V1 embeddings

[maxdebayser]

Description

This PR enables embedding models on vllm V1. In contrast with the V1 GPU implementation, here I added a separate model runner because for most of the embedding models there is no need for continuous batching. To avoid code repetition, I refactored the input batch and model runner classes into a class hierarchy with common base classes.

@gmarinho2 contributed a test that verifies that the returned embeddings don't change with batch size.

[github-actions]

👋 Hi! Thank you for contributing to vLLM support on Spyre.
Just a reminder: Make sure that your code passes all the linting checks, otherwise your PR won't be able to be merged. To do so, first install the linting requirements, then run format.sh and commit the changes. This can be done with uv directly:

uv sync --frozen --group lint --active --inexact

Or this can be done with pip:

uv pip compile --group lint > requirements-lint.txt
pip install -r requirements-lint.txt
bash format.sh

Now you are good to go 🚀

[yannicks1]

I am through 7/10 files, need to continue later... This is a huge PR:)

[yannicks1]

hmm, might there be another, more elegant way to distribute the method _get_padded_batch_size to both SpyrePoolingModelRunner, and StaticBatchingSpyreModelRunner ? Also, the name WarmupShapesMixin sounds a bit weird to me 😄 . Tagging our most experience SWE here:) @joerunde @tjohnson31415

[maxdebayser]

Well, mixin classes are one of the standard ways to solve this problem. But maybe there are better options.

[yannicks1]

Oh, it is just that I have never seen them before:) my apologies, learning new things 😄 Naming is fine too, now that I better understand. I initially read it like "Warmup shapes mixing" rather than "Warmup shapes mix-in"...

[joerunde]

Python doesn't have a standard way to do object composition so there are quite a few different ways to do it, each with their own drawbacks. This is a classic overview worth a read if you have a couple minutes: https://python-patterns.guide/gang-of-four/composition-over-inheritance/

My two cents is that mixins are better than filling up a utils package with a bunch of shared methods, and then once you have a bunch of mixins you can usually find a pattern to apply instead that doesn't use multiple inheritance.

[yannicks1]

thanks a lot. will read this in a quiet minute:)

[maxdebayser]

All tests are passing now after the changes from the first round of reviews.