* 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 dynamic_block_qunat_proto.h
* \brief
*/
#ifndef OPS_BUILT_IN_OP_PROTO_INC_NN_QUANTIZE_H_
#define OPS_BUILT_IN_OP_PROTO_INC_NN_QUANTIZE_H_
#include "graph/operator_reg.h"
namespace ge {
* @brief Online quantizes the input tensor per block.
* @par Inputs:
- x: A tensor of type float16, bfloat16 or float32. Shape must be 2-dimensional or 3-dimensional.
* @par Attributes:
- min_scale: (Optional) Minimum scale value for quantization. Must be a positive float.
* Defaults to 0.0.
- round_mode: (Optional) Quantization rounding mode. Valid values:
* - "rint": Supported for FLOAT8_E5M2/FLOAT8_E4M3FN
* - "round": Supported for HIFLOAT8 only
* Defaults to "rint".
- dst_type: (Optional) Target data type enum value:
* - 2: INT8
* - 34: HIFLOAT8
* - 35: FLOAT8_E5M2
* - 36: FLOAT8_E4M3FN
* Defaults to 35 (FLOAT8_E5M2).
- row_block_size: (Optional) Number of elements per block in -2 dimension.
* Only support 1, 128, 256, 512. Defaults to 1.
- col_block_size: (Optional) Number of elements per block in -1 dimension.
* Only support 64, 128, 192, 256. Defaults to 128.
- dst_type_max: (Optional) Maximum Value of the target data type.
* Only effective when dst_type is 34(HIFLOAT8), supporting values 0.0, 15.0, 56.0, 224.0, 32768.0. Defaults to 0.0.
* @par Outputs:
- y: Quantized tensor with same shape as input x. Data type depends on dst_type.
- scale: Scale tensor of type float. Shape is [ceil(x.rows/row_block_size), ceil(x.cols/col_block_size)] or [B,
ceil(x.rows/row_block_size), ceil(x.cols/col_block_size)].
* @par Third-party framework compatibility:
* Custom operator with no direct mapping in Caffe/ONNX/TensorFlow/PyTorch.
*/
REG_OP(DynamicBlockQuant)
.INPUT(x, TensorType({DT_FLOAT16, DT_BF16, DT_FLOAT}))
.OUTPUT(y, TensorType({DT_INT8, DT_HIFLOAT8, DT_FLOAT8_E4M3FN, DT_FLOAT8_E5M2}))
.OUTPUT(scale, TensorType({DT_FLOAT}))
.ATTR(min_scale, Float, 0.0)
.ATTR(round_mode, String, "rint")
.ATTR(dst_type, Int, DT_FLOAT8_E5M2)
.ATTR(row_block_size, Int, 1)
.ATTR(col_block_size, Int, 128)
.ATTR(dst_type_max, Float, 0.0)
.OP_END_FACTORY_REG(DynamicBlockQuant)
}
#endif
