MMDeploy versus torch compile, export, tensorrt features

[matthost]

Newer versions of PyTorch support compilation, export, and tensorrt.

Since the mm libraries return a pytorch model (for example I use mmdetection), it seems very convenient to use these features instead of mmdeploy.

Is that fine, and would there be benefit or issues to doing so?

[matthost]

One issue is the methods in mmdet for example are not necessarily traceable. I am seeing if can replace the impls with inspiration from the impls here which were written to be traced.

It would be nice if instead of a separate package that overwrites the orginal impls, the original methods were updated to be traceable. That would lose some control flow that's here depending on the compilation config but with torch 2 compile methods probably that is no longer needed or can have separate impl as needed in the original package?