feat: add a pre-AOT lowering pass to remove detach ops

[zewenli98]

Description

Added a pre-AOT lowering pass to remove detach ops

Fixes #2657

Type of change

• New feature (non-breaking change which adds functionality)

Checklist:

• [x] My code follows the style guidelines of this project (You can use the linters)
• [x] I have performed a self-review of my own code
• [x] I have commented my code, particularly in hard-to-understand areas and hacks
• [ ] I have made corresponding changes to the documentation
• [ ] I have added tests to verify my fix or my feature
• [ ] New and existing unit tests pass locally with my changes
• [x] I have added the relevant labels to my PR in so that relevant reviewers are notified

[apbose]

Overall looks good to me. Is there any use case for which this lowering pass has been added?

[zewenli98]

Overall looks good to me. Is there any use case for which this lowering pass has been added?

Yes, in the vLLM, without this lowering pass, it causes some errors during compilation.

[zewenli98]

I've implemented your fix (refactored it a bit) as a part of https://github.com/pytorch/TensorRT/pull/2756/files. Can you take a look and make these changes? We need to add this pass to both torch.compile and torch.export as well.

Where can I find your implementation? The link seems not including your change.

[peri044]

I've implemented your fix (refactored it a bit) as a part of https://github.com/pytorch/TensorRT/pull/2756/files. Can you take a look and make these changes? We need to add this pass to both torch.compile and torch.export as well.

Where can I find your implementation? The link seems not including your change.

My bad, here's the correct link : https://github.com/pytorch/TensorRT/pull/2763

[zewenli98]

Hi @peri044 I checked your PR and found you made a lot of changes like pre_export_lowering and post_lowering with the help of PassManager. I will rebase on your PR after it gets merged.

Besides, I noticed that _pretraced_backend() in backends.py and compile() in _compiler.py are pretty similar. Why do we need the separate functions?

[peri044]

Hi @peri044 I checked your PR and found you made a lot of changes like pre_export_lowering and post_lowering with the help of PassManager. I will rebase on your PR after it gets merged.

Besides, I noticed that _pretraced_backend() in backends.py and compile() in _compiler.py are pretty similar. Why do we need the separate functions?

We use the same function now https://github.com/pytorch/TensorRT/pull/2763. Closing this PR now