[V0 deprecation] Deprecate use_v1 parameter

[wangxiyuan]

Purpose

use_v1 for get_attn_backend_cls is useless now. Since this is an interface impact change, let's deprecate it and remove it in the next release.

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request deprecates the use_v1 parameter in get_attn_backend_cls across various platform implementations. The changes are consistent and well-implemented. To provide backward compatibility for plugins, a try-except block is introduced in vllm/attention/selector.py. My review includes a suggestion to make this backward compatibility logic more robust by using inspect.signature instead of catching a broad TypeError, which will prevent masking potential bugs in plugins.

[heheda12345]

@hmellor what's the suggested way of marking an argument as will be deprecated?

[hmellor]

Ordinarily it's more user facing APIs that we are deprecating (config fields, CLI arguments, API fields). We don't really have a lot of prescedent for the plugin interface.

I think the approach that's been taken is reasonable as it ensures backwards compatibility. The only notes I would make are:

• Say in which version the deprecated argument will be removed in the warning
• Add a !!! warning to the plugin docs for extra visibility (also detailing the version it will be removed in)

[wangxiyuan]

@hmellor added the removal version to 0.12.0.

For plugin doc, I just notice that there no clear doc for hardware plugin. Only one I found is this: https://docs.vllm.ai/en/latest/design/plugin_system/#compatibility-guarantee I'd like to write a detail one for hardware plugin, include interface design, how it works, how to implement one and so on in a following PR.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @wangxiyuan.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[hmellor]

Deprecation message must be present in at least one whole minor revision. (i.e. it will be present for the entirety of v0.12.0)

Suggested change

	"be removed in v0.12.0. Please remove it from your plugin code."
	"be removed in v0.13.0 or v1.0.0, whichever is soonest. "
	"Please remove it from your plugin code."

[hmellor]

I'd like to write a detail one for hardware plugin, include interface design, how it works, how to implement one and so on in a following PR.

For now, could you please add the warning to the section you linked. And then it can be superseded with the new document you write.

[wangxiyuan]

@hmellor docs is here. #28532 I'll link it later.

[hmellor]

LGTM. I'll wait to get some more eyes on this because it's a significant change

[Yikun]

LGTM, I search all get_attn_backend_cls in here all is changed.

BTW,

@yaochengji tpu need to remove V0 support:
https://github.com/vllm-project/tpu-inference/blob/015cb2b2dc3b8fb2fe8d2e2037f2d63034daf64b/tpu_inference/platforms/tpu_platform.py#L65

@xuechendi Feel free to remove assert:
https://github.com/vllm-project/vllm-gaudi/blob/e3c20ec22d556af414fabbff12c20dc8070ec14f/vllm_gaudi/platform.py#L46

And other plugin just need remove use_v1 args

[Yikun]

++! This really a nice practice to help plugin smoothly complete interface changes. Let's remove if "use_v1" in sig.parameters in next release