Theme by the Executable Book Project

BiasAddBackward¶

General¶

BiasAddBackward operation computes the gradients on the bias tensor for BiasAdd operator. This op accumulates all the values from \(\diffdst\) into the channel dimension, the axis depends on the layout of \(\src\) tensor.

Operation attributes¶

Attribute Name	Description	Value Type	Supported Values	Required or Optional
data_format	Controls how to interpret the shape of `diff_dst` and `diff_bias` .	string	`NCX` , `NXC` (default)	Optional

Execution arguments¶

The inputs and outputs must be provided according to below index order when constructing an operation.

Inputs¶

Index	Argument Name	Required or Optional
0	`diff_dst`	Required

Outputs¶

Index	Argument Name	Required or Optional
0	`diff_bias`	Required

Supported data types¶

BiasAddBackward operation supports the following data type combinations.

Diff_dst	Diff_bias
f32	f32
bf16	bf16
f16	f16

BiasAdd Clamp