%cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<10x2xbf16>} : () -> tensor<10x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x10xf32>) -> tensor<1x10xbf16>
  %1 = "tf.MatMul"(%0, %cst) {device = "", transpose_a = false, transpose_b = false} : (tensor<1x10xbf16>, tensor<10x2xbf16>) -> tensor<1x2xbf16>

  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_4:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_5:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>

  // MLIR @main function corresponds to the TF function "main_original".
  OwningOpRef<ModuleOp> module_op = ParseModuleOpString(R"mlir(
    func.func private @main(%arg: tensor<1x2xf32>) -> (tensor<1x2xf32>) attributes {tf._original_func_name = "main_original"} {
      return %arg : tensor<1x2xf32>
    }
  )mlir");
  ASSERT_TRUE(module_op);

  absl::flat_hash_map<FunctionName, FunctionAlias> function_aliases;

  constexpr absl::string_view kModuleWithCompositeDotGeneral = R"mlir(
    func.func @main(%arg0: tensor<1x2xf32>) -> tensor<1x3xf32> {
      %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3xf32>} : () -> tensor<2x3xf32>
      %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>

    return %1 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>

func.func @tfl_wrapped_jax_random_uniform(%arg0: tensor<2xui32>) -> tuple<tensor<1x2xf32>> {
  // This is a fake jax random uniform body.
  %0 = stablehlo.constant dense<0.0> : tensor<2xf32>
  %1 = "stablehlo.reshape"(%0) : (tensor<2xf32>) -> tensor<1x2xf32>
  %2 = "stablehlo.tuple"(%1) : (tensor<1x2xf32>) -> tuple<tensor<1x2xf32>>
  func.return %2 : tuple<tensor<1x2xf32>>

  func.return %dq : tensor<1x2xf32>
}

// -----

// CHECK-LABEL: @convert_quantfork_qdq_to_stablehlo_uniform_qdq
// CHECK-SAME: %[[ARG0:.*]]: tensor<1x3xf32>
// CHECK-SAME: %[[ARG1:.*]]: tensor<3x2xf32>
func.func @convert_quantfork_qdq_to_stablehlo_uniform_qdq(%arg0: tensor<1x3xf32>, %arg1: tensor<3x2xf32>) -> tensor<1x2xf32> {

    return %2 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>

  %0 = arith.constant dense<[[1.0, 2.0]]> : tensor<1x2xf32>
  %1 = "tfl.batch_matmul"(%arg0, %0) {adj_x = false, adj_y = true, asymmetric_quantize_inputs = false} : (tensor<4x128x2xf32>, tensor<1x2xf32>) -> tensor<4x128x1xf32>
  func.return %1 : tensor<4x128x1xf32>
  // CHECK: %[[CONST_WEIGHT:.*]] = arith.constant
  // CHECK-SAME: [1.000000e+00, 2.000000e+00]
  // CHECK-SAME: tensor<1x2xf32>

  // CHECK-DAG:  %[[CST_1:.*]] = "tfl.pseudo_const"() <{value = dense<0.000000e+00> : tensor<4x4xf16>}> : () -> tensor<4x4xf16>
  // CHECK-DAG:  %[[DQ_0:.*]] = "tfl.dequantize"(%[[CST_0]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>
  // CHECK-DAG:  %[[DQ_1:.*]] = "tfl.dequantize"(%[[CST_1]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>