feat(attn): add paged-decode tuple 16,128,64,false,true,false

[jikunshang]

Cherry-pick from #524
Muse-Glimmer-30B uses local attention with qgroup=16, head=128, page=64. The causal=false/local=false/sink=false sibling is already compiled for GLM/Seed. Without the local=true specialization the first paged-decode warmup raises:

RuntimeError: Paged decode kernel tuple not compiled for this
configuration. 16,128,64,false,true,false

Format: qgroup,headsize,pagesize,causal,local,sink.

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.

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.

Purpose

Test Plan

Test Result

(Optional) Documentation Update

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

[Copilot]

Pull request overview

Adds a missing paged-decode kernel tuple to the default Xe2 paged-decode configuration so models requiring qgroup=16, head_size=128, page_size=64, causal=false, local=true, sink=false (e.g., Muse-Glimmer-30B) don’t fail warmup with a “tuple not compiled” runtime error.

Changes:

• Add the 16,128,64,false,true,false tuple to paged_decode_default.conf.
• Document the tuple addition inline with a short model note.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.