RFE: move to native safetensors checkpoint format

[asterbini]

To avoid security issues with the pickled checkpoints in ckpt format, that allows the execution of unsafe code,
please move to the safetensors format, that does not.

[liuliu]

Thanks. But this is a wrong understanding actually. We don't use PyTorch nor Pickle format. The .ckpt file is in SQLite3 format for weight storage, it is as secure as safetensors if not more so. .ckpt suffix is an artifact from old deep learning days where we call any saved format for weight checkpointing as "checkpoint files" hence .ckpt suffix. That doesn't imply a particular on-disk format. It is just PyTorch got really popular and people start to make this wrong association. (In long term, we might want to rename it to .sqlite3 as suffix, however, that requires a lot of changes on our backend too given files stored on our cloud storage already in .ckpt suffix).

[asterbini]

Beautiful! sorry, I did'nt know.