[DNM][WIP]attn: convert boolean template params to runtime dispatch

[jikunshang]

UPDATE: we found perf regression(~15%), we will not merge this, just make it as a reference for now and will root cause.

Remove 5 boolean template parameters (Paged, Causal, Local, Sink, SoftmaxLSE) from chunk_prefill and 3 (Causal, Local, Sink) from paged_decode kernel instantiation chains. Boolean flags are now passed through the existing args structs and evaluated at runtime with regular if instead of if constexpr.

Policy (tile-size) templates are preserved as they must remain compile-time for SYCL register allocation.

Results (default config build):

• Build time: 15m21s → 6m45s (-56%)
• Template source files: 43 → 17 (-60%)
• Full-build template files: 600 → 60 (-90%)
• Binary size: ~unchanged (43M → 42M)
• Performance: no regression (3276 tests pass)
• Config files no longer need bool columns (backward compatible)

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

This PR refactors Xe2 attention kernel instantiation to reduce compile-time blowup by moving several boolean “feature” template parameters (paged/causal/local/sink/softmax_lse) into runtime flags carried through existing argument/parameter structs, while preserving compile-time policy templates that affect SYCL register allocation.

Changes:

• Replace boolean template parameters in chunk_prefill and paged_decode instantiation chains with runtime bool fields and if branches.
• Simplify generated extern/instantiation headers and CMake kernel-generation logic to build per-policy kernels (not per-boolean-combination kernels).
• Update mainloop/epilogue argument plumbing so runtime feature flags reach device code.

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
csrc/xpu/attn/xe_2/paged_decode.hpp	Pass runtime flags through kernel/mainloop/epilogue argument structs; remove bool template params from config path.
csrc/xpu/attn/xe_2/paged_decode_utils.hpp	Remove bool-combination dispatch; gate only on enabled policies.
csrc/xpu/attn/xe_2/paged_decode_kernel_template.cpp.in	Generate explicit instantiations per policy (no bool macro axes).
csrc/xpu/attn/xe_2/paged_decode_extern.hpp.in	Extern template declarations per enabled policy only.
csrc/xpu/attn/xe_2/paged_decode_extern.hpp	Extern template declarations per policy only (non-generated header).
csrc/xpu/attn/xe_2/paged_decode_enabled_policies.hpp.in	Drop policy+bool tuple traits; keep per-policy enablement.
csrc/xpu/attn/xe_2/paged_decode_configure.cmake	Parse config as (qgroup, headsize, pagesize) and generate one TU per policy; ignore trailing bools.
csrc/xpu/attn/xe_2/kernel/paged_decode_kernel.hpp	Add runtime feature flags into params; convert some if constexpr feature branches to runtime if.
csrc/xpu/attn/xe_2/kernel/chunk_prefill_kernel.hpp	Move paged/causal/local/sink/softmax_lse decisions to runtime flags in params.
csrc/xpu/attn/xe_2/fmha_xe2.cpp	Remove prior runtime restriction around softmax_lse specialization and simplify dispatch calls.
csrc/xpu/attn/xe_2/collective/chunk_prefill_mainloop.hpp	Remove bool template axes and plumb runtime flags into mainloop params; convert masking/paging logic to runtime.
csrc/xpu/attn/xe_2/collective/chunk_prefill_epilogue.hpp	Remove sink bool template axis; plumb sink as runtime parameter and branch in epilogues.
csrc/xpu/attn/xe_2/chunk_prefill.hpp	Pass runtime flags to kernel/mainloop/epilogue args and wire has_softmax_lse.
csrc/xpu/attn/xe_2/chunk_prefill_utils.hpp	Remove bool-combination dispatch; gate only on enabled policies.
csrc/xpu/attn/xe_2/chunk_prefill_kernel_template.cpp.in	Generate explicit instantiations per policy (no bool macro axes).
csrc/xpu/attn/xe_2/chunk_prefill_extern.hpp.in	Extern template declarations per enabled policy only.
csrc/xpu/attn/xe_2/chunk_prefill_extern.hpp	Extern template declarations per policy only (non-generated header).
csrc/xpu/attn/xe_2/chunk_prefill_enabled_policies.hpp.in	Drop policy+bool tuple traits; keep per-policy enablement.
csrc/xpu/attn/xe_2/chunk_prefill_configure.cmake	Parse config as head sizes and generate one TU per policy; ignore trailing bools.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[rogerxfeng8]

What's the wheel size with the PR using the full config?

[jikunshang]

What's the wheel size with the PR using the full config?

67MB https://github.com/vllm-project/vllm-xpu-kernels/actions/runs/28481312527?pr=444

[jikunshang]

UPDATE: we found perf regression(~15%), we will not merge this, just make it as a reference for now and will root cause.