%1 = arith.constant dense<[[2.3, 3.4, 4.5, 5.5]]> : tensor<1x4xf32>
    %2 = "tf.Add"(%0, %1) : (tensor<1x4xf32>, tensor<1x4xf32>) -> tensor<1x4xf32>
    %3 = tensor.cast %2 : tensor<1x4xf32> to tensor<1x?xf32>
    func.return %3 : tensor<1x?xf32>
}

func.func @func_0_CPU(%arg0: tensor<10x10x10xf32>, %arg1: tensor<10xf32>) -> tensor<10x10x10xf32> attributes {tac.device = "CPU", tac.interface_name = "func_0"} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  %1 = "tfl.mul"(%0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>

module {
  func.func @wrap_composite_func(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
    %0 = "tf.PartitionedCall"(%arg0, %arg1) <{f = @composite_conv2d_with_relu6_fn}> {_tfl_quant_trait = "fully_quantizable"}
          : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

  func.func @no_composite_func(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {

// PerChannel-DAG: %[[scale2:.*]] = "tf.Const"() <{value = dense<{{[0-9\.Ee\+\-]+}}> : tensor<6xf32>}> : () -> tensor<6xf32>
// PerChannel-DAG: %[[zp1:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<3xi32>}> : () -> tensor<3xi32>
// PerChannel-DAG: %[[zp2:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<6xi32>}> : () -> tensor<6xi32>

                             %arg2: tensor<5xf32> {tf._user_specified_name = "b1"},
                             %arg3: tensor<3920x10xf32> {tf._user_specified_name = "w2"},
                             %arg4: tensor<10xf32> {tf._user_specified_name = "b2"}) -> tensor<10xf32>

  // CHECK-NEXT:  buffer: 6
  // CHECK-NEXT:  } ]
  // CHECK-NEXT:  signature_defs: [ ]
  // CHECK-NEXT:}

  %0 = "tfl.pseudo_const" () {value = dense<-1.23697901> : tensor<32xf32>} : () -> tensor<32xf32> loc("Const")

    quantization_max_val = 127 : i64
  } : (tensor<10xf32>, tensor<f32>, tensor<i32>) -> tensor<10x!tf_type.qint8>
  %1 = "tf.UniformDequantize"(%0, %scale, %zp) {
    quantization_axis = -1 : i64,
    quantization_min_val = -128 : i64,
    quantization_max_val = 127 : i64
  } : (tensor<10x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<10xf32>
  return %1 : tensor<10xf32>
})mlir";

  %0 = stablehlo.constant dense<2.000000e+00> : tensor<4x2x3x3xf32>  // weight
  %1 = stablehlo.constant dense<3.000000e+00> : tensor<4xf32>  // bias
  %2 = stablehlo.constant dense<0.000000e+00> : tensor<1x4x5x5xf32>  // relu
  %3 = stablehlo.broadcast_in_dim %1, dims = [1] : (tensor<4xf32>) -> tensor<1x4x5x5xf32>

  EXPECT_THAT(num_bytes, Eq(3));
}

TEST_F(GetSizeInBytesTest, Float32ConstOpSizeInBytes) {
  constexpr absl::string_view kConstOpExpr =
      R"mlir(%cst = "tf.Const"() {value = dense<3.0> : tensor<4xf32>} : () -> tensor<4xf32>)mlir";

  Block block{};
  TF::ConstOp int_tensor_const_op = ParseConstOp(kConstOpExpr, block, ctx_);

  const int64_t num_bytes = GetSizeInBytes(int_tensor_const_op);
  EXPECT_THAT(num_bytes, Eq(16));
}

// expected-error@-1 {{'tf_executor.Merge' op expects all operands to be broadcastable with output type but got 'tensor<4xf32>' vs 'tensor<8xf32>'}}
    tf_executor.fetch %value : tensor<8xf32>
  }
  func.return %result : tensor<8xf32>
}

// -----

// Check that merge data inputs of variant type are broadcastable to the output