Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
conv2d groupOpt preload场景没有对最后一个groupOpt的singleGroups进行刷新,导致最后一块groupIter搬运指令越界,搬多了 算子类型 Conv2D, 数据类型float16 fmap(NCHW):[8, 408, 28, 28] weight(NCHW):[408, 24, 3, 3] group属性:17 stride属性:2 dilation属性:全为1 pad属性:全为0 1 0 1 带bias
Ascend950PR_9589 基于最新包构建
1)按算子规格输入构造input、weight、属性等输入 2)在对应950PR环境(32核)执行相关整网用例(通路可以使用pytorch) 3)开启oom验证
功能正确,精度正常
oom越界报错
Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
Describe the current behavior / 问题描述 (Mandatory / 必填)
conv2d groupOpt preload场景没有对最后一个groupOpt的singleGroups进行刷新,导致最后一块groupIter搬运指令越界,搬多了
算子类型 Conv2D, 数据类型float16
fmap(NCHW):[8, 408, 28, 28]
weight(NCHW):[408, 24, 3, 3]
group属性:17
stride属性:2
dilation属性:全为1
pad属性:全为0 1 0 1
带bias
Environment / 环境信息 (Mandatory / 必填)
Ascend950PR_9589
基于最新包构建
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
1)按算子规格输入构造input、weight、属性等输入
2)在对应950PR环境(32核)执行相关整网用例(通路可以使用pytorch)
3)开启oom验证
Describe the expected behavior / 预期结果 (Mandatory / 必填)
功能正确,精度正常
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
oom越界报错
Special notes for this issue/备注 (Optional / 选填)