must be unique in the schema. Models are schema level objects, so normal grants apply.

  • Custom model: when no built in type fits, you wrap your own code and supply every dependency yourself. log_model fails if a custom model cannot run in a warehouse when that platform is requested.
  • Arguments: common ones are conda_dependencies, pip_requirements, sample_input_data or signatures, metrics and target_platforms.
  • target_platforms: WAREHOUSE, SNOWPARK_CONTAINER_SERVICES or both. Models that need a GPU go to SPCS.
  • Versions: a model can have many versions, with a default version and aliases such as DEFAULT, FIRST and LAST.
PrivilegeAllows
CREATE MODEL on the schema (or owning it)Log a model
USAGE on the modelRun inference in a warehouse
READ on the modelSPCS inference and viewing metadata

Warehouse deployments have a size ceiling (the docs state 15 GB total model size), which is one reason large models go to SPCS.

Snowpark Container Services

SPCS is a managed container platform inside Snowflake. You package the application as an OCI image and Snowflake runs it.

ObjectRole
Image repositoryStores your images (CREATE IMAGE REPOSITORY)
Compute poolVM nodes that run services; some machine types have GPUs (CREATE COMPUTE POOL)
Service specificationYAML that defines the containers
ServiceLong running workload started with CREATE SERVICE; restarted if a container exits
Job serviceFinite run started with EXECUTE JOB SERVICE; not restarted

Because you control the whole container, you can bring custom runtimes and libraries. Typical workloads are model serving and GPU training. Network policies control ingress and external access integrations control egress.

For online inference the registry can create the service for you with create_service. Batch inference can run through run_batch.

Model Registrylogged modelcreate_serviceCompute poolGPU nodesrunsInference serviceendpointcalled byAI_COMPLETE or app

Using a hosted model from Cortex functions

AI_COMPLETE can use a text generation or vision language model deployed as an SPCS inference service by passing the service name as the model argument. The documentation labels this open preview. Such models have their own access requirements, separate from SNOWFLAKE.CORTEX_USER.

Exam tip

If the scenario needs Cortex style SQL calls on a model Snowflake does not host, the answer is a model on SPCS, called through the model argument of AI_COMPLETE.

Choosing

NeedChoose
Versioning, governance and sharing of a modelModel Registry
GPU, custom runtime, long running serviceSnowpark Container Services
Hosted model with no opsCortex AI functions
Watch out

Registry models do not run themselves on GPUs. A GPU needs an SPCS compute pool with GPU capable machines.

Common mistakes

Using SPCS to store model versions. Version control belongs in the Model Registry.

Logging a warehouse only model that needs a GPU. Set target_platforms to include SPCS.

Forgetting READ. SPCS inference needs READ on the model, not only USAGE.

Treating a service like a job. Services restart on exit; job services finish.

Expecting Cortex roles to cover your own model. SPCS hosted models have their own access requirements.

What to remember for the exam

  • Model Registry: log_model, versions, aliases, schema level object, CREATE MODEL.
  • Custom model: you supply all dependencies.
  • target_platforms: WAREHOUSE, SNOWPARK_CONTAINER_SERVICES or both.
  • SPCS: image repository, compute pool, service spec, CREATE SERVICE.
  • GPU and custom runtimes need SPCS.

Check yourself

  • What is logging a model?

    Adding it to the Model Registry as a new version of a schema level object.

  • Which privilege lets a role use a model on SPCS?

    READ on the model.

  • When must a model go to SPCS rather than a warehouse?

    When it needs a GPU, a custom runtime or exceeds warehouse limits.

Topic 6 of 13 in Snowflake for Gen AI

Test yourself on Snowflake for Gen AI

The full Snowflake GES-C02 question bank has 585 questions, with timed practice exams and an explanation for every answer.

This study guide is independent and is not affiliated with, endorsed by, or sponsored by Snowflake. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.