When HF_HUB_OFFLINE is set to a truthy value (1, true, yes, on),
download_model() should treat local_files_only=True to avoid any
network calls. Currently, even with the local-cache-first pass (which
may fail due to missing metadata), the retry loop still calls
download_files_from_huggingface() without local_files_only, which
triggers model_info() — a network API call that immediately fails in
offline mode. This causes an unnecessary fallback to GCS download from
storage.googleapis.com.
By setting local_files_only=True when HF_HUB_OFFLINE is enabled:
1. The HF local cache pass works if the model is cached
2. The retry loop skips the network-dependent HF path entirely
3. retrieve_model_gcs() only checks for local fast-* directories
4. No network calls are attempted at all
The truthy value check aligns with huggingface_hub's own parsing of
HF_HUB_OFFLINE, which accepts "1", "true", "yes", "on" (case-insensitive).
This is critical for air-gapped / restricted environments where both
HuggingFace and Google Cloud Storage are unreachable.
Made-with: Cursor
