Support torch.compile for RF

[albertz]

I'm not really sure whether that is possible because we have our own Tensor class which wraps around the torch.Tensor, and similarly all the PyTorch functions are wrapped inside RF.

However, I found out that there is TensorDict, which also wraps around PyTorch tensors, and it is explicitly stated that this is compatible with torch.compile, so maybe it is possible. In that case, it behaves like a dict, but does not inherit from dict, it just inherits from collections.abc.MutableMapping, see here.

So, let us discuss here possibilities and options to support torch.compile directly with the RF PyTorch backend.

Related:

• Speedup eager execution: Speedup eager execution #1402
• Scripting/tracing/torch.compile for PyTorch in general (not specific about RF): PyTorch torch.compile, scripting, tracing for faster computation (specifically training) #1436