* 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.
*/
#ifndef OPS_BUILT_IN_OP_PROTO_INC_GROUPED_QUANT_MAX_OPS_H_
#define OPS_BUILT_IN_OP_PROTO_INC_GROUPED_QUANT_MAX_OPS_H_
#include "graph/operator_reg.h"
namespace ge {
* @brief Quantizes the input tensor x with the given scale and calculates the
* absolute maximum value (amax) of x.
*
* @par Inputs:
* @li x: A required tensor of type FLOAT32, FLOAT16 or BFLOAT16, specifying the input tensor
* to be quantized. The format must be ND. Shape supports 2D to 8D.
* @li scale: A required 1D tensor of type FLOAT32 with shape [num_groups], specifying the
* quantization scaling factor. The format must be ND.
* @li group_list: A required 1D tensor of type INT64 with shape [num_groups], specifying the
* cumulative group boundaries along dim-0. The format must be ND.
*
* @par Attributes:
* @li round_mode: An optional string, specifying the rounding mode for the cast operation.
* Default value: "rint". Valid values depend on dst_type:
* - For FLOAT8_E5M2 (35) or FLOAT8_E4M3FN (36): only "rint" is supported.
* - For HIFLOAT8 (34): "round" or "hybrid" is supported.
* @li dst_type: An optional int64, specifying the output quantized data type.
* Default value: 35 (FLOAT8_E5M2).
* Valid values: 34 (HIFLOAT8), 35 (FLOAT8_E5M2), 36 (FLOAT8_E4M3FN).
*
* @par Outputs:
* @li y: A required tensor of type HIFLOAT8, FLOAT8_E5M2 or FLOAT8_E4M3FN,
* specifying the quantized output. The format must be ND.
* Shape is the same as input "x". Data type is determined by dst_type.
* @li amax: A required 1D tensor of type FLOAT32, FLOAT16 or BFLOAT16 with shape [num_groups],
* specifying the absolute maximum value of input "x".
* Data type must be the same as input "x". The format must be ND.
*
* @par Third-party framework compatibility
* It is a custom operator on Ascend NPU.
*/
REG_OP(GroupedQuantMax)
.INPUT(x, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.INPUT(scale, TensorType({DT_FLOAT}))
.INPUT(group_list, TensorType({DT_INT64}))
.OUTPUT(y, TensorType({DT_HIFLOAT8, DT_FLOAT8_E5M2, DT_FLOAT8_E4M3FN}))
.OUTPUT(amax, TensorType({DT_FLOAT, DT_FLOAT16, DT_BF16}))
.ATTR(round_mode, String, "rint")
.ATTR(dst_type, Int, DT_FLOAT8_E5M2)
.OP_END_FACTORY_REG(GroupedQuantMax)
}
#endif
