Hot take 2: easier install

[ArthurConmy]

I think that we will get less collaborators with the current install setup. I know how to install poetry etc but without quick easy pip install -e . (and pip install -U torch==1.13.1 when my GPU isn't compatible!) there's an "ugh field" for me before I contribute.

Open to disagreeement here, just like the other issue!

[alan-cooney]

The problem with using pip is that pip doesn't check for conflicts properly, so you end up getting a load of issues with dependency conflicts down the line that are very hard to debug. But if it's a huge issue we could probably work something out.

I would say that whilst poetry isn't as commonly used as pip, it is more commonly used by big libraries for their dev setup

[alan-cooney]

Closing for now as I'm happy it's the least worst option

[ArthurConmy]

I was hoping for a solution where we use poetry, but the pyproject.toml is used by pip by default and works fine - this doesn't work with our setup as pip does not install everything needed. Is this also infeasible?

[alan-cooney]

For people using the tool but not working on the codebase, they can just pip install sparse_autoencoder.

[alan-cooney]

(as in it's on pypi)

[jbloomAus]

Yeah, this confuses me. You can install with pip, it's just that for devs, you're always going to want a higher standard and poetry is the least bad option here.

[ArthurConmy]

I don't think pypi will help much for the next ~month as I expect anyone using the library will want to modify internals in order to train SAEs with their modifications

[alan-cooney]

Yeah I think that's fair for the next month, but keen to just keep as is to avoid issues down the road with converting back to poetry. Hopefully it's fine for most people as it only takes a few minutes to get setup with poetry