Use deformable_detr kernel from the Hub

[danieldk]

What does this PR do?

Remove the deformable_detr kernel from kernels/ and use the pre-built kernel from the Hub instead.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@ArthurZucker @LysandreJik

[github-actions]

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. When it is ready for review, please click the Ready for review button (at the bottom of the PR page).

[ArthurZucker]

Super super super nice! 🚀 let's just wait for @LysandreJik 's opinion on adding a new core dep!

[ArthurZucker]

That's the only place I'd want to wait @LysandreJik's input on! IMO this is the longterm plan for sure, wondering if we are gonna do that slowly (first optional soft dep) or not!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

Let's GO!!! 🤗

[ArthurZucker]

perfect!

[qubvel]

Hey @danieldk! Super nice to see kernels integrated! Is there a way to disable the kernel path dynamically? RT-DETR's compile full graph and torch.export are broken now, they should go through the torch path.

[ArthurZucker]

Ah normally if cuda is not available it should fallback to noraml forwar + if not available should do nothing

[ArthurZucker]

(@qubvel export can be fast test no? sorry that it broke)

[qubvel]

it's too slow, even though it's on a small model

[qubvel]

we discussed with @danieldk in Slack, the idea is to have a way to control which path to go through, smth like

if self.disable_custom_kernels or is_torchdynamo_compiling():
    with use_kernel_mapping({}, inherit_mapping=False): # inherit_mapping would be a new argument
        output = self.attn(
            value,
            spatial_shapes,
            spatial_shapes_list,
            level_start_index,
            sampling_locations,
            attention_weights,
            self.im2col_step,
        )
else:
    ...

[ArthurZucker]

Or we register the use kernel with torch API to allow compile ?

[qubvel]

Not aware of this but it can be an option, yes. Anyway it would be nice to be able to disable it manually in case there are any issues with the kernel + understand which path is executed (e.g. for RT-DETR to export for CoreML we need 5D tensors instead of 6D passed into the kernel, so I refactored pytorch path of deformable attention in this PR)