// CHECK-SAME: dilations = [1, 3, 4, 2]
  // CHECK-SAME: explicit_paddings = [1, 2, 5, 6, 7, 8, 3, 4]
  // CHECK-SAME: padding = "EXPLICIT"
  // CHECK-SAME: strides = [5, 7, 8, 6]
  // CHECK-SAME: (tensor<1x32x32x3xf32>, tensor<1x1x3x8xf32>) -> tensor<1x7x6x8xf32>

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: %[[RES_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%[[CONV2D]], %[[RES_PERM]])

      tf_executor.fetch %outputs_2 : tensor<*xi1>
    }
    return %0 : tensor<*xi1>
  }
  func.func private @image_sample_write_summary_summary_cond_false_710(%arg0: tensor<3x32x32x3xf32>, %arg1: tensor<*x!tf_type.resource>, %arg2: tensor<i64>) -> tensor<*xi1> attributes {tf._construction_context = "kEagerRuntime"} {
    %0 = tf_executor.graph {

    "tfl.yield"(%4, %5, %6, %7) : (tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>) -> ()
  }) : (tensor<1x64x64x32xf32>, tensor<1x64x64x32xf32>, tensor<1x64x64x32xf32>) -> (tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>)

  func.func @conv2d_unranked_filter_and_dynamic_batch(%arg0: tensor<?x32x32x3xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
    // CHECK: "tf.Conv2D"
    // CHECK-SAME: -> tensor<?x?x?x?xf32>
    %0 = "tf.Conv2D"(%arg0, %arg1) {padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<?x32x32x3xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

    ```

    will be replaced by `_xla_compile_device_type`, with its value set to the value of `device`.

    ```mlir

```

will be replaced by `_xla_compile_device_type`, with its value set to the value of `device`.

```mlir

  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<?x32x32x6xf32>, tensor<3x3x3x16xf32>) -> tensor<?x8x7x16xf32>
  func.return %0 : tensor<?x8x7x16xf32>
}