[BUG] Compatibility with weights_only torchload

[vonHartz]

Describe the bug

To prevent arbitrary code execution during unpickling, torch is moving towards using weights_only=True for torch.load.
I've tried to torch.load(weights_only=True) on some of my saved TensorDicts.

Problem is that TensorDicts are not considered save by default. And adding them manually still fails.

To Reproduce

Save a TensorDict to disk and load it back with (weights_only=True).

This will fail due to Tensordicts not being considered save by default.

I've tried the following:

torch.serialization.add_safe_globals([td.TensorDict, td.utils._StringOnlyDict])```

But still get:

```WeightsUnpickler error: Unsupported global: GLOBAL tensordict.utils._StringOnlyDict was not an allowed global by default. Please use `torch.serialization.add_safe_globals([_StringOnlyDict])` to allowlist this global if you trust this class/function.

Expected behavior

A save option to load Tensordicts would be nice.

System info

Describe the characteristic of your environment:

• td 0.83.
• torch 2.5.1
• Python 3.10

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[vmoens]

I think it's expected - weights-only will not accept any non-builtin python class (@mikaylagawarecki can you confirm? - what are our options here?)

[vonHartz]

Yes, I think that's fair. The problem is that adding TensorDict as a safe does not seem to work.