Read named tensors from source files: path -> {name: array}, one reader per SourceFormat.
The onnx reader needs the onnx extra and the torch reader the torch extra; both are
imported only when called. The safetensors reader uses NumPy alone.
read_onnxfunctionThe initializers of a self-contained .onnx model, by name.read_safetensorsfunctionThe tensors of a .safetensors file, by name (little-endian; BF16 is not supported).read_torchfunctionThe tensors of a PyTorch state dict saved with torch.save, loaded with weightsonly=True.readfunctionThe tensors of the file at path, read as format.The initializers of a self-contained .onnx model, by name.
read_onnx(path: Path) -> dict[str, npt.NDArray]The initializers of a self-contained .onnx model, by name.
Weights held in Constant nodes rather than initializers are not read; such a source tensor is
reported as not in the file.
Raises
| Type | Description |
|---|---|
ValueError | If an initializer is stored in a separate data file, which the source's sha256 does not cover. |
The tensors of a .safetensors file, by name (little-endian; BF16 is not supported).
read_safetensors(path: Path) -> dict[str, npt.NDArray]The tensors of a .safetensors file, by name (little-endian; BF16 is not supported).
The tensors of a PyTorch state dict saved with torch.save, loaded with weightsonly=True.
read_torch(path: Path) -> dict[str, npt.NDArray]The tensors of a PyTorch state dict saved with torch.save, loaded with weights_only=True.
The tensors of the file at path, read as format.
read(format: SourceFormat, path: Path) -> dict[str, npt.NDArray]The tensors of the file at path, read as format.