Closed
[Bug-Report|缺陷反馈]: ApplyCamePart1直接归约及长归约窄列场景的精度与UB规划问题 #8291
zsq_888Add on 16 days agoClosed at 15 days ago
16 days ago Added label: bug-report
16 days ago Associated pull request
16 days ago Assign qq_57796320 as responsible person
15 days ago Close issue
15 days ago Added label: resolved
Describe the current behavior / 问题描述 (Mandatory / 必填)
ApplyCamePart1在arch35路径的直接归约、长归约及窄列场景存在精度风险,UB规划也需要与各阶段同时存活的缓冲区需求保持一致:
相关代码位于optim/apply_came_part1/op_host/arch35、op_kernel/arch35及tests/assets/golden.py。
Environment Information / 环境信息 (Mandatory / 必填)
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
以上为验证步骤。PR未提供修复前的具体失败shape、随机种子、执行命令及误差数值,尚不能据此独立复现某一条失败用例。
Describe the expected behavior / 预期结果 (Mandatory / 必填)
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
PR作者说明已完成算子包编译、安装并在Ascend950上验证,GEIR图模式下三方精度比较的三个输出均为PASS。
本次仅核对PR代码差异,未重新执行编译或上板测试;PR描述未附完整原始日志、失败数值或截图,待作者补充。
Special notes for this issue/备注 (Optional / 选填)
关联PR:https://gitcode.com/cann/ops-nn/pull/13736
修复涉及UB动态规划、直接归约补偿项合并、窄列分块补偿累加及FP64 golden,共10个文件。