Pinning system dependencies in pyproject.toml

[thepowerfuldeez]

Hi! I have a question which is related to using NGC Docker image with pre-release torch2.5.0 included. Version inside docker container (which comes with pre-installed CUTLASS, CUDNN and are of correct versions) is 2.5.0a0+872d972e41.nv24.08

This version is not resolvable if I add it directly to pyproject.toml in a list of dependencies. If I change to torch2.4.0 stable then all set of requirements from this torch version will be downloaded (including cuda bindings, etc). This might break compiled CUDA libraries that were included inside docker image, and increase install time.

Current approach:

1. Use docker image nvcr.io/nvidia/pytorch:24.08-py3
2. Set UV_SYSTEM_PYTHON=1
3. Install dependencies with lower level uv pip cmd: uv pip install --system -r pyproject.toml

Problem with that approach is that library versions are not explicitly pinned, and it's not usable in production settings. I would like to add pre-release installed torch version to pyproject.toml, let is resolve dependencies when I run uv lock and then use higher-level project management command uv sync to sync with lockfile. I could still use system python to circumvent setting up required venv, since there's conda env inside docker image with installed libraries.

Required featues:

1. Setting up pre-release and installed libraries inside system python
2. Running uv lock --system and uv sync --system to use system python and not install existing libraries (how it's done with uv pip install --system at the moment.

Thank you very much!

[thepowerfuldeez]

@charliermarsh do you have an opinion?