Skip to content

feat(dcu-kernel-auto-opt): sync the DKAO task package (no variant library) - #42

Open
fms-zth wants to merge 1 commit into
HuangPuStar:next-appfrom
fms-zth:feat/dkao-remote-sync-20261013
Open

fms-zth wants to merge 1 commit into
HuangPuStar:next-appfrom
fms-zth:feat/dkao-remote-sync-20261013

Conversation

@fms-zth

@fms-zth fms-zth commented Oct 13, 2026 •

Copy link
Copy Markdown
Contributor

概要

把 DCU 开发机上的 DKAO(dcu_kernel_auto_opt)任务包同步回来(不含变体库,见下)。

变更范围:只改 metainfer/tasks/dcu_kernel_auto_opt/ 一个目录,101 文件,+19,886 / −975,0 删除。

说明:本 PR 不是自包含的。 DKAO 引用的若干公共模块(metainfer.orchestrator.hw_profile / gpu_broker / gpu_inventory / interpreter、metainfer.server.envcheck 等)在 next-app 上尚不存在,计划由后续整合 PR 补齐。因此单独合并本 PR 时 DKAO 无法运行,这是有意为之——先把任务包本身落库,公共层随后对齐。

本 PR 不再新增 variant 算子

上一轮 DKAO PR(#34 / #36)把实测得到的变体算子提交进了仓库,本次不再这样做。

原因:每个 variant 叶子是针对特定板子的实测记录,其 @@variant 头部携带 median_us / p90_us / speedup / hw_profile,而数值只对产出它的板子有意义。单机实测树约 160 个文件 / 16 万行生成 HIP——放进 PR 会把真正需要 review 的逻辑代码(约 2 万行)淹没。

本次做法:保留目录布局与上游已有样本,新增 variant/.gitignore 把 *.hip 排除在版本控制之外:

*
!.gitignore
  • 本地照常保留完整变体库,供 DKAO 运行时使用(运行时只读挂载到生成仓库的 references/<family>/)
  • 需要变体库的机器自行跑 DKAO 生成,不再依赖仓库分发
  • 已在仓库中的旧布局算子(variant/int8w8a8-gemm/...)本 PR 不动

变更内容

Harness 与策略

  • planner、gate policy、round strategy、final layout gate、contention 处理
  • error_catalog/:按类别归档的失败样本与判据(environment / generate-and-agent / measurement-and-gate / kernel-quality / process-and-infra)
  • harness_default/、WORKFLOW.md 更新

知识资产

  • skills/:按板子分册(air-k100ai / liquid-k500sm)的内核优化知识
  • baselines/:两块板子的实测基线(air-k100ai.json / liquid-k500sm.json)

工程

  • bridge/dsh/:DSH agent 桥接
  • reports/:算子汇总与诊断工具(含 concise CSV/XLSX)
  • api/、server/、tools/ 内部更新

仓库卫生

删除 28 个编辑器遗留的 *.bak-<ts> 文件(无 provenance 头部,非已采纳变体)。

验证

本 PR 只改 DKAO 包内文件,未触碰公共层。完整导入与测试验证建议在公共模块补齐后进行:

python -m pytest metainfer/tasks/dcu_kernel_auto_opt/tests

@fms-zth
fms-zth force-pushed the feat/dkao-remote-sync-20261013 branch from 21a4402 to 0c9b7f1 Compare October 13, 2026 05:37
@fms-zth fms-zth changed the title feat(dcu-kernel-auto-opt): profile-layered variant library and hardware-aware DKAO harness feat(dcu-kernel-auto-opt): sync the DKAO task package from the DCU dev container Oct 13, 2026
…brary

Bring the dcu_kernel_auto_opt task package up to date with the DCU dev
machine, but deliberately leave the measured variant library out of the
repository.

Why no variants here: each variant leaf is a per-board measurement record
(its '@@variant' header carries median_us / p90_us / speedup / hw_profile).
A number is only meaningful against the board that produced it, and a single
machine's tree is ~160 files / ~160k lines of generated HIP. Shipping that in
every DKAO PR drowns the actual review surface, so the tree now stays local:
variant/.gitignore keeps '*.hip' untracked while the layout and the upstream
samples remain in place. A machine populates its own library by running DKAO.

Included: planner / gate policy / round strategy / final layout gate /
contention handling, error_catalog, skills, baselines, bridge/dsh, reports,
WORKFLOW.md and harness_default updates.
Also drops 28 editor '*.bak-<ts>' leftovers.

Scope: only metainfer/tasks/dcu_kernel_auto_opt/.
@fms-zth
fms-zth force-pushed the feat/dkao-remote-sync-20261013 branch from 0c9b7f1 to 097a4f8 Compare October 13, 2026 05:44
@fms-zth fms-zth changed the title feat(dcu-kernel-auto-opt): sync the DKAO task package from the DCU dev container feat(dcu-kernel-auto-opt): sync the DKAO task package (no variant library) Oct 13, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant