Separate inference dependencies from training and evaluation dependencies #2

Problem

uv sync --no-dev installs all packages declared in [project.dependencies].

For users who only want to run jeff-serve with an existing checkpoint, this also installs dependencies used for training, data preparation, evaluation, visualisation, or optional tooling.

This makes the installation larger and slower than necessary for inference-only use cases.

Proposal

Please consider providing a documented minimal inference dependency profile.

Possible approaches include:

• a server optional extra;
• separate dependency groups for serving, training, and evaluation;
• a dedicated inference requirements file.

For example:

[project.optional-dependencies]
server = [
  # dependencies required to load a checkpoint and run jeff-serve
]
train = [
  # training-specific dependencies
]
evaluate = [
  # evaluation and benchmark dependencies
]

Then an inference-only setup could use something like:

uv sync --extra server --no-dev

Expected outcome

• A smaller, faster installation for checkpoint serving.
• Training and evaluation functionality remains available as opt-in dependencies.
• The existing full-development workflow remains unchanged.
• The inference profile is covered by a small smoke test that imports and starts jeff.server.