return %1 : tensor<1x3x4x2xf32>
  }

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

  func.return %dequant : tensor<1x3x4x2xf32>
}

// -----

// CHECK-LABEL: @dont_merge_quantization_followed_by_quantization
// 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]]

    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    func.return %0: tensor<1x2x2x3xf32>
  }

  func.func private @some_func(%arg0: tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32> {
    return %arg0 : tensor<1x2x2x3xf32>
  }
}

// CHECK: module
// CHECK-NOT: tf.PartitionedCall
// CHECK-NOT: some_func

// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

func.func @cast_bf16_avg_pool_to_fp32(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+00> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>

        _stablehlo_module_attrs = {}, _tfl_quant_trait = "fully_quantizable",
        device = ""
      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

  // CHECK: func.func private @qdq_for_conv_weight_per_channel_default(%[[ARG0:.+]]: tensor<1x3x4x3xf32>)
  // CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<3xf32>) -> tensor<*xf32>
    %2 = "tf.PartitionedCall"(%arg0, %cst_2) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn_1} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

// Remove TPU related ops.
func.func @tpu_conv(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
  %0 = "tf.TPUOrdinalSelector"() {device = ""} : () -> tensor<?xi32>
  %1 = "tf.TPUPartitionedCall"(%arg0, %0) {autotuner_thresh = 0 : i64, device = "", f = @tpu_func_0_optim0} : (tensor<1x3x4x3xf32>, tensor<?xi32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %2 : tensor<1x3x2x2xf32>
}

    %3 = "tf.Identity"(%2) {device = ""} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    %4 = "quantfork.stats"(%3) {layerStats = dense<[3.50919247, 6.000000e+00]> : tensor<2xf32>} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    func.return %4 : tensor<1x3x4x2xf32>
  }

  %1 = "tf.Maximum"(%0, %cst_0) : (tensor<1x3x4x2xf32>, tensor<f32>) -> tensor<1x3x4x2xf32>
  %2 = "tf.Minimum"(%1, %cst_1) : (tensor<1x3x4x2xf32>, tensor<f32>) -> tensor<1x3x4x2xf32>
  func.return %2 : tensor<1x3x4x2xf32>

    return %2 : tensor<1x3x4x2xf32>
  }

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