"vhlo.return_v1" (%421) : (tensor<1xf32>) -> ()
   }) : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  return %0 : tensor<1x320xf32>
}

//CHECK:func.func private @reduce(%arg0: tensor<1x16x16x320xf32>, %arg1: tensor<f32>) -> tensor<1x320xf32> {
//CHECK-NEXT:  %0 = "vhlo.reduce_v1"(%arg0, %arg1) <{dimensions = #vhlo.tensor_v1<dense<[1, 2]> : tensor<2xi64>>}> ({ 

  %1 = "tf.TensorListGetItem"(%0, %arg2, %arg1) {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<!tf_type.variant<tensor<1x32xf32>>>, tensor<i32>, tensor<2xi32>) -> tensor<1x32xf32>
  %2 = "tf.TensorListGetItem"(%0, %arg3, %arg1) {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<!tf_type.variant<tensor<1x32xf32>>>, tensor<i32>, tensor<2xi32>) -> tensor<1x32xf32>

  %8 = "tfl.concatenation"(%2, %0) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x2xf32>
  %9 = "quantfork.stats"(%8) {layerStats = dense<[-0.488159984, 0.189515018]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  %10 = "tfl.concatenation"(%9, %7) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x4xf32>

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x2xbf16>) -> tensor<1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  return %3 : tensor<1x2xf32>
}

// CHECK: func @cast_bf16_matmul_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>
// CHECK: %[[matmul:.*]] = "tf.MatMul"(%arg0, %[[cst]])

  %1 = "tf.ReadVariableOp"(%0) {device = "/CPU:0"} : (tensor<!tf_type.resource<tensor<1x1xf32>>>) -> tensor<1x1xf32>
  %2:2 = "tf.Split"(%cst, %arg0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x4xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)
  %3:2 = "tf.Split"(%cst, %2#0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x2xf32>) -> (tensor<1x1xf32>, tensor<1x1xf32>)

func.func private @batched_function(%arg0: tensor<1x3xf32>, %arg1: tensor<*x!tf_type.resource>) -> tensor<1x3xf32> {
  %0 = "tf.Identity"(%arg0) : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %0 : tensor<1x3xf32>
}

// -----

func.func @test_batch_function_with_invalid_symbol(%arg0: tensor<1x3xf32>, %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>>) -> () {

For example, if we have the code

```mlir
  %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
```

then running this pass with 'default-device=foobar', we get:

```mlir

    %10 = "tf.Cast"(%9) {Truncate = false, device = ""} : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %11 = "tf.Mul"(%10, %cst) {device = ""} : (tensor<1x3xf32>, tensor<f32>) -> tensor<1x3xf32>
    %12 = "tf.Relu"(%11) {device = ""} : (tensor<1x3xf32>) -> tensor<1x3xf32>
    return %12 : tensor<1x3xf32>
  }
// CHECK-LABEL: func @matmul_with_relu

  %0 = "mhlo.reshape"(%arg0) : (tensor<1x1x512xf32>) -> tensor<1x512xf32>
  %1 = "mhlo.dot"(%0, %arg1) : (tensor<1x512xf32>, tensor<512x13x!quant.uniform<i8:f32, 0.00285>>) -> tensor<1x13xf32>
  %2 = "mhlo.reshape"(%1) : (tensor<1x13xf32>) -> tensor<1x1x13xf32>
  func.return %2 : tensor<1x1x13xf32>

// CHECK:      %[[RES:.*]] = "mhlo.dot_general"(%arg0, %arg1) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<

  // CHECK: return %[[v3:.*]] : tensor<5x7xf32>
}

func.func @einsum_transpose_matmul(%arg0: tensor<2x5x7xf32>, %arg1: tensor<5x3x2xf32>) -> tensor<5x3x7xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "lbh,bkl->bkh"}: (tensor<2x5x7xf32>, tensor<5x3x2xf32>) -> tensor<5x3x7xf32>
  func.return %0 : tensor<5x3x7xf32>
  // CHECK-LABEL: einsum_transpose_matmul