support auto generate for layer_norm #52008
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
| func : layer_norm | ||
| data_type : x | ||
| backward : layer_norm_grad | ||
| optional : scale, bias |
There was a problem hiding this comment.
增加variance的intermediate属性，参考send_ue_recv算子
| @@ -847,6 +847,17 @@ | |||
| optional : master_param, skip_update, beta1_pow_out, beta2_pow_out, master_param_outs | |||
| inplace : (param -> param_out), (moment1 -> moment1_out), (moment2 -> moment2_out), (beta1_pow -> beta1_pow_out), (beta2_pow -> beta2_pow_out), (master_param -> master_param_outs) | |||
|
|
|||
| - op : layer_norm | |||
| args : (Tensor x, Tensor scale, Tensor bias, float epsilon, int begin_norm_axis) | |||
There was a problem hiding this comment.
增加epsilon begin_norm_axis的默认值
|
|
||
| // NOTE(jiahongyu): Below codes originally enclosed by PADDLE_WITH_MKLDNN | ||
| int begin_norm_axis = ctx.Attr<int>("begin_norm_axis"); | ||
| if (begin_norm_axis != |
There was a problem hiding this comment.
这个算子的GetExpectedKernelType需要invoke，可以参考算子frobenius_norm
| DECLARE_NO_NEED_BUFFER_VARS_INFERER(LayerNormGradNoNeedBufferVarInferer, | ||
| "Bias"); | ||
|
|
||
| class LayerNormCompositeGradOpMaker : public prim::CompositeGradOpMakerBase { |
There was a problem hiding this comment.
十分抱歉! 这个算子很复杂，暂时不要修改这个，这个算子任务关闭
There was a problem hiding this comment.
十分抱歉! 这个算子很复杂，暂时不要修改这个，这个算子任务关闭
收到~
|
很抱歉，经过我们的反复讨论，你的PR暂未达到合入标准，请阅读飞桨原生算子开发规范，你可以重新提交新的PR，我们先将此PR关闭，感谢你的贡献。 |
PR types
Others
PR changes
Others
Describe
#51842 case8