Skip to content

README.md

Latest commit

 

History

History
80 lines (63 loc) · 3.82 KB

README.md

File metadata and controls

80 lines (63 loc) · 3.82 KB
Raw

Publishing the Body GGUF bundle

Prepared target: LocalAI-io/sam-3d-body-dinov3-GGUF. LocalAI-io is the Hugging Face namespace; it is not the GitHub organization name. No repository creation or upload is implied by adding these files.

The model card describes the Body pose branch, known limits, source revisions and component-specific licenses. artifacts.json pins the three approved F32/F32-I32 GGUFs by size and SHA-256. These are not low-bit quants; BF16 runs from the same archives. No Objects release is included. The card links base_model to Meta's original HF repository. No finetune or quantized relation is asserted for this F32 format conversion.

Local preflight (recommended first)

Python 3.11+ and Git suffice. No HF package, credentials or network is needed:

python3 scripts/publish_gguf.py \
  --backbone /path/to/body-dinov3-f32.gguf \
  --branch /path/to/body-pose-branch-f32.gguf \
  --mhr /path/to/mhr-lod1-f32.gguf

This streams the full file hashes using bounded memory and prints the exact repository, file names, sizes, SHA-256 values and source commit/dirty status. It does not scan directories, copy model files, deserialize checkpoints, or send anything to HF. Inputs must be trusted regular files, not symlinks, and remain immutable for the entire run. Hash/size mismatches fail closed; do not update the pins merely to make a mismatched model pass. A newly converted or quantized artifact requires independent conversion/parity review first.

Upload — a separate explicit action

First review LICENSING.md and all model license texts. Commit the release source and ensure the checkout/submodule is clean. Install the optional pinned HF tooling and authenticate without putting tokens in scripts:

uv sync --project reference/python --frozen --extra download
uv run --project reference/python --frozen --extra download hf auth login
uv run --project reference/python --frozen --extra download python scripts/publish_gguf.py \
  --backbone /path/to/body-dinov3-f32.gguf \
  --branch /path/to/body-pose-branch-f32.gguf \
  --mhr /path/to/mhr-lod1-f32.gguf \
  --upload --accept-model-licenses

This requires an existing accessible model repository. Add --create-repo only when authorized to create the public repository. --repo-id namespace/name selects another destination and updates the card's download command. Existing repository visibility is not changed. No remote files are deleted.

The uploader performs HF card validation and sends one commit containing only:

  • The three explicitly supplied, approved GGUF components.
  • The reviewed model card and component-specific NOTICE.
  • SAM, DINOv3, MHR, Momentum and original-documentation Apache license texts.
  • MANIFEST.json with source revisions and file identities.
  • SHA256SUMS covering all supplied files and the generated manifest (not itself).

The original source checkout must remain clean. The remote parent commit is checked to avoid silently overwriting a concurrent repository update. Failure after an explicitly authorized repository creation may leave an empty repository; the tool does not delete it automatically. Network errors may require checking the remote commit before retrying.

Source-repository installation packages retain LICENSE, NOTICE, docs/LICENSING.md and applicable third-party license texts. The model bundle's root LICENSE is deliberately SAM, not the source project's Apache license.

Verification

uv run --project reference/python --frozen python -m unittest discover -s tests -p 'test_publication.py' -v

Tests use synthetic tiny files and mock HF calls: no weights, downloads, remote repository creation or GPU required. Before an actual release, repeat a clean source build/test and model-specific numerical/visual acceptance as appropriate.