ttnn.concat_bw

ttnn.concat_bw(grad_tensor: ttnn.Tensor, input_tensor_a: ttnn.Tensor, input_tensor_b: ttnn.Tensor, dim: int = 0, *, are_required_outputs: List[bool] | None = [True, True], memory_config: ttnn.MemoryConfig | None = None, input_grad: ttnn.Tensor | None = None, other_grad: ttnn.Tensor | None = None, queue_id: int | None = 0) → List of ttnn.Tensor

Performs backward operations for concat on input_tensor_a and input_tensor_b with given grad_tensor.

Parameters:

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

  • input_tensor_a (ttnn.Tensor) – the input tensor.

  • input_tensor_b (ttnn.Tensor) – the input tensor.

  • dim (int) – Dimension to concatenate. Defaults to 0.

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.

  • queue_id (int, optional) – command queue id. Defaults to 0.

Returns:

List of ttnn.Tensor – the output tensor.

Note

Supported dtypes, layouts, and ranks:

Dtypes

Layouts

Ranks

BFLOAT16

TILE

4

bfloat8_b/bfloat4_b is only supported on TILE_LAYOUT

Example

>>> grad_tensor = ttnn.from_torch(torch.rand([14, 1, 30, 32], dtype=torch.bfloat16), layout=ttnn.TILE_LAYOUT, device=device)
>>> tensor1 = ttnn.from_torch(torch.rand([12, 1, 30, 32], dtype=torch.bfloat16, requires_grad=True), layout=ttnn.TILE_LAYOUT, device=device)
>>> tensor2 = ttnn.from_torch(torch.rand([2, 1, 30, 32], dtype=torch.bfloat16, requires_grad=True), layout=ttnn.TILE_LAYOUT, device=device)
>>> dim = 0
>>> output = ttnn.concat_bw(grad_tensor, tensor1, tensor2, dim)