[CB] consistent max context length

[yannicks1]

We had a (known) inconsistency: Allowing the additional token in platform.validate_request, but not in scheduler.can_schedule() (#332 would allow it in the scheduler). However, as the upstream vllm version vllm-spyre is tied to does not currently allow the additional token, this PR removes the inconsistency and establishes same behavior as upstream.

Note: as soon as the upstream version tied to this repo allows the additional token, we can revert this PR and merge #332 (along with some unit tests).

cc @sducouedic

[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 🚀

[maxdebayser]

Will, vllm 0.11.0 allow this? PR #513 will enable this version and raise the lower bound to 0.10.2. Also, do you know which upstream PR changed this?

[yannicks1]

it is not for a while that this comes. As working on that upstream change, I will introduce it here when it is in a supported version. Note: we can safely merge this as it is more restrictive.

[joerunde]