Skip to content

v0.16.0

Choose a tag to compare

@github-actions github-actions released this 04 Sep 12:53
· 44 commits to main since this release
171d992
  • Model weights are served only from the cubert-gmbh Hugging Face mirrors (breaking for
    cached installs).
    Every ModelWeights registry entry now points at a public, ungated,
    commit-pinned mirror under cubert-gmbh/* (sam3, efficient-track-anything, dinov2,
    clip, adaclip) instead of the upstream repos, so download-model needs no Hugging Face
    account or token and the sandboxed child resolves the same cache folders the provisioner fills.
    The cache folder names change (models--cubert-gmbh--*), so an existing install re-downloads
    its weights once; there is no migration. New registry entries: efficienttam_s_512x512,
    efficienttam_ti_512x512, dinov2_vitb14_reg4, clip_vit_l_14_336, adaclip_all,
    adaclip_mvtec_colondb, adaclip_visa_clinicdb. Entries carry plugin and license, and
    aux_files is now a {filename: sha256} map whose companions are sha-verified too.
  • ModelWeights.resolve(name) and ModelWeights.materialize(name, dest_dir). Plugins
    resolve their weights through core instead of hardcoding an upstream repo id: resolve
    returns the cached path, downloads on a miss when online (download=None follows
    HF_HUB_OFFLINE), and raises ModelWeightsMissingError (a ModelDownloadError) naming the
    download-model command when offline; materialize hardlinks or copies the file to a fixed
    path for loaders that cannot read the Hugging Face cache layout. Neither writes to stdout.
    Plugins that call them need the hf extra (cuvis-ai-core[hf]).
  • download-model list --json emits the registry as machine-readable rows (name,
    plugin, repo_id, filename, revision, sha256, aux_files, license,
    requires_token, cache_dir_token, description) for consumers such as CuvisNEXT.
  • tools/mirror_weights.py (maintainer tool, not packaged) builds, uploads and audits the
    mirrors (plan, upload, check) and prints the registry entries to paste.
  • Removed the dead cuvis_ai_core.node.huggingface module (unreferenced, undeclared
    gradio_client / transformers dependencies, read HF_TOKEN directly).

What's Changed

  • Serve model weights from the cubert-gmbh Hugging Face mirrors (0.16.0) by @nghorbani in #73
  • Stamp the 0.16.0 changelog date by @nghorbani in #74

Full Changelog: v0.15.0...v0.16.0