* Copyright (c) 2025-2026 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 dynamic_mx_quant_proto.h
* \brief
*/
#ifndef QUANT_DYNAMIC_MX_QUANT_PROTO_H_
#define QUANT_DYNAMIC_MX_QUANT_PROTO_H_
#include "graph/operator_reg.h"
namespace ge {
* @brief Performs dynamic MX quantization on input tensor.
* Quantizes the input tensor along the specified axis using block-wise scaling factors.
* Supports various 4-bit and 8-bit floating point output formats.
* @par Inputs:
* @li x: An input tensor of type float16, bfloat16 or float.
* The shape supports at least 1 dimensions, and at most 7 dimensions.
* @par Attributes:
* @li axis: An optional int. Axis along which to quantize. Defaults to -1.
* must be in the range [-rank(input x), rank(input x)).
* @li round_mode: An optional string. Defaults to "rint".
* @li dst_type: An optional int. Declare the output y dtype. Support FLOAT4_E2M1, FLOAT4_E1M2,
* FLOAT8_E4M3FN or FLOAT8_E5M2. Defaults to FLOAT4_E2M1.
* @li blocksize: An optional int. Block size for quantization scaling factors.Defaults to 32.
* When scale_alg is 2, blocksize must be 32.
* @li scale_alg: An optional int.The algorithm for the scale in quantization.Default to 0.
* Support MxFP8/MxFP4(OCP Microscaling Formats (Mx) Specification , count 0) or MxFP8(nvidia-cuBLAS , count 1) or
MxFP4(Dynamic Dtype Range , count 2).
* @li dst_type_max: An optional Float.Max_dtype takes the maximum value of the quant_data_type, or the provided
value.Defaults to 0.
* Only support in FP4_E2M1 mode, with a valid range of 6.0 to 12.0.
* @li max_low_bound: An optional Float, indicates the maximum value limit for scale calculation. Defaults to 0. Only
support when scale_alg=1. Must be non-negative.
* @par Outputs:
* @li y: Quantized output tensor. It has the same shape and rank as input x.
* @li mxscale: An output tensor of type FLOAT8_E8M0. Shape needs to meet the following conditions: \n
* - rank(mxscale) = rank(x) + 1.
* - axis_change = axis if axis >= 0 else axis + rank(x).
* - mxscale.shape[axis_change] = (ceil(x.shape[axis] / blocksize) + 2 - 1) / 2.
* - mxscale.shape[rank(x)] = 2.
* - Other dimensions match input x.
* mxscale tensor is padded with zeros to ensure its size along the quantized axis is even.
* @attention Constraints:
* @li When dst_type is DT_FLOAT8_E5M2 or DT_FLOAT8_E4M3FN, round_mode only supports "rint".
* @li When dst_type is DT_FLOAT4_E2M1 or DT_FLOAT4_E1M2, round_mode supports "rint", "floor" and "round".
* @li If dst_type is DT_FLOAT4_E2M1 or DT_FLOAT4_E1M2, the input x last dimension of the shape must be divisible by 2.
* @li The blocksize must be a multiple of 32 (non-zero) and ≤ 1024.
* When scale_alg is 2, blocksize must be 32.
* @li The value of dst_max_value only supports 0.0 or 6.0-12.0 and is effective when scale_alg=2.
* The default value 0.0 means that maxType corresponds to the maximum value of the target data type.
* If other values are provided, mxscale is calculated based on the provided value.
* @li When x's data type is float, blocksize must be 32.
* @li When x's data type is float, the size of the quantization axis must not be less than 32.
* @par Third-party framework compatibility
* It is a custom operator. It has no corresponding operator in Caffe, ONNX, TensorFlow, or PyTorch.
*/
REG_OP(DynamicMxQuant)
.INPUT(x, TensorType({DT_FLOAT16, DT_BF16, DT_FLOAT}))
.OUTPUT(y, TensorType({DT_FLOAT4_E2M1, DT_FLOAT4_E1M2, DT_FLOAT6_E3M2, DT_FLOAT6_E2M3, DT_FLOAT8_E4M3FN,
DT_FLOAT8_E5M2}))
.OUTPUT(mxscale, TensorType({DT_FLOAT8_E8M0}))
.ATTR(axis, Int, -1)
.ATTR(round_mode, String, "rint")
.ATTR(dst_type, Int, DT_FLOAT4_E2M1)
.ATTR(blocksize, Int, 32)
.ATTR(scale_alg, Int, 0)
.ATTR(dst_type_max, Float, 0.0)
.ATTR(max_low_bound, Float, 0.0)
.OP_END_FACTORY_REG(DynamicMxQuant)
}
#endif
