Concatenate a sequence of tensors along an existing axis.

All inputs must share dtype, device, and rank; every dimension except axis must agree exactly. The output reuses freshly allocated storage — no input buffer is aliased — and the autograd node saves only per-input shapes so backward can split the gradient back into one slice per input.

Math

$y_{[\ldots,\, k,\, \ldots]} = x^{(i)}_{[\ldots,\, k - \sum_{j<i} N_j,\, \ldots]}$ where $i$ is the unique input index satisfying $\sum_{j<i} N_j \le k < \sum_{j \le i} N_j$.

Shape

Inputs of shape (D_0, ..., D_{axis-1}, N_i, D_{axis+1}, ..., D_{ndim-1}) produce output of shape (D_0, ..., D_{axis-1}, \sum_i N_i, D_{axis+1}, ..., D_{ndim-1}).

See Also

stack_op, split_op, hstack_op, vstack_op.