[FEATURE] Enables offline /score for embedding models

[gmarinho2]

Enables LLM.score() for all embedding models. The request_id consists of the request_ids of each embedding in the pair, separated by "_". The prompt_token_ids are the concatenation of all the token ids, in order and separated by the padding token when it is available. This PR is the first of two for completing the issue. The second PR will implement the same feature in the OpenAI API.

Issue: [Feature]: Enable /score endpoint for all embedding models (1/2)

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[joerunde]

@maxdebayser @gmarinho2 This looks like it only touches the offline entrypoint, but the PR title mentions /score endpoint which I had assumed meant the online REST interface.

It's not 100% clear to me from the linked issue either what was intended- is there more work planned to support the online interface or are we only aiming for offline?

[maxdebayser]

@joerunde, we're aiming for both. @gmarinho2 started with the offline API first.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @gmarinho2.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[maxdebayser]

I've left some suggestions, but it looks good to me. I think we can open this as a PR now.

[DarkLight1337]

Some initial comments.

I also suggest splitting out the logic for scoring and general embedding models into separate functions.

[DarkLight1337]

I also suggest splitting out the logic for scoring and general embedding models into separate functions.

Have you addressed this?

[DarkLight1337]

Suggested change

Avoid unnecessary line changes

[DarkLight1337]

I still see this line in the PR. Never mind, the line is for _embedding_score. Let's just merge this.

[joerunde]

Added the ready tag so the full CI run can run your tests as well. Looks like the only fastcheck failure was a network issue

[DarkLight1337]

Sorry for the confusion, looks good now.