TransdataTransposeTransdataBatchMatMulv2FusionPass

Description

Deletes redundant nodes such as TransData and Transpose between the two BatchMatMulV2 nodes.

Mode 1

Mode 2

Mode 3

Restrictions

The output data type of the first TransData must be float16.

Every node must have only one successor.

The perm list of Transpose can only be [0,2,1,3], and the last dimension must be exactly divided by 16.

The last three dimensions of the input shape of the first TransData must be the same as those of the output shape of the second TransData.

Availability

Atlas inference product