ttnn.sub_bw

ttnn.sub_bw(grad_tensor: ComplexTensor or ttnn.Tensor, input_tensor_a: ComplexTensor or ttnn.Tensor, input_tensor_b: ComplexTensor or ttnn.Tensor or Number, *, are_required_outputs: List[bool] = [True, True], memory_config: ttnn.MemoryConfig = None, input_grad: ttnn.Tensor = None, other_grad: ttnn.Tensor = None) → None

Performs backward operations for subtract of input_tensor_a and input_tensor_b or scalar with given grad_tensor. Supports broadcasting.

Parameters:

grad_tensor (ComplexTensor or ttnn.Tensor) – the input gradient tensor.
input_tensor_a (ComplexTensor or ttnn.Tensor) – the input tensor.
input_tensor_b (ComplexTensor or ttnn.Tensor or Number) – the input tensor.

Keyword Arguments:

are_required_outputs (List[bool], optional) – List of required outputs. Defaults to [True, True].
memory_config (ttnn.MemoryConfig, optional) – Memory configuration for the operation. Defaults to None.
input_grad (ttnn.Tensor, optional) – Preallocated output tensor for gradient of input_tensor_a. Defaults to None.
other_grad (ttnn.Tensor, optional) – Preallocated output tensor for gradient of input_tensor_b. Defaults to None.

Note

Supported dtypes, layouts, and ranks:

Dtypes	Layouts	Ranks
BFLOAT16, BFLOAT8_B	TILE	2, 3, 4

bfloat8_b/bfloat4_b is only supported on TILE_LAYOUT

Example

# Create gradient and input tensors for subtraction backward
grad_tensor = ttnn.from_torch(
    torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16), layout=ttnn.TILE_LAYOUT, device=device
)
tensor1 = ttnn.from_torch(
    torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16, requires_grad=True), layout=ttnn.TILE_LAYOUT, device=device
)
tensor2 = ttnn.from_torch(
    torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16, requires_grad=True), layout=ttnn.TILE_LAYOUT, device=device
)

# Compute gradients for subtraction operation
output = ttnn.sub_bw(grad_tensor, tensor1, tensor2)
logger.info(f"Subtraction backward result: {output}")

# Compute gradients for subtraction with scalar
grad_tensor = ttnn.from_torch(
    torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16), layout=ttnn.TILE_LAYOUT, device=device
)
tensor1 = ttnn.from_torch(
    torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16, requires_grad=True), layout=ttnn.TILE_LAYOUT, device=device
)
scalar = 2

# Compute gradients for subtraction operation with scalar
output = ttnn.sub_bw(grad_tensor, tensor1, scalar)
logger.info(f"Subtraction backward with scalar result: {output}")