SoftmaxGradExtFusion

Description

Fuses the Mul, ReduceSumD, and Sub operators in the graph on the left that fit the graph fusion pattern into the SoftmaxGradExt operator in the graph on the right.

Restrictions

• Input restrictions:
  • The input of the Mul_1 node shares the same input0 as the first input of the Sub node.
  • The input of the Mul_1 node shares input1 with the input of the Mul_2 node.
  • The axis parameter of ReduceSumD must be -1, and the keep_dims parameter must be True.
• Restrictions on the data format and shape:
  • The shapes of input0 and input1 must fall within the range of 4D to 8D.
  • The last two dimensions of the shapes of input0 and input1 must be 16, and the data format must be NZ.
  • The type of input2 must be scalar.
• The dynamic shape scenario is not supported.
• Restrictions on data types:
  • The data types of input0, input1, and input2 must be the same.
  • Atlas A2 training product/Atlas A2 inference product: The data type can be FLOAT16, FLOAT32, or BFLOAT16.

Availability

Atlas A2 training product/Atlas A2 inference product