* Copyright (c) 2026 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file nn_norm_ops.h
* \brief
*/
#ifndef OPS_NORM_ADA_LAYER_NORM_GRAD_PROTO_H_
#define OPS_NORM_ADA_LAYER_NORM_GRAD_PROTO_H_
#include "graph/operator_reg.h"
namespace ge {
* @brief AdaLayerNormGrad operator interface implementation \n
* @code{.c}
* Forward: out = LayerNorm(x) * (1 + scale) + shift \n
* Backward calculations: \n
* z = (x - mean) * rstd \n
* dy_g = dy * gamma * (1 + scale) \n
* temp_1 = 1/N * ∑(dy * gamma * (1 + scale)) \n
* temp_2 = 1/N * (x - mean) * rstd * ∑(dy * gamma * (1 + scale) * (x - mean) * rstd) \n
* pd_x = (dy * gamma * (1 + scale) - (temp_1 + temp_2)) * rstd \n
* pd_scale = ∑(dy * ((x - mean) * rstd * gamma + beta)) \n
* pd_shift = ∑dy \n
* pd_gamma = ∑dy * (1 + scale) * (x - mean) * rstd \n
* pd_beta = ∑dy * (1 + scale)
* @endcode
* @par Inputs:
* Seven inputs, including:
* @li dy: A tensor. The gradient tensor that represents the reverse calculation.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* The shape is [B, S, H], where B supports 0-6 dimensions.
* @li x: A tensor. First input of forward propagation.
* Must be one of the following types: float16, float32, bfloat16.
* The shape is the same as dy, which is [B, S, H], where B supports 0-6 dimensions.
* @li rstd: A tensor. Third output of forward propagation, indicates the reciprocal of the standard deviation of x.
* Must be one of the following types: float32. The format must be ND.
* Has the shape [B, S, 1], where the last dimension is fixed to 1.
* @li mean: A tensor. Second output of forward propagation, indicates the mean value of x.
* Must be one of the following types: float32. The format must be ND.
* Has the shape [B, S, 1], where the last dimension is fixed to 1.
* @li scale: A tensor. Indicates the adaptive scale parameter.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* The shape is [B, H] or [B, 1, H].
* @li gamma: A tensor. Indicates the normalization weight parameter.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* The shape is [H].
* @li beta: A tensor. Indicates the normalization bias parameter.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* The shape is [H].
* @par Outputs:
* Five outputs, including:
* @li pd_x: A tensor. Indicates the gradient of input x.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* Has the same type, shape and format as x.
* @li pd_scale: A tensor. Indicates the gradient of scale.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* Has the same type, shape and format as scale.
* @li pd_shift: A tensor. Indicates the gradient of adaptive offset parameter.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* Has the same type and format as scale. The shape is [B, H] or [B, 1, H], where B supports 0-6 dimensions.
* @li pd_gamma: A tensor. Indicates the gradient of gamma.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* Has the same type, shape and format as gamma.
* @li pd_beta: A tensor. Indicates the gradient of beta.
* Must be one of the following types: float16, float32, bfloat16. The format must be ND.
* Has the same type, shape and format as beta.
*/
REG_OP(AdaLayerNormGrad)
.INPUT(dy, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.INPUT(x, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.INPUT(rstd, TensorType({DT_FLOAT}))
.INPUT(mean, TensorType({DT_FLOAT}))
.INPUT(scale, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.INPUT(gamma, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.INPUT(beta, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.OUTPUT(pd_x, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.OUTPUT(pd_scale, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.OUTPUT(pd_shift, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.OUTPUT(pd_gamma, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.OUTPUT(pd_beta, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.OP_END_FACTORY_REG(AdaLayerNormGrad)
}
#endif
