// CHECK: %[[RES_INT_2:.*]] = mhlo.convert %[[RES_INT_1]] : tensor<3x2xi32>
  %result_int = "tf.Cast"(%result) {Truncate = false} : (tensor<3x2x!tf_type.qint32>) -> tensor<3x2xi32>
  // CHECK: return %[[RES_INT_2]] : tensor<3x2xi32>
  func.return %result_int : tensor<3x2xi32>

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  // CHECK-NEXT: [[r1:%.*]] = tfrt_fallback_async.executeop {{.*}} "tf.BiasAdd"([[r0]], [[result]])
  %3 = "tf.BiasAdd"(%2, %0) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], data_format = "NHWC", device = "/device:CPU:0"} : (tensor<3x3xf32>, tensor<3xf32>) -> tensor<3x3xf32>

      } : (tensor<1xf32>, tensor<f32>, tensor<i32>) -> tensor<1x!tf_type.qint8>
      %1 = "tf.UniformDequantize"(%0, %scales, %zps) {
        quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64
      } : (tensor<1x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1xf32>
      func.return %1 : tensor<1xf32>
    }
  })mlir";

  // CHECK: return %[[dq3]]
  func.return %2 : tensor<?x2xf32>

// fixed upstream).
func.func @broadcast_add(%arg0: tensor<1xi32>, %arg1: tensor<1x2xi32>) -> tensor<1x2xi32> {
  // CHECK-NEXT: %[[LHS_BCAST:.+]] = "mhlo.broadcast_in_dim"(%arg0) <{broadcast_dimensions = dense<1> : tensor<1xi64>}>
  // CHECK-NEXT: mhlo.add %[[LHS_BCAST]], %arg1
  %0 = "tf.AddV2"(%arg0, %arg1) : (tensor<1xi32>, tensor<1x2xi32>) -> tensor<1x2xi32>
  func.return %0: tensor<1x2xi32>
}

// CHECK-LABEL: func @broadcast_multi_dim_add

    "tf.AssignVariableOp"(%arg0, %3) <{validate_shape = false}> : (tensor<*x!tf_type.resource<tensor<1xf32>>>, tensor<1xf32>) -> ()
    tf_device.return
  }
  func.return
}

func.func private @write_chain_func(%arg0: tensor<1xf32>) -> (tensor<1xf32>) {
  %cst = "tf.Const"() <{value = dense<[[0, 1]]> : tensor<1x2xi32>}> : () -> tensor<1x2xi32>

  // CHECK: %[[OLD_SLICE2:.*]] = "tf.Slice"(%[[READ2]],
  // CHECK: %[[RESHAPE2:.*]] = "tf.Reshape"(%[[VALUE]],
  // CHECK: %[[ADD2:.*]] = "tf.AddV2"(%[[RESHAPE2]], %[[OLD_SLICE2]]) : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  // CHECK: %[[UPDATE2:.*]] = "tf.XlaDynamicUpdateSlice"(%[[READ2]], %[[ADD2]],

  func.return %emb : tensor<2x2xf32>

// CHECK-NEXT: tfl.pseudo_qconst
// CHECK-NEXT: tfl.gather
// CHECK-NEXT: tfl.dequantize
}

// CHECK-LABEL: no_pushdown_multiple_inputs
func.func @no_pushdown_multiple_inputs(%arg0: tensor<1000x1000x!quant.uniform<i8:f32, 7.812500e-03>>, %arg1: tensor<1000x1000xf32>) -> tensor<2000x1000xf32> {