// retrieving shapes of tensors they wrap until the custom device tensor's
  // shape is explicitly requested where possible.
  int (*num_dims)(void* data, TF_Status* status);

  // Computes the axis length at `dim_index`.
  int64_t (*dim)(void* data, int dim_index, TF_Status* status);

  void (*deallocator)(void* data);

  // Summarizes the value of this tensor. The caller takes ownership of the

TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "Min", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, input, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, axis, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

      } else {
        // Generalize differing axis lengths to "variable"/"unknown".
        for (int axis_index = 0; axis_index < combined_shape.dims();
             ++axis_index) {
          int64_t axis_length = combined_shape.dim_size(axis_index);
          if (axis_length != component_shape.dim_size(axis_index)) {
            axis_length = -1;
          }
          TF_RETURN_IF_ERROR(

        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This
        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,