as title

Right now the process is to keep doing this until the torch.load eventually succeeds

torch.load(f, weights_only=True)
# Fails with Unsupported GLOBAL: {bla} ... and says to use `add_safe_globals` if it is safe
torch.add_safe_globals([bla])

We should see whether we can get all the GLOBALs in the checkpoint in one go in an efficient manner to make the process more low friction. Not sure whether this is possible without code execution though.

A possibility is that we could parse the pickletools.dis("data.pkl") for GLOBAL

Thanks @vmoens for suggesting this :)

Versions

main

cc @mruberry