[Misc] Allow LoRA to adaptively increase rank and remove possible_max_ranks

[JinhyunBang]

• Removes possible_max_ranks to serve LoRA adapters with an arbitrary rank.
• When provided with a LoRA adapter with rank larger than the previously set max rank, adaptively increase the matrix sizes rather than throwing an error.

Related Issues: #2847, #3310, #3934

[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]

@mgoin do you know why the LoRA sizes are fixed like that? Is it because we have compiled ops for this?

cc @jeejeelee

[jeejeelee]

When provided with a LoRA adapter with rank larger than the previously set max rank, adaptively increase the matrix sizes rather than throwing an error.

I'm curious whether the changes are compatible with CUDA graphs. Have you tested this?

[jeejeelee]

@mgoin do you know why the LoRA sizes are fixed like that? Is it because we have compiled ops for this?

See:#1804. IIUC, I thinks due to fixed LoRA slots.

[mgoin]

I agree it would be nice to support larger ranks if the user needs it, but this should maybe be a set argument. I'm not sure if this will work as-is, so I need to see some testing added