CircleCI: disable hugginface xet - #3316

Motivation

Currently CircleCI fails because of hanging connections with HuggingFace using xet.

Loading pretrained weights from Hugging Face hub (timm/resnet50_gn.a1h_in1k)
HTTP Request: GET https://huggingface.co/api/models/timm/resnet50_gn.a1h_in1k/xet-read-token/b5b1dcfc1a526af275ec440b8d697c66c90a98b1 "HTTP/1.1 200 OK"
model.safetensors:   0% 70.5k/102M [00:01<24:59, 68.1kB/s]
Too long with no output (exceeded 10m0s): context deadline exceeded

Xet is Hugging Face’s new storage backend designed to replace the aging Git LFS (Large File Storage). While Git LFS treats every file as a single "blob," Xet breaks files into small "chunks."

In restricted CI environments (like CircleCI, GCP, or Kaggle), the Xet "chunked" download often stalls. Because Xet opens many simultaneous connections to different data chunks, CI firewalls or NAT gateways sometimes flag this as suspicious activity or simply drop the connections when they hit a certain threshold.

Changes

This PR disables Xet:

environment:
    HF_HUB_DISABLE_XET: 1

and sets caching for the huggingface weights, to avoid repeated downloading. This also improves stability in case of a bad network connection.

- restore_cache:
    name: Restore Hugging Face cache
    keys:
      - hf-weights-v1-{{ checksum "pyproject.toml" }}
      - hf-weights-v1-

More information

• HuggingFace Xet: https://huggingface.co/join/xet
• Similar issue: Xet-backed downloads stall on GCP (works when HF_HUB_DISABLE_XET=1) huggingface/xet-core#800