implicitly_broadcast_tensors.reserve(inputs.size());  // not tight
  for (const auto& input : inputs) {
    if (absl::holds_alternative<TFE_TensorHandle*>(input)) {
      if (operation_name == std::string("_EagerConst")) {
        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This

// E.g. it could know whether we're in eager mode or graph mode, keeps track
// of gradient tapes, etc.
typedef struct TF_ExecutionContext TF_ExecutionContext;

// A TF_AbstractTensor is an input to an operation. E.g. it could be a union
// type of eager and graph tensors. It is also the result of executing an
// operation.
typedef struct TF_AbstractTensor TF_AbstractTensor;

// invalidate old TF_Operation* pointers.
typedef struct TF_Operation TF_Operation;

// Represents a specific input of an operation.
typedef struct TF_Input {
  TF_Operation* oper;
  int index;  // The index of the input within oper.
} TF_Input;

// Represents a specific output of an operation.
typedef struct TF_Output {
  TF_Operation* oper;

        *   Using `functools.partial` with an invalid `tf.function` input
    *   `tf.function` now enforces input parameter names to be valid Python identifiers. Incompatible names are automatically sanitized similarly to existing SavedModel signature behavior.
    *   Parameterless `tf.function`s are assumed to have an empty `input_signature` instead of an undefined one even if the `input_signature` is unspecified.

    if (grad_inputs[0]) {
      grad_inputs[0]->Ref();
    }
    return absl::OkStatus();
  }
};

// Computes:
//
// @tf.custom_gradient
// def f(input):
//   def grad(grads):
//     return grads[0]
//   return tf.exp(input), grad
// outputs = [f(inputs[0])]
Status ExpWithPassThroughGrad(AbstractContext* ctx,
                              absl::Span<AbstractTensorHandle* const> inputs,

      "      op: 'Add'"
      "      input: 'read0:value:0'"
      "      input: 'read1:value:0'"
      "      attr {"
      "        key: 'T'"
      "        value {"
      "          type: DT_FLOAT"
      "        }"
      "      }"
      "    }"
      "    node_def {"
      "      name: 'add2'"
      "      op: 'Add'"
      "      input: 'add1:z:0'"
      "      input: 'read2:value:0'"
      "      attr {"

    TFE_OpReset(op_.get(), operation_name, device_.c_str(), status);
    if (TF_GetCode(status) != TF_OK) return;
  }
  TFE_OpAddAttrs(op_.get(), attributes);
  for (int input_index = 0; input_index < inputs.size(); ++input_index) {
    TFE_OpAddInput(op_.get(), inputs[input_index], status);
    if (TF_GetCode(status) != TF_OK) return;
  }
  std::vector<TFE_TensorHandle*> unwrapped_results(expected_max_outputs);

  TF_DeleteStatus(status);
  return th;
}

TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "Min", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, input, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, axis, status);

                         int input_index, bool use_function,
                         AbstractTensorHandle** numerical_grad) {
  vector<AbstractTensorHandle*> theta_inputs(inputs.size());
  for (int i{}; i < inputs.size(); ++i) {
    theta_inputs[i] = inputs[i];
  }

  AbstractTensorHandle* theta =
      theta_inputs[input_index];  // parameter we are grad checking

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

// Return a SendOp op `op_name` with send input tensor `in` and attributes
// `send_device`, `recv_device`, and `send_device_incarnation` set.