TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  executed = false;
  TFE_Execute(op.get(), &var_handle, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  ASSERT_TRUE(executed);
  auto handle_cleaner = tensorflow::gtl::MakeCleanup(
      [var_handle]() { TFE_DeleteTensorHandle(var_handle); });

  TFE_DeleteTensorHandle(h0_task0);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TF_DeleteStatus(status);
}

// Read the value of variable `var` and save it into `out_value`.

  parallel_device.Execute(context.get(), handle_inputs, "ReadVariableOp",
                          TFE_OpGetAttrs(read_op.get()),
                          /*expected_max_outputs=*/1, status.get());
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);
  TF_SetStatus(status.get(), TF_OK, "");

  // Check that ops still run successfully on the device.
  parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),

                                                  TF_Status* status) {
  tensorflow::EagerContext* context =
      tensorflow::ContextFromInterface(tensorflow::unwrap(ctx));
  auto collective_executor_handle = context->GetCollectiveExecutorHandle();
  collective_executor_handle->get()->StartAbort(status->status);
}

TF_CAPI_EXPORT extern void TFE_CollectiveOpsCheckPeerHealth(
    TFE_Context* ctx, const char* task, int64_t timeout_in_ms,

  TFE_DeleteTensorHandle(h0);
  TFE_DeleteTensorHandle(h1);
  TFE_DeleteTensorHandle(h2);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_ContextRemoveFunction(ctx, "AddVariablesFunction", status);
  TFE_DeleteContext(ctx);

ParallelDevice::Execute(TFE_Context* context,
                        const std::vector<ParallelTensor*>& inputs,
                        const char* operation_name,
                        const TFE_OpAttrs* attributes, int expected_max_outputs,
                        TF_Status* status) const {
  std::vector<PartialTensorShape> expected_output_shapes(expected_max_outputs);

      "SetAttrFunctionList has not been "
      "implemented yet.");
}
Status Execute(AbstractOperation* op_, AbstractContext* ctx,
               absl::Span<AbstractTensorHandle*> retvals, int* num_retvals,
               ForwardOperation* forward_op_, Tape* tape,
               const GradientRegistry& registry) {
  TF_RETURN_IF_ERROR(op_->Execute(retvals, num_retvals));
  for (int i = 0; i < *num_retvals; i++) {

  bool valid = true;
  int64_t expected_stride = 1;
  for (int i = ndim - 1; i >= 0; --i) {
    // Empty tensors are always compact regardless of strides.
    if (shape_arr[i] == 0) return true;
    // Note that dimensions with size=1 can have any stride.
    if (shape_arr[i] != 1 && stride_arr[i] != expected_stride) {
      valid = false;
    }
    expected_stride *= shape_arr[i];
  }
  return valid;
}

    TFE_OpAddAttrs(op.get(), attributes);
    int expected_outputs = TFE_OpGetOutputLength(op.get(), "outputs", status);
    if (TF_GetCode(status) != TF_OK) return result;
    if (expected_outputs != parallel_device.num_underlying_devices()) {
      std::string message(absl::StrCat(
          "The parallel device ", parallel_device_name, " expected ",
          parallel_device.num_underlying_devices(),

#include "tensorflow/c/c_api_experimental.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {