%0 = "tf.Conv2D"(%arg0, %cst) {data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x3xf32>) -> tensor<1x3x2x3xf32>
  %1 = "tf.AddV2"(%0, %cst_0) : (tensor<1x3x2x3xf32>, tensor<1x3x2x3xf32>) -> tensor<1x3x2x3xf32>
  func.return %1 : tensor<1x3x2x3xf32>
}

module {
  func.func @matmul(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

    %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>

    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} {

// 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} {

// CHECK:           %[[VAL_2:.*]] = "tfl.transpose"(%[[VAL_0]], %[[VAL_1]]) : (tensor<1x3x6x6xf32>, tensor<4xi32>) -> tensor<1x6x6x3xf32>
// CHECK:           %[[VAL_3:.*]] = arith.constant dense<0> : tensor<4x2xi32>
// CHECK:           %[[VAL_4:.*]] = "tfl.pad"(%[[VAL_2]], %[[VAL_3]]) : (tensor<1x6x6x3xf32>, tensor<4x2xi32>) -> tensor<1x6x6x3xf32>