%0 = "tf.MatMul"(%input, %filter_new) {
  } : (tensor<8x9xf32>, tensor<9x10xf32>) -> tensor<8x10xf32>
  return %0 : tensor<8x10xf32>
})mlir";
  constexpr absl::string_view kProgram = R"mlir(
func.func @main(%input: tensor<8x9xf32>, %filter: tensor<9x10xi8>) -> tensor<8x10xf32> {
  %filter_scale = "tf.Const"() { value = dense<0.0235> : tensor<f32> } : ()
  -> tensor<f32>

func.func @softmax(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %0 = "tf.Softmax"(%arg0) : (tensor<8x16xf32>) -> tensor<8x16xf32>
  func.return %0 : tensor<8x16xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexSoftmax", custom_option = #tfl<const_bytes : "0x07536F66746D617800161207536F66746D61781A002A070A0154120230013200000221191414042801">}> : (tensor<8x16xf32>) -> tensor<8x16xf32>

  %group_assignment = "tf.Const"() { value = dense<[[0],[1]]> : tensor<2x1xi32> } : () -> tensor<2x1xi32>
  // CHECK: "mhlo.all_reduce"
  // CHECK{LITERAL}: replica_groups = dense<[[0], [1]]> : tensor<2x1xi64>
  // CHECK-NOT: channel_handle
  // CHECK: mhlo.add
  %0 = "tf.XlaAllReduce"(%input, %group_assignment) {reduce_op = "Add", mode = "CrossReplica"} : (tensor<f32>, tensor<2x1xi32>) -> tensor<f32>
  func.return %0 : tensor<f32>
}

// -----

  // CHECK-DAG: %[[VAL0:.+]] = "tfl.pseudo_qconst"() <{qtype = tensor<16x{{.+}}>, value = dense<1> : tensor<16xi32>}>
  // CHECK-DAG: %[[VAL1:.+]] = "tfl.pseudo_qconst"() <{qtype = tensor<16x{{.+}}>, value = dense<2> : tensor<16xi32>}>
  // CHECK-DAG: %[[VAL2:.+]] = "tfl.dequantize"(%[[VAL0]])
  // CHECK-DAG: %[[VAL3:.+]] = "tfl.dequantize"(%[[VAL1]])

    %14 = stablehlo.subtract %12, %13 : tensor<8x16x16xi32>  // q1 - z1
    %15 = stablehlo.convert %11 : (tensor<8x16x4xi8>) -> tensor<8x16x4xi32>
    %16 = stablehlo.broadcast_in_dim %6, dims = [0, 1, 2] : (tensor<1x1x1xi32>) -> tensor<8x16x4xi32>
    %17 = stablehlo.subtract %15, %16 : tensor<8x16x4xi32>  // q2 - z2
    // Corresponds to einsum expression: b i j, b j d -> b i d

  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<16x8xf32>) -> tensor<8x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({
  ^bb0(%arg1: tensor<f32>, %arg2: tensor<f32>):
      %3 = stablehlo.maximum %arg1, %arg2 : tensor<f32>
      stablehlo.return %3 : tensor<f32>
  }) {window_dimensions = array<i64: 2, 2>, window_strides = array<i64: 2, 2>} : (tensor<8x16xf32>, tensor<f32>) -> tensor<4x8xf32>
  return %2 : tensor<4x8xf32>
}

    "tfl.yield"(%1#0, %1#1, %1#2) : (tensor<1x?x16xf32>, tensor<1x!tf_type.string>, tensor<1xi32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable"} : (tensor<1x!tf_type.string>) -> (tensor<1x?x16xf32>, tensor<1x!tf_type.string>, tensor<1xi32>)

  %cst = arith.constant dense<[[1], [2]]> : tensor<2x1xi32>

  // CHECK: %[[CST:.*]] = arith.constant dense<2> : tensor<1xi32>
  // CHECK: return %[[CST]]
  %0 = "tfl.rank"(%cst) : (tensor<2x1xi32>) -> tensor<1xi32>
  func.return %0 : tensor<1xi32>
}

// CHECK-LABEL: @rank_input_known_rank
func.func @rank_input_known_rank(%arg0 : tensor<2x1xi32>) -> tensor<1xi32> {

  // CHECK:  %5 = "tfl.arg_max"(%arg0, %cst) : (tensor<1x32x1xf32>, tensor<1xi32>) -> tensor<1x1xi32>
  // CHECK:  return %4, %5 : tensor<1x1xf32>, tensor<1x1xi32>
}

// CHECK-LABEL:   func @convert_pytorch_argmax
func.func @convert_pytorch_argmax(%arg0: tensor<1x9xi32>) -> tensor<1xi32> {
  %0 = mhlo.constant dense<0> : tensor<i32>

    // CHECK-NOT: tf.Multinomial
    %samples = "tf.Const"() { value = dense<10> : tensor<i32> } : () -> tensor<i32>
    %1 = "tf.Multinomial"(%arg0, %samples) {seed = 0, seed2 = 0}: (tensor<2x4xf32>, tensor<i32>) -> tensor<2x10xi32>
    func.return %1 : tensor<2x10xi32>
  }

  // CHECK-LABEL: @set_dynamic_dimension_size