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};

  ASSERT_EQ(errors::OK, s.code()) << s.message();
}

TEST_P(UnifiedAPI, TestUnknownShapeTracing) {
  if (!UseFunction()) {
    GTEST_SKIP() << "Tracing only test.";
  }
  if (UseMlir()) {
    // TODO(b/173074167): Remove this.
    GTEST_SKIP() << "MlirTensor::Shape is not implemented yet.";
  }
  AbstractContextPtr ctx(BuildFunction("test_fn"));
  AbstractTensorHandlePtr x;
  {

TEST(CAPI, Session_Min_GPU) {
  const string gpu_device = GPUDeviceName();
  // Skip this test if no GPU is available.
  if (gpu_device.empty()) return;

  RunMinTest(gpu_device, /*use_XLA=*/false);
}

TEST(CAPI, Session_Min_XLA_GPU) {
  const string gpu_device = GPUDeviceName();
  // Skip this test if no GPU is available.
  if (gpu_device.empty()) return;

  opts.prefix = prefix;

  for (int i = 0; i < dst_inputs.size(); ++i) {
    opts.input_map[ToTensorId(src_inputs[i])] =
        TensorId(dst_inputs[i].node()->name(), dst_inputs[i].index());
  }
  opts.skip_mapped_nodes = true;

  for (const tensorflow::Operation& op : control_deps) {
    opts.control_dependencies.push_back(op.node()->name());
  }

  for (int i = 0; i < nreturn_nodes; ++i) {

    op_stack.pop_back();
    auto op_it = state.op_tape.find(op);
    if (op_it == state.op_tape.end()) {
      // It is possible for ops to end up on the stack if they are unrelated to
      // the target; we should just skip them.
      continue;
    }
    auto trace = std::move(op_it->second);
    state.op_tape.erase(op_it);
    std::vector<Gradient*> out_gradients;
    out_gradients.reserve(trace.output_tensor_info.size());

      advanced use cases can also write their own custom
      `tf.train.experimental.ShardingCallback`s.

* `tf.train.CheckpointOptions`
    * Added `experimental_skip_slot_variables` (a boolean option) to skip
    restoring of optimizer slot variables in a checkpoint.

*   `tf.saved_model.SaveOptions`

    * `SaveOptions` now takes a new argument called

// and set the device name. It's effectively `TFE_OpSetDevice`, but it is faster
// than separately calling it because if the existing op has the same
// `raw_device_name`, it skips parsing and just leave as it is.
TF_CAPI_EXPORT extern void TFE_OpReset(TFE_Op* op_to_reset,
                                       const char* op_or_function_name,
                                       const char* raw_device_name,

#     avx_linux:              Build with avx instruction set on linux.
#     avx_win:                Build with avx instruction set on windows
#
# Other build options:
#     short_logs:       Only log errors during build, skip warnings.
#     verbose_logs:     Show all compiler warnings during build.
#     monolithic:       Build all TF C++ code into a single shared object.
#     dynamic_kernels:  Try to link all kernels dynamically (experimental).