✨ vllm support for 0.11.1 release

[joerunde]

Description

Upgrades vllm to 0.11.1, adding backwards compatibility code where necessary.

This PR:

• Updates the default vllm install to 0.11.1
• Retains the lower bound of 0.10.2
• Adds a new entry in the backwards compatibility tests to maintain test coverage of 0.11.0
• Changes the uv.lock settings to install vllm from source instead of from cuda wheels

There was one really fun change here where the type of sampled_token_ids changed, but was then changed back for 0.12.0.

TODO: There is still a problem with running qunatized models. I'm not sure what's going on there, as neither the torch version nor modeling code changed, but we're getting an error from torch 🤔

[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 🚀

[joerunde]

Installing vllm this way (with VLLM_TARGET_DEVICE=empty) leaves out extra cuda-only dependencies from the uv.lock, since the published vllm wheels on pypi are only built for cuda.

[tjohnson31415]

Somehow you make backwards compatiblity elegant

[tjohnson31415]

It looks like we could just import and use _get_extra_args() from the spyre_worker to reduce code duplication.

[joerunde]

private imports!!

but yeah, for a test file that's probably fine

[tjohnson31415]

Is this a TODO for this PR to fix before merging?

[joerunde]

oh- maybe 🤔

I think I put the TODO in because the lazy import was suuuper ugly, but I do think the import has to stay lazy or we'll hit a circular import :(. The TODO here might be to just remove the TODO and replace with a comment about why this is the way it is

[tjohnson31415]

TODO: There is still a problem with running qunatized models. I'm not sure what's going on there, as neither the torch version nor modeling code changed, but we're getting an error from torch

This PR bumps fms-model-optimizer to 0.7.0 in uv.lock. I confirmed the quantized model tests fail after upgrading 0.6.0 -> 0.7.0. Installing fms-mo from main resolved the torch error in my dev pod.

[joerunde]

Alright @tjohnson31415, looks like we are 🟢 for now. Thanks for the fms-mo hint, I validated that fms-mo 0.7.0 still works on spyre and it's just the cpu execution that's broken. I've bumped here to the latest main commit, which also appears to work fine on spyre too.

Let's talk on Monday- maybe we should get a new official fms-mo release instead of pinning a commit, and then I'm not entirely sure with our current release cadence whether we'd want to bump the actual vllm install to 0.11.1 or flip this around and just add a compatibility test for 0.11.1 and keep the uv.lock at 0.11.0. Then either way we should get the currently-good set of spyre unit tests run on this before merging