"//tensorflow/core/common_runtime/eager:tensor_handle",
            "//tensorflow/core/profiler/lib:traceme",
            "@com_google_absl//absl/algorithm:container",
            "@com_google_absl//absl/types:span",
            "@com_google_absl//absl/types:variant",
        ],
    }) + [
        ":abstract_tensor_handle",
        "//tensorflow/c:c_api_macros_hdrs",
        "//tensorflow/core:gpu_runtime",

              unwrapped_handle);
      tensorflow::ImmediateExecutionTensorHandle* result;
      status->status = custom_device_handle->device()->Pack(
          absl::Span<tensorflow::ImmediateExecutionTensorHandle*>(
              tensorflow::unwrap(handles), *num_handles),
          &result);
      return tensorflow::wrap(result);
    }
    tensor_handles.push_back(

        retvals[i] = tensorflow::unwrap(outputs[i]);
        retvals[i]->Ref();
        TFE_DeleteTensorHandle(outputs[i]);
      }
    }
    return status.status;
  }

  tensorflow::Status Pack(absl::Span<ImmediateExecutionTensorHandle*> handles,
                          ImmediateExecutionTensorHandle** result) override {
    TF_Status status;
    *result = tensorflow::unwrap(device_.pack(context_,

    *   Add support for TensorArrays to `tf.data Dataset`.
    *   Integrate Ragged Tensors with `tf.data`.
    *   All core and experimental tf.data transformations that input
        user-defined functions can span multiple devices now.
    *   Extending the TF 2.0 support for `shuffle(...,
        reshuffle_each_iteration=True)` and `cache()` to work across different
        Python iterators for the same dataset.

    TF_DeleteTensor(numerical_tensor);
    delete[] danalytical;
    delete[] dnumerical;
    outputs[i]->Unref();
  }
}

void CheckTensorValue(AbstractTensorHandle* t, absl::Span<const float> manuals,
                      absl::Span<const int64_t> dims, double abs_error) {
  TF_Tensor* analytical_tensor;
  auto s = GetValue(t, &analytical_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

  Status ComputeGradient(
      const VSpace<Gradient, BackwardFunction, TapeTensor>& vspace,
      const absl::Span<const int64_t> target_tensor_ids,
      const absl::Span<const int64_t> source_tensor_ids,
      const std::unordered_map<int64, TapeTensor>& sources_that_are_targets,
      gtl::ArraySlice<Gradient*> output_gradients, absl::Span<Gradient*> result,
      bool build_default_zeros_grads = true);

  }
};

class PassThroughGradientFunction : public GradientFunction {
 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    CHECK_EQ(grad_outputs.size(), 1);
    CHECK_EQ(grad_inputs.size(), 1);
    grad_inputs[0] = grad_outputs[0];
    if (grad_inputs[0]) {

namespace tensorflow {
namespace gradients {
namespace {
class IdentityNGradientFunction : public GradientFunction {
 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];

        "//tensorflow/c/experimental/ops:math_ops",
        "//tensorflow/c/experimental/ops:nn_ops",
        "//tensorflow/core/lib/llvm_rtti",
        "//tensorflow/core/platform:errors",
        "@com_google_absl//absl/types:span",
    ],
)

cc_library(
    name = "not_differentiable",
    srcs = ["not_differentiable.cc"],
    hdrs = [
        "not_differentiable.h",
    ],
    visibility = [
        "//tensorflow:internal",

build --define framework_shared_object=true
build --define tsl_protobuf_header_only=true

build --define=use_fast_cpp_protos=true
build --define=allow_oversize_protos=true

build --spawn_strategy=standalone
build -c opt

# Make Bazel print out all options from rc files.
build --announce_rc

# TODO(mihaimaruseac): Document this option or remove if no longer needed
build --define=grpc_no_ares=true