Input and Output

ReLU Activation Type

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float/bf16	ND	Input
output	[dim_0, dim_1, ..., dim_n]	float/bf16	ND	Output Same as the input dimension, type, and format.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float/bf16

Input

output

[dim_0, dim_1, ..., dim_n]

float/bf16

Output

Same as the input dimension, type, and format.

FAST_GELU Activation Type

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Input
output	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Output Same as the input dimension, type, and format.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float16/bf16

Input

output

[dim_0, dim_1, ..., dim_n]

float16/bf16

Output

Same as the input dimension, type, and format.

LOG Activation Type

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Input
output	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Output Same as the input dimension, type, and format.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Input

output

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Output

Same as the input dimension, type, and format.

SIGMOID Activation Type

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Input
output	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Output Same as the input dimension, type, and format.
Only the float16 data type is supported on Atlas 200I/500 A2 inference products.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float16/bf16

Input

output

[dim_0, dim_1, ..., dim_n]

float16/bf16

Output

Same as the input dimension, type, and format.

Only the float16 data type is supported on Atlas 200I/500 A2 inference products.

SWISH Activation Type

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Input
output	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Output Same as the input dimension, type, and format.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float16/bf16

Input

output

[dim_0, dim_1, ..., dim_n]

float16/bf16

Output

Same as the input dimension, type, and format.

GELU Activation Type

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Input
output	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Output Same as the input dimension, type, and format.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Input

output

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Output

Same as the input dimension, type, and format.

SWIGLU_FORWARD Activation Type

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Input When activationType is ACTIVATION_SWIGLU_FORWARD, the value of the last dimension must be a multiple of 32 for the Atlas inference products.
output	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Output Has the same dimension, type, and format as the input. The size of dim is half of the input size.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Input

When activationType is ACTIVATION_SWIGLU_FORWARD, the value of the last dimension must be a multiple of 32 for the Atlas inference products.

output

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Output

Has the same dimension, type, and format as the input. The size of dim is half of the input size.

SWIGLU_BACKWARD Activation Type

Parameter	Dimension	Data Type	Format	Description
x1	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Input If activationType is set to ACTIVATION_SWIGLU_BACKWARD, the Atlas inference products is not supported.
x2	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Input The dimension, type, and format are the same as those of x1. The size of dim is twice that of x1.
output	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Output The dimension, type, and format are the same as those of x1. The size of dim is twice that of x1.

Parameter

Dimension

Data Type

Format

Description

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Input

If activationType is set to ACTIVATION_SWIGLU_BACKWARD, the Atlas inference products is not supported.

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Input

The dimension, type, and format are the same as those of x1. The size of dim is twice that of x1.

output

[dim_0, dim_1, ..., dim_n]

float16/float/bf16

Output

The dimension, type, and format are the same as those of x1. The size of dim is twice that of x1.

FASTER_GELU_FORWARD Activation Type

Parameter	Dimension	Data Type	Format	Description
input	[dim_0, dim_1, ..., dim_n]	float/float16/bf16	ND	Input tensor
output	[dim_0, dim_1, ..., dim_n]	float/float16/bf16	ND	Output tensor. Has the same dimension, data type, and data format as the input tensor.

Parent topic: ActivationOperation