External vLLM instance

[haydn-jones]

Same thing as #309 but a bit fewer changes. Happy to close mine if we're going with that one, just want the feature added.

Tested with

python -m olmocr.pipeline ./localworkspace --markdown --pdfs ./tests/gnarly_pdfs/ambiguous.pdf --external-vllm-url http://10.0.0.175:30025

Should probably have called it something other than external-vllm-url as presumably will work with anything that exposes the OpenAI API (e.g. sglang)

[jakep-allenai]

I like this version, it looks very clean. What about calling it --server ?

[haydn-jones]

Updated it to --server and minimally updated the README.

[jakep-allenai]

Looks good thank you

[jakep-allenai]

Let me run the linters and then merge

[jakep-allenai]

Can you just run isort . and black . please?
And then I think pytest is throwing an error due to the added arg.

[haydn-jones]

Sure will do.

[haydn-jones]

Added the mock arg and ran black.

[jakep-allenai]

Thank you!

[Zoe-Leaf]

This change has been merged? Why don't I see this parameter in help?

[jakep-allenai]

I haven't made a new release yet, you'll have to install from the git repo instead of pip.

[Zoe-Leaf]

Gotcha, thanks

[simonmoesorensen]

I haven't made a new release yet, you'll have to install from the git repo instead of pip.

First of all, big fan of your work @jakep-allenai ! The model performs outstandingly

I ran into the same confusion as @Zoe-Leaf

In my team we use a main and develop branch approach. The main represents the code in its currently released state, and develop is for upcoming features and the default branch for PRs. When we release a new version of the code, we merge from develop -> main.
This helps us distinguish between the released and development codebase.

Do you think that something similar could be implemented here?

[Zoe-Leaf]

I haven't made a new release yet, you'll have to install from the git repo instead of pip.

First of all, big fan of your work! The model performs outstandingly

I ran into the same confusion as @Zoe-Leaf

In my team we use a main and develop branch approach. The main represents the code in its currently released state, and develop is for upcoming features and the default branch for PRs. When we release a new version of the code, we merge from develop -> main. This helps us distinguish between the released and development codebase.

Do you think that something similar could be implemented here?

This is what I want to say, too! So as to avoid such misunderstandings