Document uv-with-Jupyter workflows

[charliermarsh]

Summary

This is a work-in-progress as I actually want to change some behaviors here.

Closes #6329.

[charliermarsh]

Actually, it looks like this is solvable using kernel ipython kernel install (#6329 (comment)).

[charliermarsh]

I sort of think we should not be discovering the tool environment here, though. Modifying the project environment isn't great, but modifying the tool environment is even worse! It means subsequent users of the tool have these same modifications. I'd propose that we skip tool and ephemeral environments when considering active virtual environments. \cc @zanieb

[charliermarsh]

It would be nice if we could find a way to make this work without requiring that you install ipykernel into the project environment itself. (We don't want users to have to add ipykernel as a project dependency just to interact with their project in a notebook.)

[bluss]

Why is that not wanted?

But the 'Uv Run' kernelspec I posted in the other issue seems to be the most normal way to achieve exactly that - it runs the kernel using uv run --with ipykernel. Maybe a vscode extension is a better user interface, most of my questions are around the interface and if it's worth it (Because vscode already picks up .venv natively, is it worth the fuss with kernelspecs or other ways to do it - or can those external configuration ways of working be put on the long term plan, that is working on better uv + vscode integration to make these cases easy or automatic in the long run.)

[charliermarsh]

Well, if I'm working on a project, and I want to play with that project in a notebook, I don't want to add notebook dependencies as dependencies to the project. I don't want them recorded in the lock, etc.

[bluss]

I'm sure there are various opinions are about this, but jupyter notebook is the legacy interface and jupyterlab is the current interface. I don't know what relation other people have to this, would it give a better impression to use jupyterlab here? Command jupyter lab.

[charliermarsh]

Yeah I went back and forth. I'm fine to use jupyter lab, no big deal.

[bluss]

I think that's the best, it's what jupyter promotes as #1

[radames]

Just one note, following a fresh new setup with vscode

uv init my-notebook
cd my-notebook
uv add --dev ipykernel
uv run ipython kernel install --user --name=uv
code .

!uv pip install diffusers
/usr/bin/sh: 1: uv: not found

The notebook kernel couldn't find the correct env path for uv nor pip , not sure if it's specific to my env however uv is set correct via my .zshrc and .profile . "$HOME/.cargo/env". Doing the same steps with pip doesn't lead to an issue

[charliermarsh]

Good call, I experienced the same (though I don't see that when running Jupyter directly).

[radames]

indeed uv run --with jupyter jupyter lab works, however it looks like there are still some PATH issues. pyproject.toml is not an error, just started it from a fresh dir

[charliermarsh]

Do you mean that !pip list and !nvidia-smi are missing?

[charliermarsh]

Does !nvidia-smi work in other Jupyter notebooks for you?

[radames]

Does !nvidia-smi work in other Jupyter notebooks for you?

yes, starting a jupyter lab from a new conda env, the env path are working correctly

conda create -n temp python=3.10 
pip install jupyterlab
jupyter lab  

[charliermarsh]

I haven't been able to reproduce that unfortunately.

[zanieb]

Future: I wonder if we want uvx --with-project jupyter lab?

[charliermarsh]

Yeah I kinda like that. The repeated jupyter would be nice to avoid.

[zanieb]

Could be cool to have a command that emits the project name, so you could do --name=$(uv ...) as a one-liner that actually works in any project.

[zanieb]

Small comments