[Fix] Tests breaking with vLLM:Main

[sducouedic]

Closes #305

The changes in upstream that are breaking tests with vLLM:main branch are twofolds:

1. This PR introduces an import which is incompatible with torch 2.5.1
2. Previously, when model_config.task was set to auto, the task could be automatically resolved in vLLM: after the post-init method of ModelConfig, the task field was automatically replaced from auto to .e.g. generate or embed). With the changes in This PR, the task in model_config, cannot be automatically resolved anymore when it is generate. This doesn't look like it is intended, but rather a side effect of adding new supported tasks. A cleaner (but longer) way would be to do a PR to vLLM to fix that and reintroduce the automatic resolution of auto task field.

This PR addresses both of the breaking changes:

• for 1. it installs pytorch==2.7.0 in github workflows when vLLM:main branch is used
• for 2: it infers the task through model_config.supported_tasks rather than model_config.task now

[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]

LGTM!

Looks like the last failure is due to cleanup not happening properly after running with tensor parallel. We can go ahead and merge this and continue to investigate that failure.

Once we have a base image up including a 2.7.1 version of torch we can standardize on that as well

[maxdebayser]

LGTM