Make DL Packages Optional

[AlexMV12]

Feature request

Make the packages needed for DL (e.g., Torch and Brevitas) optional, or create flavors of Concrete-ML that include only a subset of dependencies (e.g., concrete-ml[sklearn] which would include only the linear models).

Motivation

This would be particularly useful to reduce the size of Docker images based on Concrete-ML: for instance, adding Concrete-ML with uv or poetry produces, by default, Docker images of >6GBs. With a bit of tweak, we can force it to use the CPU version of Torch, which still produces a Docker image of >2GBs. Considering that in some cases only a subset of models is needed, it could be interesting.

However, I don't know the structure of the code, so it may break a lot of things creating different subpackages/install options. Feel free to close this if this is unpractical. Thanks.

[jfrery]

That would be great indeed but as you say, it would be pretty hard to do this since a lot of things would break without torch / brevitas. If you feel like deep diving into this we would be happy to review your PR!

[AlexMV12]

Hello @jfrery ,
Thanks for your answer. I took some time to think about it, and I wanted to put on the table the possibility to think about a concreteml[client] version. This would include only the part needed for quantization and encryption (and obviously dequantization and decryption).
Unfortunately I don't feel confident enough to tackle this in your library, I'll just leave here the idea. Thanks again!