fix(gguf): Make GGUFMoEMethod.apply() parameters optional

[kitaekatt]

Summary

• Makes top_k and renormalize keyword arguments with defaults in GGUFMoEMethod.apply()
• These parameters were required but never passed by the caller nor used in the method body

Problem

GGUF MoE models (e.g., Qwen3-30B-A3B-Instruct-GGUF) fail with:

TypeError: GGUFMoEMethod.apply() missing 2 required positional arguments: 'top_k' and 'renormalize'

Root Cause

• FusedMoE.forward_impl() in layer.py:1946 calls quant_method.apply(layer=self, x=..., router_logits=...)
• Base class FusedMoEMethodBase.apply() only requires (layer, x, router_logits)
• GGUFMoEMethod.apply() had extra required positional args that weren't being passed
• These args were also not used in the method body (it calls layer.select_experts() without them)

Test plan

• Test with Qwen3-30B-A3B-Instruct-GGUF model
• Verify GGUF MoE inference works correctly

🤖 Generated with Claude Code

[mergify]

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

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

[gemini-code-assist]

Code Review

This pull request correctly addresses a TypeError in GGUFMoEMethod.apply() by making the top_k and renormalize parameters optional. According to the pull request description, these parameters were required but not always passed by the caller, and as they are not used within the method body, making them optional with default values is a safe and effective fix. The change is minimal and directly resolves the bug affecting GGUF MoE models. The implementation looks good.

[kitaekatt]

Closing this PR as it's no longer needed.

Upstream PR #29066 ([MoE][Refactor] Remove most arguments to FusedMoEMethodBase.apply) resolved the underlying issue by removing the extra arguments from the method signature entirely.

Thanks to the vLLM team for the comprehensive MoE refactor!