Bugfix device map detr model

[pedrogengo]

What does this PR do?

Fixes #26700 #23145

Who can review?

@LysandreJik @SunMarc

[SunMarc]

Hi @pedrogengo thanks for making detr model compatible with device_map ! The PR looks good overall. I don't think that we need to add tests for device_map because we already have them. You just need to make sure that you can pass the accelerate tests for all models. This will trigger the following tests: test_disk_offload, test_cpu_offload and test_model_parallelism. Please use 2 GPUs to tests them. If you don't have them, i can run the tests for you!

[pedrogengo]

Hi @SunMarc :) I don't have 2 GPUs unfortunately. Can you run on your side? In the meantime I can remove the tests I added + fix the style issue

[pedrogengo]

@SunMarc addressed your comment on my side. Could you test?
I also noticed a bad pattern maybe on check_copies.py. At first, I tried this:

# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrPreTrainedModel with DeformableDetr->Deta,DeformableDetrConvEncoder->DetaBackboneWithPositionalEncodings

And it was breaking the CI, because it first replace DeformableDetr to Deta, and after search for the next pair, which was DeformableDetrConvEncoder->DetaBackboneWithPositionalEncodings. However, as we first replaced, it turns that in the code DeformableDetrConvEncoder was changed to DetaConvEncoder. Maybe we can sort by length and start replacing from the largest string, and also replace with a placeholder in both codes to avoid replacing more than one time. Just something that I saw haha I can propose a PR for that, but I don't think it is that much important.

[SunMarc]

LGTM ! I've tested the accelerate tests on 2GPUs and they passed. Feel free to raise an issue or a PR to fix the pattern on check_copies.

[amyeroberts]

Thanks for fixing @pedrogengo, and thanks @SunMarc for running the accelerate tests!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.