func.func @testConv2dShapeInferenceDynamic(%arg0: tensor<1x?x?x128xf32>, %arg1: tensor<128x3x3x128xf32>, %arg2: tensor<128xf32>) -> tensor<1x?x?x128xf32> {
  // CHECK: "tfl.conv_2d"(%arg0, %arg1, %arg2) <{dilation_h_factor = 2 : i32, dilation_w_factor = 2 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32}> : (tensor<1x?x?x128xf32>, tensor<128x3x3x128xf32>, tensor<128xf32>) -> tensor<1x?x?x128xf32>

// CHECK-LABEL: matmulBiasAdd
func.func @matmulBiasAdd(%arg0: tensor<64xf32>, %arg1: tensor<8x32xf32>, %arg2: tensor<32x64xf32>) -> (tensor<*xf32>) {
  // CHECK: %[[VAL_3:.*]] = "tf._FusedMatMul"(%arg1, %arg2, %arg0) <{epsilon = 0.000000e+00 : f32, fused_ops = ["BiasAdd"], transpose_a = false, transpose_b = false}> : (tensor<8x32xf32>, tensor<32x64xf32>, tensor<64xf32>) -> tensor<*xf32>

// CHECK: return %[[v5]] : tensor<?x?x8xf32>
}

func.func @einsum_no_reshape(%arg0 : tensor<?x?x8x128xf32>, %arg1 : tensor<1x?x8x128xf32>) -> tensor<?x?x8x?xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {device = "", equation = "BQNH,BTNH->BQNT"} : (tensor<?x?x8x128xf32>, tensor<1x?x8x128xf32>) -> tensor<?x?x8x?xf32>
  func.return %0 : tensor<?x?x8x?xf32>
// CHECK-LABEL: einsum_no_reshape

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

func.func @softmax(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %0 = "tf.Softmax"(%arg0) : (tensor<8x16xf32>) -> tensor<8x16xf32>
  func.return %0 : tensor<8x16xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexSoftmax", custom_option = #tfl<const_bytes : "0x07536F66746D617800161207536F66746D61781A002A070A0154120230013200000221191414042801">}> : (tensor<8x16xf32>) -> tensor<8x16xf32>

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  %2:2 = "tf.IteratorGetNext"(%arg0) {device = "/device:CPU:0"} : (tensor<*x!tf_type.resource>) -> (tensor<128xf32>, tensor<128xf32>)
  %3:2 = "tf.IteratorGetNext"(%arg1) {device = "/device:CPU:0"} : (tensor<*x!tf_type.resource>) -> (tensor<128xf32>, tensor<128xf32>)
  // CHECK: "tf.TPUCompileSucceededAssert"(%[[COMPILE]]#0)
  "tf_device.launch"() ({

  %w = arith.constant dense<127.0> : tensor<4x12xf32>
  %b = arith.constant dense<0.0> : tensor<4xf32>
  %fc = "tfl.fully_connected"(%arg0, %w, %b) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<4x12xf32>, tensor<4xf32>) -> tensor<1x112x112x4xf32>
  func.return %fc : tensor<1x112x112x4xf32>