Prefetch Hugging Face models

flyte.prefetch.hf_model() downloads a model from the Hugging Face Hub into your own object storage and registers the result as a model artifact. It is the third way an artifact is created, alongside task outputs and publishing your own.

        
import flyte
import flyte.prefetch

flyte.init_from_config()

run = flyte.prefetch.hf_model(repo="HuggingFaceTB/SmolLM2-135M-Instruct")
run.wait()

For the full how-to, including sharding for multi-GPU inference, resources, tokens for gated repos, CLI usage, and serving the result from a vLLM or SGLang app, see Prefetching models. This page covers what a prefetch means for the artifact registry.

What gets registered

On success the platform records a model artifact for the stored weights:

The artifact name defaults to the last segment of the repo id, with . replaced by -, or set it with artifact_name. The example above registers SmolLM2-135M-Instruct.
The version is the Hugging Face commit ID.
The searchable metadata carries the model facts (framework, architecture, task, modality, serialization format) plus the source repo and commit.
The repo’s README, if it has one, is attached as the model card.

Because the version is the upstream commit, a prefetch is idempotent. Re-running it for a model that has not moved republishes the same version instead of filling the registry with duplicates, and a genuine upstream change arrives as a new version you can trigger on.

Finding a prefetched model

Retrieve it by name:

        
from flyte.remote import Artifact

model = Artifact.get("SmolLM2-135M-Instruct")

Or find it by where it came from. The source repo and commit are recorded as searchable metadata:

        
    
flyte get artifact --attr source_repo=HuggingFaceTB/SmolLM2-135M-Instruct

Those source_repo and source_commit attributes are what make a prefetched model traceable back to its Hub repo and commit.

Once prefetched, mount the model into a serving app with an artifact parameter. See Use artifacts in apps.