Improve UX by letting the user know that their `torch` and `torchcodec` versions are incompatible

[hbredin]

🚀 The feature

Improve error message when torchcodec version is not compatible with torch version.

Before, user gets an obscure segmentation fault

terminate called after throwing an instance of 'std::bad_alloc'
  what():  std::bad_alloc
Aborted (core dumped)

After, user gets a meaningful error message

Your version of `torchcodec` (0.8) is incompatible with `torch` (2.8.0)`. Please downgrade `torchcodec` to 0.7 of upgrade `torch` to `2.9.0`.

Motivation, pitch

For some reason, I (and a bunch of my pyannote.audio users) ended up with an environment with torchcodec 0.8 and torch 2.8.0 which (I understand from the docs) are incompatible.

Running import torchcodec with version 0.8 of torchcodec and version 2.8.0 of torch leads to an obscure bad_alloc error message (+ core dump).

An obvious solution would indeed be for pyannote.audio to pin torch and torchcodec versions but this usually leads to incompatibility with other libraries.

A nice error message would help users figure this out by themselves.

[scotts]

@hbredin, thanks for the suggestion! This is something we can investigate - but depending on when this error happens, it's possible there's not much we can do.

[traversaro]

@hbredin, thanks for the suggestion! This is something we can investigate - but depending on when this error happens, it's possible there's not much we can do.

An alternative is to just do the check an init/load time, and trow an early error there, without waiting for the actual error to appear. See for example https://github.com/openxla/xla/blob/cc5a4b4be6d3689d39860023b9440fc82467e661/xla/stream_executor/cuda/cuda_dnn.cc#L447-L455 .