%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>
  }

    func.return %1: tensor<*xf32>
  }

  func.func private @outer_fn(%arg0: tensor<1x2x2x3xf32>, %arg1: tensor<2x1024xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.PartitionedCall"(%arg0, %arg1) {config = "", config_proto = "", executor_type = "", f = @inner_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x1x1x3xf32>) -> tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x1x1x3xf32>
  %5 = "tfl.add"(%1, %4) {fused_activation_function = "NONE"} : (tensor<1x5x5x3xf32>, tensor<1x1x1x3xf32>) -> tensor<1x5x5x3xf32>

    return %2 : tensor<1x2x2x3xf32>
  }
  func.func private @quantize_i8(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<f32>, %arg2: tensor<i32>) -> tensor<1x3x4x3xi8> {
    %0 = "tf.Div"(%arg0, %arg1) : (tensor<1x3x4x3xf32>, tensor<f32>) -> tensor<1x3x4x3xf32>
    %1 = "tf.Round"(%0) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

  %res_p3 = "tfl.cumsum"(%arg, %axis_p3) {exclusive = false, reverse = false} : (tensor<1x2x1x3xf32>, tensor<i32>) -> tensor<1x2x1x3xf32>
  func.return %res_m4, %res_m3, %res_m2, %res_m1, %res_00, %res_p1, %res_p2, %res_p3 : tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>

// CHECK-SAME: %[[ARG_0:.*]]: tensor<1x3x4x3xf32>
func.func @dont_merge_quantization_followed_by_quantization(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32> {
  // CHECK: %[[QUANT_ARG_0:.*]] = stablehlo.uniform_quantize %[[ARG_0]]
  // CHECK: %[[DEQUANT:.*]] = stablehlo.uniform_dequantize %[[QUANT_ARG_0]]
  // CHECK: return %[[DEQUANT]]

    %2 = stablehlo.convolution(%1, %0)
      dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f],
      window = {
        stride = [1, 1], pad = [[0, 0], [1, 1]],
        lhs_dilate = [1, 1],
        rhs_dilate = [1, 1]
      }
      {
        batch_group_count = 1 : i64,
        feature_group_count = 1 : i64
      } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>)

    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    // CHECK-SKIP: tf.PartitionedCall
    // CHECK-NOSKIP: call @some_other_func
    // CHECK-NOSKIP-NOT: tf.PartitionedCall
    func.return %1: tensor<1x2x2x3xf32>
  }

  // CHECK-SKIP: func.func private @some_func
  func.func private @some_func(%arg0: tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32> attributes {tf._noinline = true} {

// Confirm a wide array of attribute survives the round-trip
func.func @main(tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32> {
^bb0(%arg0: tensor<1x6x6x16xf32>):
  // CHECK: "tfl.average_pool_2d"(%{{.*}}) <{filter_height = 3 : i32, filter_width = 6 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 3 : i32, stride_w = 1 : i32}> : (tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32>