//   }
//   ~AddGradientFunction() override {}
// };
//
// GradientFunction* AddRegisterer(const ForwardOperation& op) {
//   // More complex gradient functions can use inputs/attrs etc. from the
//   // forward `op`.
//   return new AddGradientFunction;
// }
//
// Status RegisterGradients(GradientRegistry* registry) {
//   return registry->Register("Add", AddRegisterer);
// }
class GradientFunction {
 public:

      ASSERT_EQ(errors::OK, s.code()) << s.message();
      ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr ctx_;

 public:

#include "tensorflow/core/platform/status.h"

namespace tensorflow {
namespace parallel_device {

using TensorHandlePtr = tensorflow::Safe_TFE_TensorHandlePtr;

class ParallelTensor;
class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use

    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,

#include "tensorflow/c/eager/unified_api_testutil.h"

namespace tensorflow {
namespace gradients {

/* Returns numerical grad inside `dtheta_approx` given `forward` model and
 * parameter specified by `input_index`.
 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *

absl::Status RunAndMaybeSum(AbstractContext* ctx, Model forward,
                            absl::Span<AbstractTensorHandle* const> inputs,
                            absl::Span<AbstractTensorHandle*> outputs,
                            bool use_function) {
  AbstractTensorHandle* model_outputs[1];

  // Run the model.
  TF_RETURN_IF_ERROR(
      RunModel(forward, ctx, inputs, model_outputs, use_function));

    device_threads_.emplace_back(new DeviceThread(
        devices[device_index].c_str(), is_async, in_flight_nodes_limit));
  }
}

// Necessary for a unique_ptr to a forward-declared type.
ParallelDevice::~ParallelDevice() = default;

std::unique_ptr<ParallelTensor> ParallelDevice::CopyToParallelDevice(
    TFE_Context* context, TFE_TensorHandle* tensor, TF_Status* status) const {

# 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.

            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

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