norm/layer_norm_v4/tests/assets/golden.py
meanOutOptional
rstdOutOptional
None
aclnnAddLayerNorm
aclnnLayerNorm
aclnnFastLayerNorm
layer_norm_v4/tests/assets/golden.py
norm/add_layer_norm/README.md
norm/inplace_add_layer_norm/README.md
norm/add_layer_norm/docs/aclnnAddLayerNorm.md
[y]
关联 PR:https://gitcode.com/cann/ops-nn/pull/14541
一、问题描述
norm/layer_norm_v4/tests/assets/golden.py的 aclnn golden 在meanOutOptional/rstdOutOptional为 None 时仍返回None占位,导致参考输出列表的长度与位置和算子实际输出槽位不一致,容易引起比对错位。aclnnAddLayerNorm又限制"输入的某一维不能为 0"、"x1 为空时 meanOut/rstdOut 必须为空",与 Ascend 950PR&950DT 支持空入非空出(空归约 mean=0、rstd=NaN)的实际能力不符。二、重现步骤
aclnnLayerNorm/aclnnFastLayerNorm构造仅传部分可选输出(不传 meanOut 和/或 rstdOut)的用例,加载layer_norm_v4/tests/assets/golden.py的 aclnn golden;norm/add_layer_norm/README.md、norm/inplace_add_layer_norm/README.md与norm/add_layer_norm/docs/aclnnAddLayerNorm.md的空 Tensor 约束描述。三、预期结果
[y],存在 mean/rstd 时依次追加,不使用None占位;关联 PR:https://gitcode.com/cann/ops-nn/pull/14541