"tfl.custom"(%0, %2) {custom_code = "FlexAssignVariableOp", custom_option = #tfl<const_bytes : "0x1041737369676E5661726961626C654F70003B121041737369676E5661726961626C654F701A001A002A0B0A056474797065120230012A140A0E76616C69646174655F736861706512022800320000024F3E1414042801">} : (tensor<!tf_type.resource<tensor<1x10xf32>>>, tensor<1x10xf32>) -> ()

    %13 = stablehlo.subtract %10, %12 : tensor<1x4x3xf32>  // Precalculated zp_neg.
    %14 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2] : (tensor<1x1x3xf32>) -> tensor<1x4x3xf32>  // Optional
    %15 = stablehlo.multiply %13, %14 : tensor<1x4x3xf32>  // s1 * s2
    %16 = call @uniform_quantize_1(%15, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

    %3 = "tf.Identity"(%2#1) {device = ""} : (tensor<1x3xf32>) -> tensor<1x3xf32>
    return %3 : tensor<1x3xf32>
  }

    %0 = "tf.Identity"(%cst) {device = ""} : (tensor<i1>) -> tensor<i1>
    %1 = "tf.PartitionedCall"(%output, %cst_1, %cst_0) <{config = "", config_proto = "", executor_type = "", f = @composite_matmul_with_bias_fn_1}> {_tfl_quant_trait = "fully_quantizable"} : (tensor<1x4xf32>, tensor<4x3xf32>, tensor<3xf32>) -> tensor<1x3xf32>

  func.return %24 : tensor<1x4xf32>

  func.return %24 : tensor<1x4xf32>

  // CHECK-LABEL: func.func private @merge_relu_fusion
  func.func private @merge_relu_fusion(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {
    %0 = stablehlo.constant() {value = dense<127> : tensor<4x3xi8>} : () -> tensor<4x3x!quant.uniform<i8<-127:127>:f32:1, {5.000000e-03,5.000000e-03,5.000000e-03}>>
    %1 = stablehlo.uniform_quantize %arg0 : (tensor<1x4xf32>) -> tensor<1x4x!quant.uniform<i8:f32, 6.000000e-03:-128>>
    // CHECK: call @quantized_dot_general_relu_fn

  %43 = "quantfork.stats"(%42) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x4x3xf32>) -> tensor<1x4x3xf32>
  %44 = "tfl.strided_slice"(%43, %1, %2, %3) {begin_mask = 7 : i32, ellipsis_mask = 0 : i32, end_mask = 5 : i32, new_axis_mask = 0 : i32, shrink_axis_mask = 0 : i32, offset = false} : (tensor<1x4x3xf32>, tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<1x2x3xf32>

// -----

// CHECK-LABEL: float_matmul
func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>