bool found_neg = false;
  size_t pos = 0;
  TF_Operation* oper;
  while ((oper = TF_GraphNextOperation(graph, &pos)) != nullptr) {
    if (oper == feed) {
      EXPECT_FALSE(found_placeholder);
      found_placeholder = true;
    } else if (oper == three) {
      EXPECT_FALSE(found_scalar_const);
      found_scalar_const = true;
    } else if (oper == add) {
      EXPECT_FALSE(found_add);
      found_add = true;

const char* TF_OperationName(TF_Operation* oper) {
  return oper->node.name().c_str();
}

const char* TF_OperationOpType(TF_Operation* oper) {
  return oper->node.type_string().c_str();
}

const char* TF_OperationDevice(TF_Operation* oper) {
  return oper->node.requested_device().c_str();
}

int TF_OperationNumOutputs(TF_Operation* oper) {
  return oper->node.num_outputs();
}

this file with a file generated from [`api_template.__init__.py`](https://www.github.com/tensorflow/tensorflow/blob/master/tensorflow/api_template.__init__.py)
"""
# pylint: disable=g-bad-import-order,protected-access,g-import-not-at-top

import distutils as _distutils
import importlib
import inspect as _inspect
import os as _os
import site as _site
import sys as _sys

// through custom device implementations).
//
// TODO(allenl): Currently these are black boxes, but we should have some way to
// inspect values. This would let people e.g. copy over most attributes and then
// modify some based on their values.

// A reference to an op's name -> attribute mapping
typedef struct TFE_OpAttrs TFE_OpAttrs;

semantics to match NumPy. * `tf.split` now takes arguments in a reversed order
and with different keywords. In particular, we now match NumPy order as
`tf.split(value, num_or_size_splits, axis)`. * `tf.sparse_split` now takes
arguments in reversed order and with different keywords. In particular we now
match NumPy order as `tf.sparse_split(sp_input, num_split, axis)`. NOTE: we have

# Do not sort: Copybara rule
# copybara:uncomment_begin
# # buildifier: disable=out-of-order-load
# load("//devtools/build_cleaner/skylark:action_config_test.bzl", "action_config_test")
# load("//devtools/copybara/rules:copybara.bzl", "copybara_config_test")
# load("@rules_license//rules:license.bzl", "license")
# # buildifier: enable=out-of-order-load
# copybara:uncomment_end

# copybara:comment_begin(oss-only)

    return absl::OkStatus();
  }

  tensorflow::FullTypeDef FullType() const override {
    const FullTypeDef* ft;
    mutex_lock l(graph_->mu);
    graph_->graph.NodeType(output_.oper->node.name(), &ft);
    if (ft == nullptr) {
      return FullTypeDef();
    } else {
      return *ft;
    }
  }

  TF_Output output_;

  // For LLVM style RTTI.

            'from distutils.sysconfig import get_python_lib;'
            'print(get_python_lib())'
        ])
    ]

  all_paths = set(python_paths + library_paths)
  # Sort set so order is deterministic
  all_paths = sorted(all_paths)

  paths = []
  for path in all_paths:
    if os.path.isdir(path):
      paths.append(path)
  return paths

          {"SparseSoftmaxCrossEntropyWithLogits", {1}},
          {"FusedBatchNorm", {1, 2, 3, 4}},
      });
  return m;
}

}  // namespace

// If over kMinAggregateCount gradients are accumulated and the total
// memory consumption is over kMinAggregateBytes, do an early aggregation
// so as to release the gradient tensor to save memory.
constexpr int kMinAggregateCount = 4;

    "//tensorflow/tools/lib_package:jnilicenses_generate"
}

# This test ensures that all the targets built into the Python package include
# their dependencies. It's a rewritten version of the "smoke test", an older
# Python script that was very difficult to understand. See
# https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/pip_package/pip_smoke_test.py
@test "Pip package includes all required //tensorflow dependencies" {