Snowflake Gen AI Functions · Topic 25 of 26

Running Third-Party Models with Snowpark Container Services

Host a third-party model in Snowpark Container Services: set up the environment, Docker image, specification file, compute pool and image repository.

  • 4 min
In a nutshell
  • Order: image repository, push image, compute pool, spec file, service.
  • Compute pool needs MIN_NODES, MAX_NODES and INSTANCE_FAMILY; AUTO_SUSPEND_SECS default is 3600.
  • Image path in the spec: /database/schema/repository/image:tag.

What it is

Snowflake hosts many models for Cortex, but sometimes you need a model it does not host, such as an open model from Hugging Face or your own container. Snowpark Container Services (SPCS) runs containerized workloads inside Snowflake, with optional GPUs, and keeps the data in your account.

Four objects work together. You build an image, store it in an image repository, run it on a compute pool, as a service described by a specification file.

Dockerbuild imagepushImage repositoryregistrypulled byServicespec YAMLruns onCompute poolCPU or GPU nodesendpointApplicationSQL or HTTP

How it works

Environment setup

Before anything is created, an administrator prepares roles and the account.

NeedDetail
Image repositoryCREATE IMAGE REPOSITORY privilege on the schema
Compute poolCREATE COMPUTE POOL privilege on the account
ServiceA role that can create services in the schema and use the pool
Pull and push imagesREAD to list and download, WRITE to push, on the repository
NetworkAn external access integration if the container must reach the internet, for example to download model weights

Also decide the machine type. Large models need GPU instance families, and the image must be built for the platform the pool runs (the SPCS tutorials build for linux/amd64).

Docker images

Your container image holds the model server and its dependencies. The workflow:

  1. Build the image with Docker for the right platform.
  2. Log in to the Snowflake registry. The Snowflake CLI command is snow spcs image-registry login.
  3. Tag the image with the repository URL and push it.

The registry hostname is <orgname>-<acctname>.registry.snowflakecomputing.com, in lowercase, with underscores in the account name replaced by dashes. The repository URL is the hostname plus /<database>/<schema>/<repository>. SHOW IMAGE REPOSITORIES shows it, and SHOW IMAGES IN IMAGE REPOSITORY lists what you pushed.

Exam tip

The most common trap is the platform. An image built on an Apple silicon laptop for the wrong architecture will not run on an amd64 pool.

Limits: individual images cannot be dropped, since dropping the repository removes all images in it.

Image repository

CREATE IMAGE REPOSITORY IF NOT EXISTS ml_db.public.model_repo;
SHOW IMAGE REPOSITORIES IN SCHEMA ml_db.public;

The privileges are READ (list and download), WRITE (also push) and OWNERSHIP. A service pulls the image when it starts, so the role that creates the service must be able to read the repository.

Compute pool

CREATE COMPUTE POOL llm_pool
  MIN_NODES = 1
  MAX_NODES = 2
  INSTANCE_FAMILY = GPU_NV_S;
ParameterDetail
MIN_NODESRequired; must be above 0
MAX_NODESRequired
INSTANCE_FAMILYRequired; the machine type that sets compute and credit use. GPU families run models
AUTO_RESUMEDefault TRUE
AUTO_SUSPEND_SECSDefault 3600; 0 disables automatic suspension

The instance family name above is only an example. Run SHOW COMPUTE POOL INSTANCE FAMILIES to see the families available in your region. A compute pool bills while its nodes run, so suspend it when idle.

Specification file

The service specification is YAML under a top level spec.

spec:
  containers:
  - name: model-server
    image: /ml_db/public/model_repo/llm:1.0
    resources:
      requests: {nvidia.com/gpu: 1}
      limits: {nvidia.com/gpu: 1}
  endpoints:
  - name: api
    port: 8000
    public: false
  • containers is required: name and image for each container. command overrides the Dockerfile ENTRYPOINT and args overrides CMD. env sets variables.
  • The image path is /database/schema/repository/image:tag.
  • resources has requests and limits for cpu, memory and nvidia.com/gpu. A GPU request needs a matching limit.
  • endpoints expose a port. public defaults to false, which keeps the service private.
  • Optional sections: volumes, secrets and probes for readiness and liveness.

The service is then created with CREATE SERVICE ... IN COMPUTE POOL, pointing at the spec. Access to the endpoint is granted through service roles and USAGE.

Using the hosted model with Cortex

A text generation or vision language model served from SPCS can be called with AI_COMPLETE by passing the service name as the model, in the form <db>.<schema>.<service>. The docs mark this bring your own model feature as open preview. The guardrails option is not supported there. Text embedding models hosted this way can also be used with AI_EMBED (text only, also preview).

Common mistakes

Building for the wrong CPU architecture.

Forgetting MAX_NODES or INSTANCE_FAMILY. Compute pool needs both, and MIN_NODES above 0.

Requesting a GPU without a matching limit in the spec.

Leaving the pool running. It bills for as long as nodes are active.

Watch out

Do not describe SPCS bring your own model as general availability. For AI_COMPLETE and AI_EMBED it is open preview.

What to remember for the exam

  • Order: image repository, push image, compute pool, spec file, service.
  • Compute pool needs MIN_NODES, MAX_NODES and INSTANCE_FAMILY; AUTO_SUSPEND_SECS default is 3600.
  • Image path in the spec: /database/schema/repository/image:tag.
  • READ pulls, WRITE pushes; CREATE IMAGE REPOSITORY and CREATE COMPUTE POOL are separate privileges.
  • A hosted model is called with AI_COMPLETE by service name (open preview).

Check yourself

  • Which three parameters are required in CREATE COMPUTE POOL?

    MIN_NODES, MAX_NODES and INSTANCE_FAMILY.

  • What goes in a service specification file?

    A spec with containers (name, image, resources), endpoints, and optional volumes and secrets.

  • How do you call a model served by SPCS from AI_COMPLETE?

    Pass the service name as the model, in the form database.schema.service.

Topic 25 of 26 in Snowflake Gen AI Functions

Test yourself on Snowflake Gen AI Functions

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.