After the Kokoro account takeover, I want my local AI experiments to load the exact weights I tested, not whatever “main” points to today. Here’s how to pin a model to a commit.
1. Find the commit you trust
pip install -U huggingface_hub
python3 -c "from huggingface_hub import HfApi; i=HfApi().model_info('ORG/MODEL'); print(i.sha, i.last_modified)"
2. Download that exact revision
hf download ORG/MODEL --revision <sha> --local-dir ./models/MODEL
In Python, pass revision="<sha>" to from_pretrained() or hf_hub_download(). A branch name like main can move, but a full commit hash can’t.
3. Stop silent re-downloads
export HF_HUB_OFFLINE=1
With that set, your scripts use only what’s already cached locally and won’t fetch new revisions behind your back.
4. Record what you run
sha256sum ./models/MODEL/*.safetensors > MODEL.sha256
sha256sum -c MODEL.sha256
When you’re ready to upgrade, read the commit history first, test in a VM, and only then bump the pin.
Isolate it: try new software in a disposable VM or on a spare machine first, with a snapshot you can roll back to. Don’t point it at a machine you care about.
God Bless, and May the Source Be with You!