fix RestrictedUnpickler

[wkpark]

it does no harm to remove the pytorch_lightening dependency in the safe.py (torch.load)

see also comfyanonymous/ComfyUI@735ac4c

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[w-e-w]

it does no harm to remove the pytorch_lightening dependency in the safe.py (torch.load)

NO
we do use those to load LDSR upscaller models

[wkpark]

it does no harm to remove the pytorch_lightening dependency in the safe.py (torch.load)

NO we do use those to load LDSR upscaller models

did you tested? this is a torch.load() fix. without pytorch_lightening in the unpicker, LDSR ckpt model can be loaded without an issue.

[w-e-w]

okay you are right, it dose seems to load
it would seem my assumption about how pickle works is wrong

that being said, I still don't see a point of fixing something that isn't broke
unlike comfy is not like this change will allows to remove pytorch_lightning as it still used elsewhere

reopening this letting others decide

reopening this letting others decide

[wkpark]

okay you are right, it dose seems to load it would seem my assumption about how pickle works is wrong

that being said, I still don't see a point of fixing something that isn't broke unlike comfy is not like this change will allows to remove pytorch_lightning as it still used elsewhere

reopening this letting others decide

As you may have guessed, a number of my recent PRs have been related to optimising startup loading time.
some PRs are experimental (can be switched on/off), some PRs have been extracted from my recent Flux1 PR to reduce the total number of commits (not exactly related to the Flux parts)

the import time of pytorch_lightening is about ~5sec on my machine. (almost the same as imports torch)
so Im trying to reduce any dependency of pytorch_ligenting from core.

thank you for your concern!