SoftmaxSmallOpFusionPass

Description

Identifies and fuses small operators such as ReduceMaxD and Sub into the SoftmaxV2 operator.

Fusion patterns are disabled by default.

Before: After:

Restrictions

• The dynamic shape scenario is not supported.
• Input parameters of ReduceMaxD and ReduceSumD:
  • The axes parameter must be the same.
  • The value of keep_dims is true.
• The current fusion pattern does not take effect if the reduce axis is set to 1 and AReduceSumFusionPass fusion pattern is enabled.
• The current fusion pattern does not take effect if there is more than one reduce axis and the tail axis is included, the data type is FLOAT32, and ReduceMaxDFusionPass fusion pattern is enabled.
• Data type restrictions:
  • Atlas A2 training product/Atlas A2 inference product: The data type can be FLOAT32, FLOAT16, or BFLOAT16.

Availability

Atlas A2 training product/Atlas A2 inference product