#!/usr/bin/env python3
# -*- coding: UTF-8 -*-
# ----------------------------------------------------------------------------
# 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.
# ----------------------------------------------------------------------------
__golden__ = {
"kernel": {
"matmul_reduce_scatter": "matmul_reduce_scatter_golden"
}
}
import numpy as np
import torch
from torch.distributed import ReduceOp
def matmul_reduce_scatter_golden(
x1,
x2,
bias=None,
group: str = "",
reduce_op: str = "sum",
is_trans_a: bool = False,
is_trans_b: bool = False,
comm_turn: int = 0,
rank_size: int = 0,
**kwargs
):
"""
Golden 计算函数 for matmul_reduce_scatter
参数名称与 matmul_reduce_scatter_def.cpp 完全一致
计算逻辑:
1. 处理转置
2. 执行 matmul
3. 执行 reduce_scatter
"""
if is_trans_a:
x1 = x1.t()
if is_trans_b:
x2 = x2.t()
if bias is not None:
output = torch.matmul(x1, x2) + bias
else:
output = torch.matmul(x1, x2)
tensor_scatter = kwargs.get('tensor_scatter', None)
if tensor_scatter is None:
output_shape = [x1.shape[0] // rank_size, x2.shape[0] if is_trans_b else x2.shape[1]]
tensor_scatter = torch.zeros(output_shape, dtype=x1.dtype)
dist = kwargs.get('dist', None)
if dist is not None:
dist._reduce_scatter_base(tensor_scatter, output, op=ReduceOp.SUM)
output = tensor_scatter
return output