Support the "meta" device on load_file()

[lstein]

Feature request

Support the following syntax for obtaining an object containing the tensor shapes and other metadata without loading the actual tensors:

model = safetensors.torch.load_file(path, device="meta")

Motivation

This would be analogous to torch.load(path, map_location="meta") and allows the metadata to be interrogated without having to load the whole safetensors into memory.

Your contribution

Happy to test the implementation.

[Narsil]

Have you tried the following ?

with safe_open(filename, framework="pt") as f:
     print(f.keys())
     print(f.get_slice("mytensor").shape)
     print(f.metadata())

Should be pretty instant and it doesn't allocate anything

[keturn]

Is there a way to get data types from that? I don't see it exposed on PySafeSlice.

[Narsil]

Indeed it seems to be missing.