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,

                                    vector<AbstractTensorHandle*> f_outputs)
      : forward_inputs_(f_inputs), forward_outputs_(f_outputs) {
    for (auto input : forward_inputs_) {
      if (input) {
        input->Ref();
      }
    }
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

//   }
//   ~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:

            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) {

      ASSERT_NEAR(manuals[j], danalytical[j], abs_error);
    }
  }

  TF_DeleteTensor(analytical_tensor);
  delete[] danalytical;
}

Model BuildGradModel(Model forward, GradientRegistry registry) {
  return [forward_model = std::move(forward),
          grad_registry = std::move(registry)](
             AbstractContext* ctx,
             absl::Span<AbstractTensorHandle* const> inputs,

  // For LLVM style RTTI.
  static bool classof(const AbstractOperation* ptr) {
    return ptr->getKind() == kTape;
  }
  ~TapeOperation() override;

 private:
  AbstractOperation* parent_op_;
  ForwardOperation forward_op_;
  Tape* tape_;
  const GradientRegistry& registry_;
};

}  // namespace gradients
}  // namespace tensorflow

      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_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 immediate_execution_ctx_;
  GradientRegistry registry_;

# 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.