Transpose

The transpose operation permutes the dimensions of a tensor according to a permutation vector. For each output axis i, the value permutation[i] is the source axis taken from the input (so output_dim[i] = input_dim[permutation[i]], and strides are permuted the same way).

On supported cuDNN versions, the graph lowers to the backend transpose operation (CUDNN_BACKEND_OPERATION_TRANSPOSE_DESCRIPTOR).

Requirements

• Native backend transpose support requires cuDNN 9.22.0 or newer. Older toolkits may report GRAPH_NOT_SUPPORTED for transpose nodes.

C++ API

std::shared_ptr<Tensor_attributes>
transpose(std::shared_ptr<Tensor_attributes> input, Transpose_attributes attributes);

Transpose_attributes provides:

Transpose_attributes&
set_permutation(std::vector<int64_t> const& value);
Transpose_attributes&
set_name(std::string const& value);
Transpose_attributes&
set_compute_data_type(DataType_t value);  // optional; types usually follow the input

The permutation must have length equal to the input rank, contain each index 0 .. rank-1 exactly once, and describe where each output axis reads from in the input.

Python API

• pygraph.transpose(input, permutation, name="")
  • input: source tensor.
  • permutation: list of ints, same semantics as C++.
  • name: optional node name.

Example (C++)

auto X = graph.tensor(/* ... */);  // e.g. dims {2, 4}, row-major strides
auto Y = graph.transpose(
    X,
    cudnn_frontend::graph::Transpose_attributes()
        .set_name("t")
        .set_permutation({1, 0}));  // dims become {4, 2}
Y->set_output(true);