: (tensor<?x3x2x1xi8>, tensor<2x1x1x1xi8>) -> tensor<?x2x2x1xi32>
  %1 = chlo.broadcast_add %0, %zp_offset : (
      tensor<?x2x2x1xi32>, tensor<?x2x2x1xi32>) -> tensor<?x2x2x1xi32>
  %2 = chlo.broadcast_add %1, %bias : (
      tensor<?x2x2x1xi32>, tensor<1xi32>) ->tensor<?x2x2x1xi32>
  %3 = chlo.broadcast_add %2, %bias : (
      tensor<?x2x2x1xi32>, tensor<1xi32>) ->tensor<?x2x2x1xi32>
  return %3 : tensor<?x2x2x1xi32>
}

// -----

    has_token_input_output = false, module = "", platforms = [],
    version = 5 : i64
  } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>
  return %0 : tensor<1x2x2x2xf32>
}

// CHECK-LABEL: func.func @qdq_for_conv_weight_empty
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32>
// CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    }> {
      _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

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x2x1x3x!quant.uniform<i8:f32, 1.0>>, volatile} : (tensor<1x2x1x3xf32>) -> tensor<1x2x1x3x!quant.uniform<i8:f32, 1.0>>
  %5 = "tfl.dequantize"(%arg0) : (tensor<1x2x1x3x!quant.uniform<i8:f64, 1.0>>) -> tensor<1x2x1x3xf32>
  "tfl.assign_variable"(%1, %5) : (tensor<!tf_type.resource>, tensor<1x2x1x3xf32>) -> ()

  %2 = "tf.Relu6"(%1) {device = ""} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
  return %2 : tensor<1x2x2x2xf32>
}
// CHECK-LABEL: @composite_conv2d_with_bias_and_relu6_fn_1
// CHECK-NOT: "tf.CalibrationStatisticsSaver"

// -----

// Check the IfOp is set to stateful.

  // Use a four dimension sharding (devices=[1,1,1,1]0)
  // Since the input tensor only has three dimensions, we expect this to fail.
  %0 = "tf.XlaSharding"(%arg0) { _XlaSharding = "\08\03\1A\04\01\01\01\01\22\01\00" } : (tensor<1x2x3xi32>) -> tensor<1x2x3xi32>
  %1 = "tf.A"(%0) : (tensor<1x2x3xi32>) -> (tensor<1x2x3xi32>)
  func.return %1: tensor<1x2x3xi32>
}

// -----

// CHECK-LABEL: func @check_retval_sharding_errors

// CHECK-NEXT: %[[CONV2D:.*]] = "tf.Conv2D"(%arg0, %[[CONST]]) <{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>
// CHECK-NEXT: %[[ADD:.*]] = "tf.AddV2"(%[[CONV2D]], %[[CONST_0]]) : (tensor<1x3x2x3xf32>, tensor<1x3x2x3xf32>) -> tensor<1x3x2x3xf32>

func.func @depth_to_space(%arg0: tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tf.DepthToSpace"(%arg0) {block_size = 2: i64,  data_format = "NHWC"}: (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
  func.return %0 : tensor<1x2x2x1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexDepthToSpace", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
// CHECK: return %[[CUSTOM_0]] : tensor<1x2x2x1xf32>
}

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x2x2x3xf32>

    %1 = "tf.PartitionedCall"(%arg0, %cst, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_conv2d_with_bias_and_relu6_fn_1} : (tensor<1x2x2x3xf32>, tensor<2x2x3x2xf32>, tensor<2xf32>) -> tensor<*xf32> loc(callsit...