Restore is_torch_greater_or_equal_than for backward compatibility

[tlrmchlsmth]

This PR restores booleans such as is_torch_greater_or_equal_than_1_13 for backward compatibility reasons. These were removed in #35358, and I've set them to True as Tranformers not requires PyTorch >= 2.0

For context, I am trying to upgrade vLLM to use transformers 4.48.0 in vllm-project/vllm#12120. One of the issues I'm running into is that MiniCPM3-4B
's checks is_torch_greater_or_equal_than_1_13 here https://huggingface.co/openbmb/MiniCPM3-4B/blob/e5715484011d723a1892db91da8b59d979d14aee/modeling_minicpm.py#L63-L65.

I've put up a PR to remove that check in MiniCMP3-4B but it seems likely that there are other models that will hit the same issue, I think it would be good to restore these at least for the short term.

cc @ydshieh

[Rocketknight1]

@ArthurZucker WDYT? I had to do some emergency PRs to MiniMax because of this, so maybe keeping them as static True makes sense for backward compatibility

[tlrmchlsmth]

@Rocketknight1 FYI the link to my PR that does the same thing for MiniCPM3 is https://huggingface.co/openbmb/MiniCPM3-4B/discussions/39

[ydshieh]

OK to restore for backward compatibility, but let's do it as before and not just set to True to avoid unpleasant surprise.
Thank you for bringing this issue and the PR

[tlrmchlsmth]

OK to restore for backward compatibility, but let's do it as before and not just set to True to avoid unpleasant surprise. Thank you for bringing this issue and the PR

Sounds good -- updated.

[ydshieh]

@Rocketknight1 I just merged this PR

[Rocketknight1]

Yes, sounds good! Thanks @ydshieh!

[noldorj]

I still happens in
Torch version: 2.5.1+cu124
Transformers version: 4.48.1

[ydshieh]

Hmm, I don't think the fix is included in the patch.

Would it be good for you to using (for now) the dev version or 4.48 or an older version v4.47.1?

Sorry for the inconvenience.

[tlrmchlsmth]

For vLLM, we're on 4.45.2, but would like to pick up 4.48 for bamba support

[ArthurZucker]

yep working on the patch!

[ArthurZucker]

will probably ship now and have another one for the num items in batch issue

[abidkhan484]

For vLLM, we're on 4.45.2, but would like to pick up 4.48 for bamba support

Same error with transformers 4.48.1 and torch 2.5.1 here.

[ydshieh]

For vLLM, we're on 4.45.2, but would like to pick up 4.48 for bamba support

Same error with transformers 4.48.1 and torch 2.5.1 here.

the fix in this PR is not included in 4.48.1. We will have a new patch.