Skip to main content

Scenario: Hugging Face Model Ingestion

This guide covers the end-to-end workflow for using Anvil's built-in Hugging Face integration to ingest and store ML models.

Key Concept: Ingestion is an asynchronous background job. You initiate the job with the CLI, and Anvil's backend workers handle the download and storage process. You can then use the CLI to monitor the job's status.

1. Admin: Grant Hugging Face Permissions​

First, an administrator must grant the client's app the necessary permissions to manage Hugging Face keys and initiate ingestion jobs.

# Grant permission to manage HF keys and start ingestions
admin policy grant --app-name data-science-app --action hf_key:create --resource "*"
admin policy grant --app-name data-science-app --action hf_key:delete --resource "*"
admin policy grant --app-name data-science-app --action hf_key:read --resource "*"
admin policy grant --app-name data-science-app --action hf_ingestion:create --resource "*"

2. Client: Manage Hugging Face API Keys​

The client needs to add their Hugging Face API token to Anvil. This token is stored securely and associated with a friendly name.

# Add a personal HF token to Anvil and give it a memorable name
anvil hf key add --name my-hf-key --token hf_abc...xyz --note "Personal Read Token"

# List available HF keys to confirm it was added
anvil hf key ls

To remove the key later:

anvil hf key rm --name my-hf-key

3. Client: Prepare Destination and Start Ingestion​

Before starting the ingestion, the client must ensure a destination bucket exists and that the administrator has granted object:write permissions for it.

# admin policy grant --app-name data-science-app --action object:write --resource "models/*"

The ingest start command will return a unique ID for the ingestion job.

Expected Output:

ingestion id: ingest_12345...

4. Client: Monitor the Ingestion Job​

The client uses the job ID to check the status of the asynchronous ingestion process.

# Check the status of the background job
anvil hf ingest status --id ingest_12345...

Expected Output:

state=Completed queued=0 downloading=0 stored=14 failed=0 error=

If necessary, the job can be cancelled before it completes:

anvil hf ingest cancel --id ingest_12345...

Once complete, the model files will be available as objects in the models bucket.