Merged
docs(gmm): 补充 GMMFR、Swiglu 与 InplaceAdd 转置融合指南 #15168
zhoushaolongAdd on 7 days ago
docs(gmm): 补充 GMMFR、Swiglu 与 InplaceAdd 转置融合指南 #15168
Merged
Pull Request has been successfully merged by CANN-robot.
(Thank you for zhoushaolong's contribution.)7 days ago Created pull request, commit e733f298
atomgit-bot
7 days ago Comments:
7 days ago Comments:
变更摘要
本 PR 为文档更新,在 gmm 目录下新增两篇融合 Pass 说明文档,分别对应 GroupedMatmulFinalizeRoutingTransposeFusionPass 与 GroupedMatmulSwigluQuantTransposeFusionPass。两篇文档结构一致,均说明了该融合 Pass 支持融合 Transpose、Reshape 节点的六种模式(以配图形式给出),并写明使用约束和支持的昇腾型号,用于向用户说明这两个融合 Pass 的生效形态与适用范围。
主要改动
- 新增
GroupedMatmulFinalizeRoutingTransposeFusionPass文档: 新增GroupedMatmulFinalizeRoutingTransposeFusionPass.md,以“融合模式”章节列出该 Pass 支持的六种融合Transpose、Reshape节点的模式,并逐条引用对应示意图。 - 新增
GroupedMatmulSwigluQuantTransposeFusionPass文档: 新增GroupedMatmulSwigluQuantTransposeFusionPass.md,同样以六种融合模式说明该 Pass 的融合形态,并引用同名系列的配图。 - 明确使用约束差异:
GroupedMatmulFinalizeRoutingTransposeFusionPass的使用约束限定为仅当权重(w)数据格式为 ND 时生效;GroupedMatmulSwigluQuantTransposeFusionPass的使用约束标注为无。 - 声明支持的硬件型号: 两篇文档均以
<!-- npu="950" id1 -->标记块声明支持Ascend 950PR/Ascend 950DT。 - 新增融合模式示意图资源: 文档通过相对路径引用
docs/zh/figures/下的GroupedMatmulFinalizeRoutingTransposeFusionPass_1.png至_6.png与GroupedMatmulSwigluQuantTransposeFusionPass_1.png至_6.png系列图片,用于呈现六种融合模式。
🧩 本次 Review 已使用自定义 Skill:


Incorrect?
7 days ago Added label: cann-cla/yes
CANN-robot
7 days ago Comments:
7 days ago Comments:
Thanks for your pull-request.
The full list of commands accepted by me can be found at here.
You can get sig-info at here.
You can self-configure the PR merge rules for this repository. For more details, please refer to Here.
For more, you also can visit HICANN.
PR Approval Progress
✅ Congratulations! All modules have met the lgtm and approve requirements.
Module Approval Details
| module | lgtm status | approve status |
|---|---|---|
| docs | ✅ 宋恺, wang-minbo (2/2) | ✅ 宋恺, wang-minbo (2/1) |
| gmm | ✅ 陈康, kknan, wuyi_huawei (3/2) | ✅ 陈康 (1/1) |
💡 Tip:
- Committer can comment
/approveor/lgtm- Commenting
/approveimplies both code review (lgtm) and intent to merge (approve)
CLA Signature Pass
zhoushaolong, thanks for your pull request. All authors of the commits have signed the CLA. 👍


7 days ago Set yangziqi_007,cheng_zhiwei,chenkang30,juyangokok,monologue815,chaotang233,wuyi_huawei,chen-linxin4,yu-xinjie62,jiang-lirui,Allan_Yu,wang-minbo,xtqh,songkai111,liudan12,Wei_NaChuan,zhangshuoleiokok,shi-rui,cc-z,kdy18482276080,kknan,wenxin_fight,mabing1118,Julius47,zzzzzzyf_ as approver
7 days ago Set yangziqi_007,cheng_zhiwei,chenkang30,juyangokok,monologue815,chaotang233,wuyi_huawei,chen-linxin4,yu-xinjie62,jiang-lirui,Allan_Yu,wang-minbo,xtqh,songkai111,liudan12,Wei_NaChuan,zhangshuoleiokok,shi-rui,cc-z,kdy18482276080,kknan,wenxin_fight,mabing1118,Julius47,zzzzzzyf_ as approver
88 messages are collapsed here。 See more
4 days ago Added label: lgtm
chen-kang30
4 days ago Comments:
4 days ago Comments:
/lgtm
/approve


4 days ago Added label: approved
4 days ago Closed associated issue
4 days ago Merged pull request
描述
补充三个既有 GMM 转置融合 Pass 的用户指南与模式配图,便于了解可融合结构、约束和适用型号;本 PR 不新增或修改融合 Pass 实现。
GroupedMatmulFinalizeRoutingTransposeFusionPass.mdGroupedMatmulSwigluQuantTransposeFusionPass.mdQuantGroupedMatmulInplaceAddTransposeFusionPass.md关联的Issue
关联 Issue #7064:https://gitcode.com/cann/ops-transformer/issues/7064
测试
不涉及
文档更新
新增以下指南及
docs/zh/figures/中对应配图:gmm/grouped_matmul_finalize_routing/docs/GroupedMatmulFinalizeRoutingTransposeFusionPass.mdgmm/grouped_matmul_swiglu_quant_v2/docs/GroupedMatmulSwigluQuantTransposeFusionPass.mdgmm/quant_grouped_matmul_inplace_add/docs/QuantGroupedMatmulInplaceAddTransposeFusionPass.md类型标签