Closed
[Bug-Report|缺陷反馈]: INTrainingReduceV2动态oom、极值归约及格式边界异常 #7835
raoliang_sacAdd on 21 days agoClosed at 18 days ago
21 days ago Added label: bug-report
21 days ago Assign raoliang_sac as responsible person
21 days ago Change title to "[Bug-Report|缺陷反馈]: INTrainingReduceV2动态oom、极值归约及格式边界异常", originally titled "[Bug-Report|缺陷反馈]: INTrainingReduceV2动态oom异常及非法格式错误码不准确".
21 days ago Change title to "[Bug-Report|缺陷反馈]: INTrainingReduceV2动态oom、极值归约及格式边界异常", originally titled "[Bug-Report|缺陷反馈]: INTrainingReduceV2动态oom异常及非法格式错误码不准确".
21 days ago Modify description of issue
18 days ago Close issue
18 days ago Added label: resolved
问题描述
INTrainingReduceV2(Ascend 950 / arch35)扫描发现动态oom异常、极值sum精度失败及格式诊断问题。复核还确认原BIN匹配能拦截非法输出格式,直接DYN Tiling缺少对应的Host保护。
FLT_MAX, FLT_MAX, -FLT_MAX, -FLT_MAX应输出sum=0,原实现为NaN。原扫描网页:https://cann.cannbot.cn/harnessmcp-tests/evaluation/validation/upload-20261030191350-intrainingreducev2-25aa800e?tab=testcases
修复
对应PR为400000单独注册ARA结构,加入全R共享二进制缩放并保留AR高/低分量,修正输入格式/零R诊断并补齐两个输出的ND保护;TestSpec保持Torch FP64参考,保留原精度标准。README保留有效能力,补必要说明;新增88条专项生成资产及59条Host元数据边界与预期清单。
推导阶段保留合法动态标记和GE空张量传播;native执行的dtype/format/shape域通过原型、OpDef、匹配、Tiling或不支持类型的编译阶段约束。产品并集及UpdateMask自动递减等误报已按源码/API/实测核实,不为消除告警破坏已有能力。
验证
CANN 9.3.0,Ascend950PR_9579与真实A100,官方TTK a996a41594dc9e81430d7f169b0ad42f917d16e9。TTK统一单进程,构建最多8任务。
已逐项归档网页24项及原始53条映射,无剩余已确认缺陷;不将metadata或预期异常计为精度PASS,不降低阈值。完整冻结输入、日志、分布分析及逐项闭环报告保留在任务交接目录。
关联 PR:https://gitcode.com/cann/ops-nn/pull/13239