Add support for vllm model implementation

[SilverSoldier]

Changing vllm_spyre to use vllm's own model implementation instead of fms.

Status: Tested on llama and fails at self.attn() in LlamaAttention.

NOTE: Attention, KV Cache and Paging implementation is totally incomplete.

[SilverSoldier]

@tdoublep FYI, these are the changes we made in vllm-spyre side apart from the vllm changes mentioned:

• import fusion only if compilation config is set
• replace unsupported ops with their supported equivalents

[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]

Tests fail as expected 😉

This looks pretty cool!

[maxdebayser]

On spyre the decoder models are currently supported using the foundation-model-stack and the continuous batching feature that is currently being implemented relies on that library. However, encoder (embedding) models are supported in v0 using the transformers library. I'm currently working on encoder model support in V1 and I think that it would make a lot of sense to use the approach of this PR instead of continuing to run transformers.