torch:transpose

(torch:transpose a &optional axes)

Differentiable transpose: with no axes the matrix transpose (a vector passes through, like linalg:transpose); with an axes list the rank-n permutation (out-dims[k] = dims[axes[k]], a negative axis counting from the end). The backward pass applies the inverse permutation to the gradient. The matrix transpose and an axes list that exchanges exactly the last two axes ('(0 2 1) on a stack) return a view: no copy is made, torch:matmul reads the source in place and sends the gradient straight to it, and any other reader materializes the transpose once.