* Copyright (c) 2025 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_quantize.h
* \brief
*/
#ifndef OPS_BUILT_IN_OP_PROTO_INC_FLAT_QUANT_H_
#define OPS_BUILT_IN_OP_PROTO_INC_FLAT_QUANT_H_
#include "graph/operator_reg.h"
namespace ge {
* @brief Flatness matters for LLM quantization. \n
*
* @par Inputs:
* @li x: A tensor. The original data input. 3-D with shape [K, M, N]. Must be one of the following types:
* float16, bfloat16. The format support ND.
* @li kronecker_p1: A tensor. Input calculation matrix 1. 2-D with shape [M, M]. The value of M is same as input "x".
* Must be one of the following types: float16, bfloat16. Has the same type as input "x".
* The format support ND.
* @li kronecker_p2: A tensor. Input calculation matrix 2. 2-D with shape [N, N]. The value of N is same as input "x".
* Must be one of the following types: float16, bfloat16. Has the same type as input "x".
* The format support ND. \n
*
* @par Outputs:
* @li out: When the output data type is int4,the shape is consistent with the input parameter x.When the output data
type is int32,the shape is [K,M,N/8].When the output data type is float4_e2m1,the shape is [K,M*N].
* @li quant_scale: When the output data type is int4 or int32,the shape is [K].When the output data type is
float4_e2m1,the shape is [K,ceilDiv(M*N,64),2].
* The values of K,M,and N are same as that of "x". \n
* @par Attributes:
* clip_ratio: An optional float. Used to control the quantization cropping ratio. Defaults to 1. \n
* dst_dtype: An optional int. Used to control the quantization dst_type. Defaults to 29. \n
*/
REG_OP(FlatQuant)
.INPUT(x, TensorType({DT_FLOAT16, DT_BF16}))
.INPUT(kronecker_p1, TensorType({DT_FLOAT16, DT_BF16}))
.INPUT(kronecker_p2, TensorType({DT_FLOAT16, DT_BF16}))
.OUTPUT(out, TensorType({DT_INT4, DT_FLOAT4_E2M1}))
.OUTPUT(quant_scale, TensorType({DT_FLOAT, DT_FLOAT8_E8M0}))
.ATTR(clip_ratio, Float, 1)
.ATTR(dst_dtype, Int, DT_INT4)
.OP_END_FACTORY_REG(FlatQuant)
}
#endif
