ZConfusionSoftmaxGradFusionPass

Description

Fuses the Mul, ReduceSumD/ReduceSum (supported only by Ascend 950PR/Ascend 950DT), and Sub operators that match the pattern into the ConfusionSoftmaxGrad operator.

Constraints

• The fusion pattern does not take effect for dynamic shapes.
• The fusion pattern does not take effect when the two inputs of the Mul operator have different shapes.
• The fusion pattern does not take effect when the input0 of the Mul operator and the input0 of the Sub operator are from different nodes.
• The fusion pattern does not take effect when input0 is not the first input of the Sub operator.
• The fusion pattern does not take effect when the axis attribute of ReduceSumD/ReduceSum does not correspond to the tail axis index of the input shape.
• The fusion pattern does not take effect when the last axis of the input shape for ReduceSumD/ReduceSum is greater than 30000.
• The fusion pattern does not take effect when the data type is not supported by the fused operator.
• The data type and format of input0 must be the same as those of input1.
• The fusion pattern does not take effect if the reduce axis is set to 1 and the AReduceSumFusionPass fusion pattern is enabled.
• For Ascend 950PR/Ascend 950DT, if keep_dims of ReduceSum is False, this fusion pattern does not take effect.
• For Ascend 950PR/Ascend 950DT, if the input axis of ReduceSum is not a const node, this fusion pattern does not take effect.

Applicable Products

Atlas A2 training products/Atlas A2 inference products

Ascend 950PR/Ascend 950DT