programs that TensorFlow executes. TensorFlow programs are encoded as
computation
[**graphs**](https://developers.google.com/machine-learning/glossary/#graph).
Since models are practically programs that TensorFlow executes, using untrusted
models or graphs is equivalent to running untrusted code.

If you need to run untrusted models, execute them inside a

cat <<EOF
IMPORTANT: These tests ran under docker. This script does not clean up the
container for you! You can delete the container with:

$ docker rm -f tf

You can also execute more commands within the container with e.g.:

$ docker exec tf bazel clean
$ docker exec -it tf bash
EOF

    if (TF_GetCode(s) != TF_OK) {
      return;
    }
  }
  int num_outputs = unwrap(o)->expected_num_outputs;
  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Execute(
             absl::MakeSpan(reinterpret_cast<AbstractTensorHandle**>(
                                unwrap(o)->outputs.data()),
                            unwrap(o)->outputs.size()),
             &num_outputs));
}

2. Running the CI scripts locally, as explained below
3. **Google employees only**: Google employees can use an internal-only tool
called "MLCI" that makes testing more convenient: it can execute any full CI job
against a pending change. Search for "MLCI" internally to find it.

You may invoke a CI script of your choice by following these instructions:

```bash
cd tensorflow-git-dir

  exit 1
fi

# "TFCI_MACOS_BAZEL_TEST_DIR_PATH" specifies the directory that Bazel should use
# when running tests. Each test will be executed in a separate subdirectory
# inside this directory. TF Mac builds need ~150 GB of disk space to be able to
# run all the tests. Since TFCI Mac VMs execute Bazel test commands in a
# partition with insufficient storage, we specify the

      TF_RETURN_IF_ERROR(fn_op->AddInput(input));
    }
    int retvals = outputs.size() - null_indices.size();
    std::vector<AbstractTensorHandle*> fn_outputs(retvals);
    TF_RETURN_IF_ERROR(fn_op->Execute(
        absl::Span<AbstractTensorHandle*>(fn_outputs.data(), fn_outputs.size()),
        &retvals));
    int skipped_indices = 0;
    for (int i = 0; i < outputs.size(); i++) {
      if (!null_indices.contains(i)) {

  GradientRegistry registry;
  s = RegisterGradients(&registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto tape = std::make_unique<Tape>(/*persistent=*/false);
  s = Execute(check_numerics_op.get(), ctx.get(), absl::MakeSpan(outputs),
              &num_retvals, &forward_op, tape.get(), registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

  string read_message;

  explicit TFE_Executor(tensorflow::EagerExecutor* executor)
      : owned_executor(nullptr), unowned_executor(executor) {}

  tensorflow::EagerExecutor* executor() {
    return owned_executor == nullptr ? unowned_executor : owned_executor.get();
  }

  std::unique_ptr<tensorflow::EagerExecutor> owned_executor;
  tensorflow::EagerExecutor* unowned_executor;
};

// `struct_size` is used for struct member compatibility check between core TF
// and plug-ins with the same C API minor version. More info here:
// https://github.com/tensorflow/community/blob/master/rfcs/20200612-stream-executor-c-api/C_API_versioning_strategy.md
#define TF_VALIDATE_STRUCT_SIZE(STRUCT_NAME, STRUCT_OBJ, SIZE_VALUE_NAME) \
  do {                                                                    \

// a map (`tensorflow::eager::TensorTape`) from the wrapped tensor to the id of
// the op that produced it (or -1 if this tensor was watched using
// `GradientTape::Watch`.) The op_id is simply a unique index assigned to each
// op executed under the tape. A separate map (`tensorflow::eager::OpTape`)
// maintains the map from `op_id` to a `OpTapeEntry` which stores the `op_type`,
// inputs and outputs and the gradient function These data structures combined