//   |      dz|          |
    //   |        |          |
    //   |     Const_3       |
    //   |                   |
    //   |        ^          |
    //   |       z|          |     // MatMul Forward Graph
    //   |        |          |
    //   |      MatMul       |
    //   |     /       \     |
    //   |    ^         ^    |
    //   |    |         |    |
    //   |---x|        y|----|

    your response for more than 2 weeks.

**4. Approved**

-   Once the PR is approved, it gets `kokoro:force-run` label applied and it
    initiates CI/CD tests.
-   We can't move forward if these tests fail.
-   In such situations, we may request you to make further changes to your PR
    for the tests to pass.
-   Once the tests pass, we now bring all the code into the internal code base,

            backwards. See PR
            [47925](https://github.com/tensorflow/tensorflow/pull/47925).
        *   `tf.image.crop_and_resize` gradient w.r.t. either `image` or
            `boxes`. See PR
            [48905](https://github.com/tensorflow/tensorflow/pull/48905).
        *   `tf.sparse.sparse_dense_matmul` forwards. See PR

    forward_grads.resize(output_tensors.size());
    TF_RETURN_IF_ERROR(ForwardpropFromTape(
        op_type, output_tensors, backward_function_getter,
        backward_function_deleter, in_grads, absl::MakeSpan(forward_grads)));
  } else {
    TF_RETURN_IF_ERROR(
        (*forward_function)(in_grads, &forward_grads, use_batch_));
  }
  for (int i = 0; i < forward_grads.size(); ++i) {
    if (forward_grads[i] != nullptr) {

# This is the same as the official TensorFlow builds.
# See https://developer.nvidia.com/cuda-gpus#compute
# `compute_XY` enables PTX embedding in addition to SASS. PTX
# is forward compatible beyond the current compute capability major
# release while SASS is only forward compatible inside the current
# major release. Example: sm_80 kernels can run on sm_89 GPUs but
# not on sm_90 GPUs. compute_80 kernels though can also run on sm_90 GPUs.