[Help needed] How to pass `weights_only` param to `tuner`

[phoeenniixx]

Description & Motivation

Copied from ar.ipynb tutorial of pytorch-forecasting.

# find optimal learning rate
from lightning.pytorch.tuner import Tuner

res = Tuner(trainer).lr_find(
    net, train_dataloaders=train_dataloader, val_dataloaders=val_dataloader, min_lr=1e-5
)
print(f"suggested learning rate: {res.suggestion()}")
fig = res.plot(show=True, suggest=True)
fig.show()
net.hparams.learning_rate = res.suggestion()

Here the trainer is fitted inside lr_find.
And because it doesn't have weights_only param, it is failing.
we pass weights_only param in trainer.fit. Is there any way to pass this param here in lr_find?

Another Doubt

I also have one more doubt (not exactly related to ptf but out of curiosity), why the default value of weights_only param is kept as None and not True like in torch.load(see torch.load doc here) ? Is this intentional as it may break some code downstream?

See the default values of weights_only param in trainer.fit here and load_from_checkpoint param here

Thanks a lot for this amazing package! I am really a fan of your work. I would really appreciate you helping us with this!

Pitch

No response

Alternatives

No response

Additional context

No response

cc @lantiga

[bhimrazy]

Hi @phoeenniixx,
maybe you could try setting the PyTorch env vars to force the load behavior — not fully sure it covers this case, but it might unblock things in the meantime:

export TORCH_FORCE_WEIGHTS_ONLY_LOAD=1

https://github.com/pytorch/pytorch/blob/ba56102387ef21a3b04b357e5b183d48f0afefc7/torch/serialization.py#L1439C3-L1446

[bhimrazy]

As for the second question about why weights_only=None instead of defaulting to True, I think this was kept for backwards compatibility with older versions.

@matsumotosan please feel free to correct me or add any further thoughts — I might be missing some context.