Model artifacts
InferCrane resolves a Hugging Face repository and requested revision throughhuggingface_hub.HfApi.model_info. The resulting commit SHA is attached once to the immutable
deployment revision and passed to vLLM with --revision. Retries reuse the persisted artifact;
they do not resolve a mutable branch again.
Persisted evidence includes the repository, requested reference, immutable commit, canonical model
identity, approximate Hub storage when returned, and grounded library/pipeline metadata. Cache state
is unknown unless the execution backend can measure the worker cache. The control-plane host cache
is never presented as worker cache evidence.
The production image installs huggingface_hub and hf_xet in an isolated virtual environment so
their dependency versions cannot alter SkyPilot’s environment. Private or gated repositories use
the standard Hugging Face token environment/configuration; InferCrane does not implement a model
download protocol.