AbstractTensorHandlePtr two_eps;
    {
      AbstractTensorHandle* two_eps_raw = nullptr;
      TF_RETURN_IF_ERROR(TestScalarTensorHandle<float, TF_FLOAT>(
          ctx, 2 * epsilon, &two_eps_raw));
      two_eps.reset(two_eps_raw);
    }

    // Initialize theta[i] + epsilon.
    memcpy(thetaPlus_data.data(), TF_TensorData(theta_tensor),

      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_OpSetAttrType(handle_op.get(), "dtype", TF_FLOAT);
  TFE_OpSetAttrShape(handle_op.get(), "shape", /*dims=*/nullptr, /*num_dims=*/0,
                     status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  auto outputs =

    absl::Status s = TestTensorHandleWithDims<float, TF_FLOAT>(
        ctx_.get(), A_vals, A_dims, 2, &A_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    A.reset(A_raw);
  }
  float B_vals[] = {.5f, -1.0f, 1.0f, 1.0f};
  int64_t B_dims[] = {2, 2};
  AbstractTensorHandlePtr B;
  {
    AbstractTensorHandle* B_raw;
    absl::Status s = TestTensorHandleWithDims<float, TF_FLOAT>(
        ctx_.get(), B_vals, B_dims, 2, &B_raw);

    ctx.reset(ctx_raw);
  }

  AbstractTensorHandlePtr t;
  {
    AbstractTensorHandle* x_raw = nullptr;
    absl::Status s =
        TestScalarTensorHandle<float, TF_FLOAT>(ctx.get(), 1.0f, &x_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    t.reset(x_raw);
  }

  AbstractOperationPtr check_numerics_op(ctx->CreateOperation());
  ForwardOperation forward_op;

  // data.
  TF_Tensor* t =
      TF_NewTensor(TF_FLOAT, nullptr, 0, nullptr, 0, &NoOpDeallocator, nullptr);
  ASSERT_TRUE(t == nullptr);
}

TEST(CAPI, AllocateTensor) {
  const int num_bytes = 6 * sizeof(float);
  int64_t dims[] = {2, 3};
  TF_Tensor* t = TF_AllocateTensor(TF_FLOAT, dims, 2, num_bytes);
  EXPECT_EQ(TF_FLOAT, TF_TensorType(t));
  EXPECT_EQ(2, TF_NumDims(t));

  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  auto* arg0 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  auto* arg1 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Create a first "Add" computing `arg0 + arg1`.

  dtype.bits = TF_DataTypeSize(data_type) * 8;
  switch (data_type) {
    case TF_DataType::TF_BOOL:
      dtype.code = DLDataTypeCode::kDLBool;
      break;
    case TF_DataType::TF_HALF:
    case TF_DataType::TF_FLOAT:
    case TF_DataType::TF_DOUBLE:
      dtype.code = DLDataTypeCode::kDLFloat;
      break;
    case TF_DataType::TF_INT8:
    case TF_DataType::TF_INT16:
    case TF_DataType::TF_INT32:

  TF_SetAttrType(placeholder_desc, "dtype", TF_FLOAT);
  TF_Operation* placeholder_op = TF_FinishOperation(placeholder_desc, status);
  if (TF_GetCode(status) != TF_OK) return;
  TF_Output x{placeholder_op, 0};

  TF_OperationDescription* reduce_desc =
      TF_NewOperation(body.get(), "CollectiveReduce", "CollectiveReduce");
  TF_SetAttrType(reduce_desc, "T", TF_FLOAT);
  TF_SetAttrInt(reduce_desc, "group_size", group_size);

  TF_Status* status = TF_NewStatus();
  TFE_Op* op = TFE_NewOp(ctx, "ReadVariableOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpAddInput(op, var, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  int num_retvals = 1;
  TFE_Execute(op, out_value, &num_retvals, status);