// `ctx_0`, `ctx_1`, `ctx_2` contains `remote_device`.
  {
    const std::vector<std::string>& device_names = ListDeviceNames(ctx_0);
    ASSERT_TRUE(std::find(device_names.begin(), device_names.end(),
                          remote_device) != device_names.end());
  }

  {
    const std::vector<std::string>& device_names = ListDeviceNames(ctx_1);

  TF_ExecutionContext* ctx = TF_NewEagerExecutionContext(opts, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  TFE_DeleteContextOptions(opts);

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Build an abstract input tensor.
  TFE_Context* eager_ctx = TF_ExecutionContextGetTFEContext(ctx, status.get());

  TFE_Executor* old_executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_Executor* executor = TFE_NewExecutor(
      /*is_async=*/async, /*enable_streaming_enqueue=*/true,
      /*in_flight_nodes_limit=*/0);
  TFE_ContextSetExecutorForThread(ctx, executor);

  TFE_TensorHandle* m = TestMatrixTensorHandle(ctx);
  TFE_Op* matmul = MatMulOp(ctx, m, m);
  TFE_TensorHandle* retvals[2] = {nullptr, nullptr};
  int num_retvals = 2;

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

    {
      AbstractContext* ctx_raw = nullptr;
      status_ =
          BuildImmediateExecutionContext(std::get<1>(GetParam()), &ctx_raw);
      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically

Status TestTensorHandleWithDims(AbstractContext* ctx, const T* data,
                                const int64_t* dims, int num_dims,
                                AbstractTensorHandle** tensor) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TFE_Context* eager_ctx =
      TF_ExecutionContextGetTFEContext(wrap(ctx), status.get());
  TF_RETURN_IF_ERROR(StatusFromTF_Status(status.get()));

#define C_OPKERNELCONTEXT_REQUIRES_OK(CTX, C_STATUS, __VA_ARGS__) \
  do {                                                            \
    ::tensorflow::Status _s(__VA_ARGS__);                         \
    if (!TF_PREDICT_TRUE(_s.ok())) {                              \
      ::tensorflow::Set_TF_Status_from_Status(C_STATUS, _s);      \
      TF_OpKernelContext_Failure(CTX, C_STATUS);                  \

  TestHandleFromDLPack(status, ctx, {1}, {1});
  TestHandleFromDLPack(status, ctx, {1}, {0});
  TestHandleFromDLPack(status, ctx, {4, 1, 2}, {2, 1, 1});
  TestHandleFromDLPack(status, ctx, {4, 1, 2}, {2, 0, 1});
  TestHandleFromDLPack(status, ctx, {4, 3, 1}, {3, 1, 1});
  TestHandleFromDLPack(status, ctx, {4, 3, 1}, {3, 1, 0});
  // Test that empty tensors can have any strides.
  TestHandleFromDLPack(status, ctx, {4, 0, 2}, {0, 2, 1});

  ImmediateTensorHandlePtr dim(imm_ctx->CreateLocalHandle(minus_1.get()));
  AbstractTensorHandle* expand_dims_outputs;
  TF_RETURN_IF_ERROR(
      ops::ExpandDims(ctx, vec, dim.get(), &expand_dims_outputs, "ExpandDims"));
  TF_RETURN_IF_ERROR(
      ops::Mul(ctx, expand_dims_outputs, mat, &outputs[0], "Mul"));
  expand_dims_outputs->Unref();
  return absl::OkStatus();
}

    return nullptr;
  }

  TensorReference tensor_ref(*tensor);  // This will call buf_->Ref()
  auto* tf_dlm_tensor_ctx = new TfDlManagedTensorCtx(tensor_ref);
  tf_dlm_tensor_ctx->reference = tensor_ref;

  DLManagedTensor* dlm_tensor = &tf_dlm_tensor_ctx->tensor;
  dlm_tensor->manager_ctx = tf_dlm_tensor_ctx;
  dlm_tensor->deleter = &DLManagedTensorDeleter;
  dlm_tensor->dl_tensor.device = tf_dlm_context;

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

    {
      AbstractContext* ctx_raw = nullptr;
      status_ =
          BuildImmediateExecutionContext(std::get<1>(GetParam()), &ctx_raw);
      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically