[Bugfix] Fix fully sharded LoRAs with Mixtral

[n1hility]

Fixes a regression introduced by #9008 , which leads to an assertion error when --fully-sharded-loras is enabled with an adaptor that includes a gate target:

vllm/lora/models.py:352: in __init__
    self._create_lora_modules()
vllm/lora/models.py:506: in _create_lora_modules
    self.register_module(module_name, new_module)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <vllm.lora.models.LRUCacheLoRAModelManager object at 0x7ff5ea843e10>, module_name = 'model.layers.0.block_sparse_moe.gate'
module = ReplicatedLinear(in_features=4096, output_features=8, bias=False)

    def register_module(self, module_name: str, module: "BaseLayerWithLoRA"):
>       assert isinstance(module, BaseLayerWithLoRA)
E       AssertionError

This occurs because Mixtral includes a ReplicatedLinear layer for the MoE gate, and #7081 marked it as @_not_fully_sharded_can_replace. Since these are per-GPU and the implementation looks safe I assume this was just an unintentional copy of the decorator.

This PR removes it so that ReplicatedLinearWithLora will replace ReplicatedLinear regardless of whether fully shared LoRAs are enabled, and updates the test scenario to cover both values.

Let me know if I am missing anything.

Thanks!

[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.

🚀

[jeejeelee]

Thanks for your fixing. Currently Mixtral is not being tested on CI. I assume you have already tested it locally.

[n1hility]

Thanks for your review @jeejeelee! I ran the test LoRA Mixtral tests locally (before and after change) as well as test interaction with a booted vllm instance using Mixtral with LoRA adapters. I can look for and run additional Mixtral tests in the vllm testsuite locally if these aren’t always run.

[n1hility]

Thanks for your review @jeejeelee! I ran the test LoRA Mixtral tests locally (before and after change) as well as test interaction with a booted vllm instance using Mixtral with LoRA adapters. I can look for and run additional Mixtral tests in the vllm testsuite locally if these aren’t always run.

@jeejeelee These all passed too ( I ran the other Mixtral tests outside of tests/lora (in addition to the ones in tests/lora))

It looks like the reported CI failures in entrypoints-test are also failing on main so those are unrelated.