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>

    %0:2 = tf_executor.island wraps "tf.RecvTPUEmbeddingActivations"() {config = "test_config_recv_embedding"} : () -> tensor<512x256xf32> loc("RecvTPUEmbedding")
    %1:1 = tf_executor.island wraps "tf.SendTPUEmbeddingGradients"(%0) {N = 1 : i64, NN = 0 : i64, config = "test_config_send_embedding", operandSegmentSizes = array<i32: 1, 0>} : (tensor<512x256xf32>) -> () loc("SendTPUEmbedding")
    tf_executor.fetch
  }
  func.return
}

func.func @fakeQuantFollowedByReshape(tensor<1x2xf32>, tensor<f32>, tensor<f32>) -> (tensor<2x1xf32>) {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<f32>, %arg2: tensor<f32>):
  %cst_0 = arith.constant dense<[2, -1]> : tensor<2xi64>
  %0 = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2) {num_bits = 3, narrow_range = false} : (tensor<1x2xf32>, tensor<f32>, tensor<f32>) -> tensor<1x2xf32>
  %1 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>

      return %1 : tensor<1x3xf32>
    }

    func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> {
      %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>

    // expected-remark@+1 {{Cost: 2}}
    %1 = "tf.ReadVariableOp"(%0) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<!tf_type.resource<tensor<512x512xf32>>>) -> tensor<512x512xf32>
    // 262657 = 1 + 512 + 512 * 512
    // expected-remark@+1 {{Cost: 262657}}

  %2 = "tf.Const"() {value = dense<0.0> : tensor<2x2xf32>} : () -> tensor<2x2xf32>
  %3 = "tf.Const"() {value = dense<0.0> : tensor<4x4xf32>} : () -> tensor<4x4xf32>
  "tf.SendTPUEmbeddingGradients"(%2, %3) {_tpu_embedding_layer = "call1", config = "\0A\0B\0C\0D", operandSegmentSizes = array<i32: 2, 0>} : (tensor<2x2xf32>, tensor<4x4xf32>) -> ()

// Legacy mode re-calculates zero point when it's changed due to subtle difference in scale.
func.func @ZeroPointLegacy(%arg0: tensor<1x2xf32>) -> tensor<1x2xf32>  {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[-1.0, 1.20779215]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  func.return %0 : tensor<1x2xf32>
// CHECK: %1 = "tfl.dequantize"(%0) : (tensor<1x2x!quant.uniform<i8:f32, 0.0086580084819419707:-12>>)

      func.func @main(%arg: tensor<1x2xf32> {tf_saved_model.index_path = ["input_tensor:0"]}) -> (tensor<1x2xf32> {tf_saved_model.index_path = ["output_tensor:0"]}) attributes {tf.entry_function = {inputs = "input_tensor:0", outputs = "output_tensor:0"}, tf_saved_model.exported_names = ["main"]} {
        %0 = tf_executor.graph {
          tf_executor.fetch %arg : tensor<1x2xf32>
        }
        return %0 : tensor<1x2xf32>
      }

      %2 = "tf.XlaCallModule"(%arg0, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_entry_function = @composite_dot_general_fn_1, _original_entry_function = "composite_dot_general_fn_1", _tfl_quant_trait = "fully_quantizable", _quantization_method = "weight_only_ptq { }"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>
      return %2 : tensor<?x2xf32>