# non-interactive shells as well.
  shopt -s expand_aliases
else
  echo '==TFCI==: Error: Cannot find path to grealpath or gstat'
  echo 'TF CI scripts require GNU core utilties to be installed. Please make'
  echo 'sure they are present on your system and try again.'
  exit 1
fi

# "TFCI_MACOS_BAZEL_TEST_DIR_PATH" specifies the directory that Bazel should use

                                           size_t plugin_size) {
  if (plugin_ops == nullptr) return nullptr;

  size_t copy_size = std::min(plugin_size, sizeof(T));
  auto core_ops = std::make_unique<T>();
  memset(core_ops.get(), 0, sizeof(T));
  memcpy(core_ops.get(), plugin_ops, copy_size);
  return core_ops;
}

// Registers one filesystem from the plugin.
//

  outputs[0]->Unref();
}

TEST_P(CppGradients, TestSparseSoftmaxCrossEntropyWithLogitsGrad) {
  if (UseFunction()) {
    // TODO(b/168850692): Enable this.
    GTEST_SKIP() << "Can't take gradient of "
                    "SparseSoftmaxCrossEntropyWithLogits in tracing mode.";
  }

  // Score
  float X_vals[] = {1.0f, 2.0f, 3.0f, -5.0f, -4.0f, -3.0f, 2.0f, 0.0f, -1.0f};
  int64_t X_dims[] = {3, 3};

Eager mode lets users write imperative-style statements that can be easily
inspected and debugged and it is intended to be used during the development
phase.

As part of the differences that make Eager mode easier to debug, the [shape
inference
functions](https://www.tensorflow.org/guide/create_op#define_the_op_interface)
are skipped, and any checks implemented inside the shape inference code are not
executed.

namespace internal {
struct AbstractTensorHandleDeleter {
  void operator()(AbstractTensorHandle* p) const {
    if (p != nullptr) {
      p->Unref();
    }
  }
};
}  // namespace internal

// TODO(b/185908092): Make AbstractTensorHandlePtr an IntrusivePtr.
using AbstractTensorHandlePtr =
    std::unique_ptr<AbstractTensorHandle,
                    internal::AbstractTensorHandleDeleter>;

}  // namespace tensorflow

Status SetAttrFunctionList(AbstractOperation*, const char* attr_name,
                           absl::Span<const AbstractOperation*> values,
                           ForwardOperation*);

// Make the call to `Tape::RecordOperation`.
Status Execute(AbstractOperation*, AbstractContext*,
               absl::Span<AbstractTensorHandle*> retvals, int* num_retvals,
               ForwardOperation*, Tape*, const GradientRegistry&);

}

absl::Status SetPjRtCBufferToTensor(PJRT_Buffer* c_buffer,
                                    xla::PjRtCApiClient* c_api_client,
                                    Tensor* tensor) {
  auto buffer = std::make_unique<xla::PjRtCApiBuffer>(c_api_client, c_buffer);
  tensorflow::AsyncValueTensor* av_tensor =
      tensorflow::AsyncValueTensor::FromTensor(tensor);
  if (av_tensor == nullptr) {
    TF_ASSIGN_OR_RETURN(

//
// Very similar to TF_OperationDescription with some differences:
// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList
// (2) TF_ColocateWith, TF_AddControlInput etc. do not make sense.
// (3) Implementation detail: Avoid use of NodeBuilder/NodeDefBuilder since
//     the additional sanity checks there seem unnecessary;
typedef struct TFE_Op TFE_Op;

      ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
    lru_list_.push_front(key);
    Entry entry{timer_seconds_(), value, lru_list_.begin()};
    auto insert = cache_.insert(std::make_pair(key, entry));
    if (!insert.second) {
      lru_list_.erase(insert.first->second.lru_iterator);
      insert.first->second = entry;
    } else if (max_entries_ > 0 && cache_.size() > max_entries_) {

# so I am not sure how to verify that the dependencies themselves are valid for
# the moment.
if [[ "$TFCI_BAZEL_COMMON_ARGS" =~ gpu|cuda ]]; then
  echo "Checking to make sure tensorflow[and-cuda] is installable..."
  "$python" -m pip install "$(echo *.whl)[and-cuda]" $TFCI_PYTHON_VERIFY_PIP_INSTALL_ARGS