Skip to main content

​
Hugging Face models

InferCrane uses huggingface_hub and hf_xet rather than a custom transfer protocol. Before provisioning, it resolves a repository and mutable reference such as main to an immutable commit when the provider exposes that evidence.
The resulting ModelArtifact can include repository identity, immutable revision, approximate size, and grounded compatibility/cache metadata. Missing provider evidence remains unavailable; InferCrane does not infer it.

​
Private and gated repositories

Register a reference to a fine-grained read token already injected into the control-plane environment. The API and database receive the reference, never the token value:
Pass the returned ID as model_secret_reference_id when creating the deployment. InferCrane resolves that credential only at the artifact boundary, verifies repository access, and persists the immutable revision plus the reference ID. It does not persist or return the token. Artifact access and runtime delivery are separate controls. RunPod should receive a provider-native secret name, Kubernetes should use a namespace-scoped Secret, and offline AWS/GCP workers should use an immutable pre-materialized model cache. A successful Hub resolution does not claim that an unconfigured runtime can download the repository.

Model artifacts

See the persisted identity and evidence model.