Merged
fix: 加固INTrainingReduceV2、Relu6Grad和NormalizeBBox #13239
raoliang_sacAdd on 21 days ago
fix: 加固INTrainingReduceV2、Relu6Grad和NormalizeBBox #13239
Merged
Pull Request has been successfully merged by CANN-robot.
(Thank you for raoliang_sac's contribution.)21 days ago Created pull request, commit c7f316d1
21 days ago Modify description of pull request
21 days ago Added label: cann-cla/yes
CANN-robot
21 days ago Comments:
21 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 |
|---|---|---|
| */*/README.md | ✅ 杨露, 陈展熹, 杨阳 (3/2) | ✅ 杨露 (1/1) |
| activation | ✅ 陈展熹, 杨阳 (2/2) | ✅ 陈展熹 (1/1) |
| norm | ✅ 陈展熹, 杨阳 (2/2) | ✅ 陈展熹 (1/1) |
| vfusion | ✅ 陈展熹, 杨阳 (2/2) | ✅ 陈展熹 (1/1) |
💡 Tip:
- Committer can comment
/approveor/lgtm- Commenting
/approveimplies both code review (lgtm) and intent to merge (approve)
CLA Signature Pass
raoliang_sac, thanks for your pull request. All authors of the commits have signed the CLA. 👍


21 days ago Set zengjuan,u010470851,chaotang233,kevin_huang1234,crystalhu,yangyang016,fanqirui,gitee-duhuiping,wangrui_,renruhai,su-yueming,zhajianqing123,yourealize,chenqi317,wang-xing001,liubo75,tangweiwei2,gitcode-chenjiao,chenxingyu18,liujie12345678,east_yang,pingchuantang,Chen_HaoWen,qianzehong,fengying555682,caiwenwen,yanglu-1 as approver
21 days ago Set zengjuan,u010470851,chaotang233,kevin_huang1234,crystalhu,yangyang016,fanqirui,gitee-duhuiping,wangrui_,renruhai,su-yueming,zhajianqing123,yourealize,chenqi317,wang-xing001,liubo75,tangweiwei2,gitcode-chenjiao,chenxingyu18,liujie12345678,east_yang,pingchuantang,Chen_HaoWen,qianzehong,fengying555682,caiwenwen,yanglu-1 as approver
68 messages are collapsed here。 See more
yanglu-1
18 days ago Comments:
18 days ago Comments:
/lgtm
/approve


CANN-robot
18 days ago Comments:
18 days ago Comments:
The following users do not have permission to comment /lgtm or /approve on any module in this PR:
/approve: 杨阳


18 days ago Added label: approved
18 days ago Closed associated issue
18 days ago Merged pull request
描述
合并交付INTrainingReduceV2、Relu6Grad、NormalizeBBox三个算子的质量修复。基于官方master
22de9811f0c5d04e9f879fc4b2fae3dfcc135e12,仅修改这三个算子,共32个文件。734fc4a0f55ee67d657efc8ad79805e91783e37d仅提取activation/relu6_grad下7个文件。InferShape补充rank上限及已知维广播校验;InferDataType检查支持集合和类型一致性;补齐UT及6条精简ST。数值Kernel计算逻辑不变。该PR中的SGD修改不在本次范围内。825bef907433a4a2d75cfce7709dfe1ab950dada,共13个文件。高宽保持FP32,FP16坐标提升后相除并在输出时舍入,修复合法INT32大除数先转FP16溢出的问题;同步UB规划,TilingData布局不变;修正数学/GPU参考,补齐shape、batch和dtype校验、UT及8条精简ST。两个移植算子的完整目录Git tree与来源PR一致,原INTrainingReduceV2目录与
3dbff215a145c0561ff9be221a544bd1898d19c5一致。已保留备份,先squash为单提交并验证同一原始基点下完整binary/mode diff和Git tree不变,再rebase最新官方master;无冲突,最终相对该master只有一个提交。关联的Issue
测试
本轮合并验证:2026-11-03,受测HEAD
9bd2a73e5a8f527d676ad8eaf3c41b002967e430,官方master基线22de9811f0c5d04e9f879fc4b2fae3dfcc135e12。使用任务内独立CANN 9.3.0和指定统一入口/workspace/CI-skills/ci_local_check.sh,仅选择三个算子;双SoC串行执行,构建有效并行度不超过8。本轮没有重新执行TTK/NPU/GPU精度、内存检查或示例。下列为来源PR的历史验收,保留原受测版本和环境;目录内容一致的检查不表述为重新上板:
a996a41594dc9e81430d7f169b0ad42f917d16e9。19条原失败冻结输入三方通过;网站L1 486条完成验证,其中1条CPU参考资源退出后按原尺寸独立复测通过;137条原动态oom失败、24条Kernel L0、24条GEIR L0、88条专项及28条最小/稳定性探针通过。最后Host版本26条正常Kernel、12条GE回归通过;59条元数据边界和14条负测符合预期,单独判定,不计作精度PASS。网页24项及原始53条逐项核对归档。boxes=1000、高宽65536时FP16错误输出0,正确值为0.0152587890625;修正参考后旧Kernel仍失败。修复后26条Kernel用例的dyn/cst/bin、26条GE、4条动态shape/rank及同图多rank复用通过;21条注册回调与13条负向诊断符合预期。最终原文C++示例FP16/FP32编译运行并核对输出,A3存量包4条GE兼容用例通过。来源TTK执行均显式
--pc=1,不放宽精度阈值。本次保留每个算子单份精简ST CSV,分别20、6、8条;复核字段、启用状态、shape规模及无外部输入文件依赖,没有新增重复用例或大shape。本轮本地CI通过的同一SHA已同步至PR源分支,PR回查为单提交。线上CI触发请求已提交并回查成功;等待平台回报,不沿用旧SHA的线上通过状态。
文档更新
保留三个算子的README增量整改,遵守现行《文档修改与示例验证指引》及官方README模板:不删减已核实有效的产品、dtype、format和调用入口,产品差异放在既有约束说明的无序列表中,不新增重复编译运行章节。
本次整合没有额外改写README、示例或示例构建配置。
类型标签
AI/Agent生成声明