Vendored libraries

packaging

Source: https://github.com/pypa/packaging/

PyPI: https://pypi.org/project/packaging/

Vendored version: 23.2.0

Instructions to update:

• Copy the file packaging/version.py and all files that it is depending on
• Check if the licensing has changed from the BSD / Apache dual licensing and update the license files accordingly

quack

This is a subset of the full quack library, currently vendoring the following implementation paths:

• RMSNorm
• Lower-level GEMM epilogue implementation dependencies used by PyTorch-owned adapters

Note: There are a couple of patchsets applied to make the library vendorable - at a high level:

• Change exports from absolute quack.module to relative .module
• Rename cache directories / worker module paths so this copy is independent of any external quack package
• Remove custom-op registration from RMSNorm; PyTorch owns the public operator adapters

Source: https://github.com/Dao-AILab/quack

The pinned upstream commit is the PINNED_SHA constant in tools/vendoring/quack/vendor.sh (__version__ in the generated vendored package records the upstream version). That constant is the single source of truth; do not duplicate the pin here.

Instructions to update:

Edit PINNED_SHA in tools/vendoring/quack/vendor.sh to the new commit, then re-render (no SHA is passed):

tools/vendoring/quack/vendor.sh

# Or, to reuse an existing local clone instead of fetching:

tools/vendoring/quack/vendor.sh --src /path/to/local/quack

Instructions to update the subset of quack being vendored:

• In the vendor.sh script:
  • Update the files to be copied (FILES)
  • Update the rewrite_imports methods is there are more patterns required
• Add any extra patchsets needed in tools/vendoring/quack/patches