[Model] Support for fairseq2 Llama

[MartinGleize]

This PR adds support for all Llama models produced with fairseq2. Its checkpoint format is mostly the same as the original Llama checkpoint format, up to layer renaming.
The PR provides a model with custom weight loading, and adds Fairseq2LlamaForCausalLM to the model registry.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[robertgshaw2-neuralmagic]

Could you add a simple example to the weight loader tests so we can confirm this wont break in the future?

[MartinGleize]

Could you add a simple example to the weight loader tests so we can confirm this wont break in the future?

Thank you for the advice, much appreciated. Just so that I'm clear, best way to do what you said is to add a HF model here: https://github.com/vllm-project/vllm/blob/main/tests/weight_loading/models.txt ?

[ywang96]

Thanks for the contribution! As @robertgshaw2-neuralmagic mentioned it would be great if you can add an example huggingface model repo in https://github.com/vllm-project/vllm/blob/main/tests/weight_loading/models.txt.

I'm also curious if it's worth adding Fairseq2LlamaForCausalLM to https://github.com/vllm-project/vllm/blob/main/docs/source/models/supported_models.md#list-of-text-only-language-models, but I'll leave it to your judgement!

[ywang96]

Is there a reason why you needed to redefine this?

vllm/vllm/model_executor/models/llama.py

Lines 553 to 554 in 8c3230d

	def _init_model(self, vllm_config: VllmConfig, prefix: str = ""):
	return LlamaModel(vllm_config=vllm_config, prefix=prefix)

[MartinGleize]

Good point, I had used this as a debugging point but forgot to remove it.