Typed Hub errors. Functions still return anyhow::Result at the boundary
so callers keep one error type, but raise these variants so messages stay
consistent and matchable via .to_string().
Checkpoint file extensions this tool knows how to read, in the order
preferred when a repo ships more than one format and the caller doesn’t
request a specific one.
Canonical HuggingFace repo per friendly model name.
Single source of truth for CLI --model defaults and Python convert()
dispatch, so the two cannot drift when defaults change.
Where the canonical, always-F32 GGUF for repo_id lives once converted once. Every
per-model convert command checks this path first: if present, it recasts straight
from this cached GGUF to whatever dtype was requested instead of re-downloading and
re-converting from HuggingFace.
Delete a model’s cached files (canonical GGUF + config) and optionally an
output GGUF. Shared by the zsfm CLI and the Python bindings so both
report identical messages on stdout/stderr.
Download a repo’s checkpoint in whatever format it’s stored in (or a
specific one, if format is given). Resolves HF’s standard sharding
indices for safetensors and bin; for every other format, if more than
one file matches, file_filter (a case-insensitive substring) is required
to pick which ones — this avoids silently pulling every alternative
quantization/variant of a multi-GB checkpoint.
Download (or locate from cache) a single arbitrary file from repo_id, resuming a partial
download if one exists. For models that don’t fit the config.json + model.safetensors
shape — e.g. Lag-Llama’s raw PyTorch Lightning .ckpt checkpoint.