func.func @dequantize_pushdown_gather_with_reduction(%arg0: tensor<2xi32>) -> tensor<2x2xf32> {
  %w = "tfl.pseudo_qconst"() {qtype = tensor<12x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>, value = dense<127> : tensor<12x2xi8>} : () -> tensor<12x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>
  %dq_w = "tfl.dequantize"(%w) : (tensor<12x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>) -> tensor<12x2xf32>

// WHOLE-PASSES-DAG:  %[[rv2:.*]] = "tfl.read_variable"({{.*}}) : (tensor<*x!tf_type.resource<tensor<1x2x3x!quant.uniform<u8:f32, {{.*}}>>>>) -> tensor<1x2x3x!quant.uniform<u8:f32, {{.*}}>>

// quantization type of other operands.
inline quant::AccumulatorScaleFunc GetUniformQuantizedTypeForBiasWithScale(
    double scale) {
  return [=](const std::vector<quant::QuantParams>& quant_params,
             const int adjusted_quant_dim,
             const bool legacy_float_scale) -> quant::QuantParams {
    if (auto qtype = mlir::dyn_cast_or_null<UniformQuantizedType>(
            quant::GetUniformQuantizedTypeForBias(

  mlir::registerTensorFlowPasses();
  mlir::quant::stablehlo::registerPasses();
  mlir::quant::stablehlo::registerBridgePasses();
  mlir::stablehlo::registerPasses();
  mlir::mhlo::registerAllMhloPasses();
  // These passes are only used for testing purposes.
  mlir::quant::stablehlo::testing::registerTestPasses();

  // Register StableHLO Quantizer pass pipelines.
  mlir::quant::stablehlo::RegisterPassPipelines();

  %4 = "tfl.reshape"(%3, %cst) : (tensor<1x112x112x32x!quant.uniform<u8:f32, 0.023528476789885875>>, tensor<2xi32>) -> tensor<1x401408x!quant.uniform<u8:f32, 0.023528476789885875>>
  %5 = "tfl.softmax"(%4) {beta = 1.000000e+00 : f32} : (tensor<1x401408x!quant.uniform<u8:f32, 0.023528476789885875>>) -> tensor<1x401408x!quant.uniform<u8:f32, 3.906250e-03>>

  %2 = "tfl.reshape"(%arg0, %0) : (tensor<4x384x32x!quant.uniform<i8:f32, 0.2:-3>>, tensor<4xi32>) -> tensor<1x4x384x32x!quant.uniform<i8:f32, 0.2:-3>>
  // CHECK-NOT: tac.device tac.inference_type
  %3 = "tfl.quantize"(%2) {qtype = tensor<1x4x384x32x!quant.uniform<i8:f32, 0.19:1>>} : (tensor<1x4x384x32x!quant.uniform<i8:f32, 0.2:-3>>) -> tensor<1x4x384x32x!quant.uniform<i8:f32, 0.19:1>>

  func.return %6 : tensor<1x112x112x32x!quant.uniform<u8:f32, 0.023528476789885875>>
}

// DEBUG-LABEL: QuantizeSplit
func.func @QuantizeSplit(%arg: tensor<4x!quant.uniform<u8:f32, 1.0>>, %cst: tensor<i32>) -> (tensor<2x!quant.uniform<u8:f32, 1.0>>,tensor<2x!quant.uniform<u8:f32, 1.0>>) {
  %0 = "tfl.dequantize"(%arg) : (tensor<4x!quant.uniform<u8:f32, 1.0>>) -> tensor<4xf32>

  func.return %0...

  %0 = "tfl.pseudo_qconst"() { qtype = tensor<3x3x!quant.uniform<i32:f32:0, {1.0, 0.5:1, 0.25:1}>>, value = dense<1> : tensor<3x3xi32>} : () -> tensor<3x3x!quant.uniform<i32:f32:0, {1.0, 0.5:1, 0.25:1}>>
  func.return %0 : tensor<3x3x!quant.uniform<i32:f32:0, {1.0, 0.5:1, 0.25:1}>>
}

func.func @qu8() -> tensor<3x!quant.uniform<u8<1:255>:f32, 1.0>> {
  // CHECK-LABEL: @qu8

// RUN: stablehlo-quant-opt %s -quant-convert-tf-quant-ops-to-mhlo | FileCheck %s

// CHECK-LABEL: func @quantized_matmul_fn
func.func @quantized_matmul_fn(%input: tensor<?x?xf32>) -> tensor<?x?xf32> {