%1:2 = "tf.IteratorGetNext"(%arg4) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<*x!tf_type.resource>) -> (tensor<2x224x224x3xf32>, tensor<2x1xf32>)

// CHECK-SAME: [[arg0:%.*]]: !tfrt_fallback.tf_tensor,
// CHECK-SAME: [[arg1:%.*]]: !tfrt_fallback.tf_tensor)
// CHECK-SAME: -> (!tfrt.chain, !tfrt_fallback.tf_tensor)
func.func @func_basic(
    %arg0: tensor<3x1xf32> {tf_saved_model.index_path = [0]},
    %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>> {tf_saved_model.index_path = [1]})
      -> (tensor<3x3xf32> {tf_saved_model.index_path = []})

        unknown_rank: true
      }
    }
  }
  attr {
    key: "dtype"
    value {
      type: DT_BOOL
    }
  }
}

# CHECK-DAG: tf.IteratorGetNext{{.+}}-> (tensor<1x8xbf16>, tensor<2x?x16xf32>, tensor<*xf64>)
# CHECK-DAG: tf.IteratorGetNextSync{{.+}}-> (tensor<*xi16>, tensor<3x24xi32>, tensor<?x4x32xi64>)

    "tfl.yield"(%1#0, %1#1, %1#2) : (tensor<1x?x16xf32>, tensor<1x!tf_type.string>, tensor<1xi32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable"} : (tensor<1x!tf_type.string>) -> (tensor<1x?x16xf32>, tensor<1x!tf_type.string>, tensor<1xi32>)

  // CHECK-DAG: %[[LOG_SOFTMAX_LOG:.*]] = "tf.Log"(%[[LOG_SOFTMAX_SUM]]) : (tensor<2x1xf32>) -> tensor<2x1xf32>
  // CHECK-DAG: %[[LOG_SOFTMAX:.*]] = "tf.Sub"(%[[LOG_SOFTMAX_SHIFTED]], %[[LOG_SOFTMAX_LOG]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>

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

func.func @cast_bf16_matmul_to_fp32(%arg0: tensor<1x10xf32>) -> (tensor<1x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<10x2xbf16>} : () -> tensor<10x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x10xf32>) -> tensor<1x10xbf16>

  %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<

  %1 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<16xf32>} : () -> tensor<16xf32>
  // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %2 = "tfl.conv_2d"(%arg0, %0, %1) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>

  // CHECK: return %[[v1]] : tensor<2x?x?x1xf32>
}

func.func @einsum_broadcast(%arg0: tensor<3x4x5xf32>, %arg1: tensor<5x6xf32>) -> tensor<3x4x6xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ijk,km->ijm"}: (tensor<3x4x5xf32>, tensor<5x6xf32>) -> tensor<3x4x6xf32>
  func.return %0 : tensor<3x4x6xf32>

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