Externally managed dependencies

• I have searched the issues of this repo and believe that this is not a duplicate.
• I have searched the FAQ and general documentation and believe that my question is not already covered.

Feature Request

In this issue/discussion, it's clear that automatically installing the correct CUDA versioned PyTorch will probably remain out of scope for Poetry until there is a standard for this and it is supported by PyPI, if ever. In that issue it was suggested that a tool like light-the-torch could be used in various ways, including hooks. However, for that to happen, Poetry would have to cooperate to some extent, essentially relinquishing control to the other tool.

[tool.poetry]
dependencies-external = ["pytorch"]

I have taken to using the current workaround myself e.g. in Dockerfiles or in CI, when needing a specific version of PyTorch:

poetry export --without-hashes > requirements.txt && \
    sed -i '/torch/d' requirements.txt && \
    python3 -m pip install -r requirements.txt && \
    # < INSTALL PYTORCH SOME OTHER WAY >>

I guess this provides a sort of specification for how this should work. It should make poetry install work as resolution happens and then the dependency is removed before it is installed. It should also avoid removing the externally managed package if it is already installed beforehand. I think having some kind of support like this within Poetry would be helpful. I think this only works at the project level i.e. when poetry install is run, and so would not change behavior at all when being used as a library.

Start by reading discussion 6680 and the project-level poetry install behavior described here, then compare it with the poetry export workaround. Done requires a settled design for declaring an external dependency, resolving it without Poetry installing it, and preserving an already-installed package.