load_safetensors

→dict of str to Tensor

load_safetensors(path: str, device: str = 'cpu')

source edit

Load a SafeTensors file into a flat state dict.

Reads the header and each tensor's bytes directly, builds a Lucid Tensor of the stored dtype — bfloat16 included — optionally relocates it to the requested device, and restores any zero-rank scalars that were promoted to shape (1,) during save_safetensors.

Parameters

pathstr

Path to a .safetensors file.

devicestr= 'cpu'

Target device for every loaded tensor: "cpu" (default) or "metal".

Returns

dict of str to Tensor

Flat state dict suitable for model.load_state_dict().

Raises

ValueError

If the file is not a well-formed SafeTensors file, or stores a dtype Lucid cannot represent (the 8-bit floats).

Notes

Unsigned integers have no Lucid dtype and widen losslessly: U8 to int16, U16 to int32, U32 to int64 (U64 to int64, refused past its range).

Examples

>>> import lucid
>>> sd = {"w": lucid.ones(3)}
>>> lucid.serialization.save_safetensors(sd, "/tmp/weights.safetensors")
>>> loaded = lucid.serialization.load_safetensors("/tmp/weights.safetensors")
>>> loaded["w"].tolist()
[1.0, 1.0, 1.0]