%cst = arith.constant dense<1> : tensor<1xi32>
  %0 = "tfl.exp"(%arg0) : (tensor<8x128xf32>) -> tensor<8x128xf32>
  %1 = "tfl.sum"(%0, %cst) {keep_dims = true} : (tensor<8x128xf32>, tensor<1xi32>) -> tensor<8x1xf32>
  %2 = "tfl.div"(%0, %1) {fused_activation_function = "NONE"} : (tensor<8x128xf32>, tensor<8x1xf32>) -> tensor<8x128xf32>
  func.return %2 : tensor<8x128xf32>

      %17 = "tf.Concat"(%cst_1, %16#5, %16#15, %16#25, %16#35) : (tensor<i32>, tensor<32x1024xf32>, tensor<32x1024xf32>, tensor<32x1024xf32>, tensor<32x1024xf32>) -> tensor<128x1024xf32>

constexpr absl::string_view kModuleStatic = R"mlir(
  module {
    func.func @main(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

    %19 = stablehlo.convert %18 : (tensor<8x16x4xi32>) -> tensor<8x16x4xf32>
    %20 = stablehlo.broadcast_in_dim %9, dims = [0, 1, 2] : (tensor<1x1x1xf32>) -> tensor<8x16x4xf32>
    %21 = stablehlo.multiply %19, %20 : tensor<8x16x4xf32>  // * s1 s2
    %22 = call @uniform_quantize_1(%21, %7, %8) : (tensor<8x16x4xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<8x16x4xi8>

  module {
    func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

// CHECK-SAME:                      %[[VAL_0:.*]]: tensor<8x128xf32>,
// CHECK-SAME:                      %[[VAL_1:.*]]: tensor<f32>) -> tensor<7x128xf32> {
// CHECK:           %[[VAL_2:.*]] = arith.constant dense<{{\[\[}}0, 0], [0, 1]]> : tensor<2x2xi64>
// CHECK:           %[[VAL_3:.*]] = "tf.PadV2"(%[[VAL_0]], %[[VAL_2]], %[[VAL_1]]) : (tensor<8x128xf32>, tensor<2x2xi64>, tensor<f32>) -> tensor<8x129xf32>