modeldiffr fingerprint¶
Print content hashes for one model without running an audit.
Use it to pin a model in a model card, to check that two uploads are the same weights, or to see which tensors a fine tune touched.
model ./checkpoints/step_4000
weights root sha256:194b404d99de3d507664c452555a641147f697491e14ee29fd2632ac8089b8b9
tensors 21 (188,800 bytes, safetensors)
vocab sha256:395b8af95f1f4d7b0d9091287c0b2073e24f2860bf71f8e77654a904c854c306 (415 tokens)
tokenizer.json sha256:db3150f82f955f62d7a95aa1768b5366cb4f83b0588ca232d35a363772f8456f
chat template sha256:85ed6a3dbe40efa25c2d790b07f9739515f9ef46126d92db2c3c886c42fa169f
Arguments¶
model¶
string, required
A Hugging Face model id or a local directory. Hub models are downloaded once into the normal Hugging Face cache.
Options¶
--revision¶
string
Branch, tag or commit for Hub models. The resolved commit is printed.
--leaves¶
flag
Also print the hash of every tensor, one per line, sorted by name. Diff two of these outputs to see which tensors changed.
--json¶
flag
Print everything as JSON, in the same shape as fingerprints.base in a report.
Supported formats¶
| Format | Supported |
|---|---|
| safetensors, single file or sharded | Yes, read directly, every dtype |
PyTorch .bin, .pt, .pth |
Yes, needs torch; gives the same root as the same weights in safetensors |
| GGUF | Not yet, arrives with the llama.cpp backend in Phase 2 |
The tokenizer hashes need transformers. Without it, only the weights are fingerprinted.