return %1 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>

func.func @mean_4d_no_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = false} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x512xf32>
  func.return %0 : tensor<1x512xf32>
}

// CHECK:       func @mean_4d_no_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {

  %weight = arith.constant dense<1.0> : tensor<144x12xf32>
  %q_weight = "quantfork.qcast"(%weight) : (tensor<144x12xf32>) -> tensor<144x12x!quant.uniform<i8:f32, 0.074855112561992565:-1>>
  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<144x12x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<144x12xf32>
  %9 = "tf.MatMul"(%7, %dq_weight) {transpose_a = false, transpose_b = false} : (tensor<*xf32>, tensor<144x12xf32>) -> tensor<*xf32>

// -----

// CHECK-LABEL: float_matmul
func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>

  // CHECK:  %5 = "tfl.arg_max"(%arg0, %cst) : (tensor<4x32x256xf32>, tensor<1xi32>) -> tensor<4x32xi32>
  // CHECK:  return %4, %5 : tensor<4x32xf32>, tensor<4x32xi32>
}

// CHECK-LABEL: func @convert_argmax_constant
func.func @convert_argmax_constant(%arg0: tensor<2x2x4xf32>) -> (tensor<2x2xf32>, tensor<2x2xi32>) {
  %0 = mhlo.constant dense<0xFF800000> : tensor<f32>

  "tfl.custom"(%0, %2) {custom_code = "FlexAssignVariableOp", custom_option = #tfl<const_bytes : "0x1041737369676E5661726961626C654F70003B121041737369676E5661726961626C654F701A001A002A0B0A056474797065120230012A140A0E76616C69646174655F736861706512022800320000024F3E1414042801">} : (tensor<!tf_type.resource<tensor<1x10xf32>>>, tensor<1x10xf32>) -> ()

  %flow_1 = "tf.TensorArrayWriteV3"(%handle, %index, %value, %flow) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<2x!tf_type.resource<tensor<1x512xf32>>>, tensor<i32>, tensor<1x512xf32>, tensor<f32>) -> tensor<f32>

  %0 = "tf_device.cluster_func"(%arg0, %arg1) {func = @_func, use_spmd_for_xla_partitioning = true, use_tpu = true, num_cores_per_replica = 2 : i64} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1:2 = "tf.TPUPartitionedOutputV2"(%0) {device = "", partition_dims = [2, 1]} : (tensor<2x2xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)

    %3 = shape.shape_of %2 : tensor<?x2xf32> -> tensor<2xindex>
    %4 = stablehlo.dynamic_broadcast_in_dim %arg2, %3, dims = [1] : (tensor<2xf32>, tensor<2xindex>) -> tensor<?x2xf32>
    %5 = stablehlo.add %2, %4 : tensor<?x2xf32>
    %6 = stablehlo.clamp %0, %5, %1 : (tensor<f32>, tensor<?x2xf32>, tensor<f32>) -> tensor<?x2xf32>
    return %6 : tensor<?x2xf32>
  }

"/job:tpu_host_worker/replica:0/task:3/device:CPU:0"}, %arg15: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "927", tf.device = "/job:tpu_host_worker/replica:0/task:0/device:CPU:0"}, %arg16: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "929", tf.device = "/job:tpu_host_worker/replica:0/task:1/device:CPU:0"}, %arg17: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "931", tf.device = "/job:tpu_host_worker/replica:0/task:2/device:CPU:0"},...