[Bugfix] Fix fully sharded LoRA bug

[jeejeelee]

FIX #10342 (link existing issues this PR will resolve)

The issue arose because the code did not properly handle cases where all three lora_a
in MergedQKVParallelLinearWithShardedLora could be None. I also handled this issue in MergedColumnParallelLinearWithShardedLoRA.

BTW, fix small typo in worker.py

ping @DarkLight1337 also cc @xyang16 Thanks for catch this bug

[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.

🚀

[DarkLight1337]

Can we add a test to avoid similar regressions?

[jeejeelee]

Can we add a test to avoid similar regressions?

I've been thinking about this.
current LoRA CI tests only run on a single GPU, which means many TP related logic cannot be verified in CI. Issues like this one cannot be catched when testing on a single GPU.

[DarkLight1337]

We can add some LoRA tests to distributed tests.

[jeejeelee]

We can add some LoRA tests to distributed tests.

I'll do that, but it will probably take some time. Regarding this bug, my local testing shows no issues, including cases where qkv_proj, gate_proj, and up_proj are None.

[DarkLight1337]

Alright, let's try to get this into the upcoming release first.