ttnn.gelu_bw

ttnn.gelu_bw(grad_tensor: ttnn.Tensor, input_tensor: ttnn.Tensor, *, variant: ttnn.GeluVariant = ``ttnn.GeluVariant.Accurate``, memory_config: ttnn.MemoryConfig = None, input_grad: ttnn.Tensor = None) → List of ttnn.Tensor

Performs backward operations for GELU on input_tensor, with given grad_tensor and variant.

Parameters:

  • grad_tensor (ttnn.Tensor)the input gradient tensor.

  • input_tensor (ttnn.Tensor)the input tensor.

Keyword Arguments:

  • variant (ttnn.GeluVariant, optional)Select GELU implementation. Defaults to ttnn.GeluVariant.Accurate.

  • memory_config (ttnn.MemoryConfig, optional)memory configuration for the operation. Defaults to None.

  • input_grad (ttnn.Tensor, optional)preallocated output tensor. Defaults to None.

Returns:

List of ttnn.Tensorthe output tensor.

Note

Supported dtypes and layouts:

Dtypes

Layouts

BFLOAT16, FLOAT32, BFLOAT8_B, BFLOAT4_B

TILE

For more details about BFLOAT8_B, refer to the BFLOAT8_B limitations.

Example

# Create sample tensors for backward GELU operation
grad_tensor = ttnn.from_torch(
    torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16), layout=ttnn.TILE_LAYOUT, device=device
)
input_tensor = ttnn.from_torch(
    torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16, requires_grad=True), layout=ttnn.TILE_LAYOUT, device=device
)

# Call the gelu_bw function with approximation method
output = ttnn.gelu_bw(grad_tensor, input_tensor, variant=ttnn.GeluVariant.Accurate)
logger.info(f"GELU Backward: {output}")