auto y = GatherV2(scope_, x, {9, 8, 7, 6}, /*axis=*/3);
  TensorShape y_shape({2, 3, 7, 4});
  RunTest(x, shape, y, y_shape);
}

TEST_F(ArrayGradTest, GatherV2Grad_BatchDim) {
  TensorShape shape({2, 100, 3});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
  GatherV2::Attrs attrs;
  attrs.batch_dims_ = 1;
  auto y =

// CHECK-NEXT: [[res:%.*]] = tfrt_fallback_async.executeop key(1) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.GatherV2"([[value]], {{.*}}, [[const]])
// CHECK-NEXT: tfrt.return [[out_chain]], [[res]] : !tfrt.chain, !tfrt_fallback.tf_tensor
func.func @gather(%indices: tensor<?xi32>,

// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.GatherV2
// CHECK-NEXT: tf.AddV2
// CHECK-NEXT: tf.AddV2
// CHECK-NEXT: tf.GatherV2
// CHECK-NEXT: tf.Mul
// CHECK-NEXT: tf.GatherV2
// CHECK-NEXT: tf.Mul
// CHECK-NEXT: tf_mlrt.tf_await
// CHECK-NEXT: tf_mlrt.tf_await
// CHECK-NEXT: tf.TensorListSetItem

  }
}


// -----
//  Decomposable Resource Ops usage
//
// CHECK-LABEL:  func.func @serving_default
// CHECK:      "tf.VarHandleOp"
// CHECK-NEXT: "tf.IfrtLoadVariable"
// CHECK-NEXT: "tf.GatherV2"
// CHECK-NEXT:  return 
//
module {
  func.func @serving_default() -> tensor<1x3xbf16> {
    %cst = "tf.Const"() <{value = dense<[1]> : tensor<1xi32>}> : () -> tensor<1xi32>

  %383 = "tf.GatherV2"(%11, %382#1, %14) {batch_dims = 0 : i64, device = ""} : (tensor<1xi64>, tensor<?xi64>, tensor<i32>) -> tensor<?xi64>
  %384 = "tf.GatherV2"(%12, %383, %14) {batch_dims = 0 : i64, device = ""} : (tensor<1xi64>, tensor<?xi64>, tensor<i32>) -> tensor<?xi64>

  if (mlir::isa<mlir::TF::StringType>(key_type.getElementType()))
    cost *= kLookupTableFindStringKeyCostScale;

  return cost;
}

// The cost function for tf.GatherV2.
int64_t InferGatherV2Cost(const CostContext& context, mlir::TF::GatherV2Op op) {
  return InferTensorSize(
      context, mlir::cast<mlir::TensorType>(op.getOutput().getType()));
}

  std::set<string> constrained_kernel = {"Diag",
                                         "DiagPart",
                                         "GatherNd",
                                         "GatherV2",
                                         "QuantizeAndDequantizeV2",
                                         "QuantizeAndDequantizeV3",
                                         "QuantizeAndDequantizeV4",

  }
  func.func private @composite_gather_fn_1(%arg0: tensor<1024x3x4x3xf32>, %arg1: tensor<1xi32>, %arg2: tensor<i32>) -> tensor<1x3x4x3xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.GatherV2"(%arg0, %arg1, %arg2) {attr_map = "0:batch_dims", batch_dims = 0 : i64, device = ""} : (tensor<1024x3x4x3xf32>, tensor<1xi32>, tensor<i32>) -> tensor<1x3x4x3xf32>
    return %0 : tensor<1x3x4x3xf32>
  }

  %cst = "tf.Const"() {device = "", value = dense<0> : tensor<i32>} : () -> tensor<i32>
  %cst_0 = "tf.Const"() {device = "", value = dense<1.0> : tensor<128x32xf32>} : () -> tensor<128x32xf32>
  %0 = "tf.GatherV2"(%cst_0, %arg0, %cst) {batch_dims = 0 : i64, device = ""} : (tensor<128x32xf32>, tensor<6xi64>, tensor<i32>) -> tensor<6x32xf32>
  return %0 : tensor<6x32xf32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<i32>

    // expected-remark@+1 {{Cost: 1}}
    %0 = "tf.Const"() { value = dense<[1]> : tensor<1xi32> } : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 300}}
    %1 = "tf.GatherV2"(%arg0, %arg1, %0) : (tensor<1x2x20xf32>, tensor<3x5xi32>, tensor<1xi32>) -> tensor<1x3x5x20xf32>
    // expected-remark@+1 {{Cost: 40}}
    func.return %1 : tensor<1x3x5x20xf32>
}