Why is torch so much faster than other formats?

[mar-muel]

Hello there!

I'm very impressed by the load speed for torch tensors, however I'm experiencing much slower loading for flax/numpy.

E.g. to load a 3.3 GB model artefact with safetensors.torch.load_file takes 0.08s (!) whereas it's on the order of 3-4s for safetensors.numpy.load_file or safetensors.flax.load_file.

Am I missing something or are these just technical limitations for other other frameworks at the moment?

Thanks for any insight 🙏