| # Tensorizer torch_compat weights_only default bypass |
|
|
| This package reproduces a Tensorizer `torch_compat` issue where PyTorch 2.6+ |
| default safe loading blocks a malicious checkpoint, but the normal |
| `tensorizer_loading()` compatibility context executes the same checkpoint |
| without the caller passing `weights_only=False`. |
|
|
| Run from the repository root: |
|
|
| ```bash |
| reproduce.py |
| ``` |
|
|
| Expected marker: |
|
|
| ```text |
| MFV_TENSORIZER_WEIGHTS_ONLY_DEFAULT_BYPASS |
| ``` |
|
|
| The marker appears only in the explicit unsafe PyTorch baseline and the default |
| Tensorizer compatibility context. It does not appear for plain `torch.load()`, |
| plain `torch.load(..., weights_only=True)`, or |
| `tensorizer_loading()` with explicit `weights_only=True`. |
|
|
