// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output
// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of

  TF_DeleteBuffer(op_list_buf);
}

class DummyKernel : public tensorflow::OpKernel {
 public:
  explicit DummyKernel(tensorflow::OpKernelConstruction* context)
      : OpKernel(context) {}
  void Compute(tensorflow::OpKernelContext* context) override {}
};

// Test we can query kernels
REGISTER_OP("TestOpWithSingleKernel")
    .Input("a: float")
    .Input("b: float")
    .Output("o: float");

                 TF_Message(first_bad_status.get()));
    return result;
  }
  // For each output of the original operation, pack the per-device
  // TensorHandles we've computed into a single parallel TensorHandle.
  std::vector<std::unique_ptr<ParallelTensor>> per_device_outputs;
  per_device_outputs.reserve(first_op_output_count);
  for (int i = 0; i < first_op_output_count; ++i) {

  TF_Status* s = TF_NewStatus();
  TF_OpKernelConstruction_GetAttrType(ctx, "foobar", &type, s);
  TF_DeleteStatus(s);
  return NULL;
}

// A compute function. This will never actually get called in this test, it's
// just nice to know that it compiles.
void compute(void* kernel, TF_OpKernelContext* ctx) {
  TF_Tensor* input;
  TF_Status* s = TF_NewStatus();
  TF_GetInput(ctx, 0, &input, s);
  TF_DeleteTensor(input);

      : forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

// Struct to be filled in to define a custom device tensor handle. Fields are
// required except where indicated.
typedef struct TFE_CustomDeviceTensorHandleMethods {
  int version = TFE_CUSTOM_DEVICE_VERSION;

  // Computes the rank of the tensor handle.
  //
  // Shapes are specified via callbacks because retrieving the shape of a tensor
  // is a blocking operation for async eager; custom devices should avoid

*   TF Core:

    *   Corrected higher-order gradients of control flow constructs (`tf.cond`,
        `tf.while_loop`, and compositions like `tf.foldl`) computed with
        `tf.GradientTape` inside a `tf.function`.
    *   Changed the default step size in `gradient_checker_v2.compute_gradients`
        to be exactly representable as a binary floating point numbers. This

        "If the intent is to treat this op as non-differentiable consider "
        "using RegisterNotDifferentiable or "
        "NotDifferentiableGradientFunction.");
  }
  return gradient_function->Compute(ctx_, output_gradients, result);
}

Status TapeVSpace::BuildOnesLike(const TapeTensor& t,
                                 AbstractTensorHandle** result) const {
  AbstractOperationPtr op(ctx_->CreateOperation());

        "Expected numeric or variant tensor, got dtype ", dtype);
  }
  return tensorflow::ops::Conj(ctx, input, output, name);
}

class AddGradientFunction : public GradientFunction {
 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // TODO(b/161805092): Support broadcasting.

*   When you contribute a new feature to TensorFlow, the maintenance burden is
    (by default) transferred to the TensorFlow team. This means that the benefit
    of the contribution must be compared against the cost of maintaining the
    feature.
*   Full new features (e.g., a new op implementing a cutting-edge algorithm)
    typically will live in