%2 = "tf.Relu"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    %3 = "tf.FakeQuantWithMinMaxArgs"(%2) {device = "", max = 4.000000e-01 : f32, min = -3.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

    %0 = "quantfork.stats"(%arg1) {layerStats = dense<[-3.54062747, 0.54742622]> : tensor<2xf32>} : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2xf32>
    %1 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 2.824783]> : tensor<2xf32>} : (tensor<1x3x2x3xf32>) -> tensor<1x3x2x3xf32>

  %3 = "tf.BiasAdd"(%2, %cst_0) {data_format = "NHWC", device = ""} : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32>
  %4 = "tf.Relu"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  %5 = "quantfork.qcast"(%4) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>

  %3 = "tf.Cast"(%2) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %4 = "tf.IdentityN"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %4 : tensor<1x3x2x2xf32>
}

// CHECK: func @cast_bf16_conv_with_bias_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

        : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32> loc(fused["Conv2D:", "test_opt_out"])
    %2 = "tf.BiasAdd"(%1, %cst_0) : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32> loc(fused["BiasAdd:", "model/bias_add"])
    %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    return %3 : tensor<1x3x2x2xf32>
  }
}

    return %0 : tensor<1x3x4x2xf32>
  }
}

// CHECK-LABEL: quantize_conv_fn
// CHECK-SAME: %[[ARG0:.+]]: tensor<1x3x4x3xf32>
// CHECK: %[[CST:.+]] = stablehlo.constant dense<3.000000e-01> : tensor<2x3x3x2xf32>
// CHECK: %[[Q:.+]] = "quantfork.qcast"(%[[CST]]) : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2x!quant.uniform<i8:f32, 6.000000e-03:-128>>

      tf_executor.fetch %outputs#1, %outputs#0 : tensor<2x2x1x2xi32>, tensor<2x2x1x2xf32>
    }
    func.return %0#0, %0#1 : tensor<2x2x1x2xi32>, tensor<2x2x1x2xf32>
  }
}

// CHECK:        name: "serving_default_x"
// CHECK-NEXT:   op: "_Arg"

  // CHECK-NOT: stablehlo.uniform_quantize
  // CHECK: %[[DEQUANT:.*]] = stablehlo.uniform_dequantize %[[CONV]]
  // CHECK: return %[[DEQUANT]]
  %cst = stablehlo.constant dense<0.4> : tensor<2x3x3x2xf32>
  %quant_cst = stablehlo.uniform_quantize %cst : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2x!quant.uniform<i8<-127:127>:f32, 0.015>>
  %quant_arg = stablehlo.uniform_quantize %arg0 : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3x!quant.uniform<i8:f32, 0.0039207626791561354:-128>>

func.func @tpu_conv(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
  %0 = "tf.TPUOrdinalSelector"() {device = ""} : () -> tensor<?xi32>
  %1 = "tf.TPUPartitionedCall"(%arg0, %0) {autotuner_thresh = 0 : i64, device = "", f = @tpu_func_0_optim0} : (tensor<1x3x4x3xf32>, tensor<?xi32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %2 : tensor<1x3x2x2xf32>
}