def Test():

  x = tf.constant([[1.0], [1.0], [1.0]])
  y = tf.compat.v1.get_variable(
      name='y',
      shape=(1, 3),
      initializer=tf.random_normal_initializer(),
      trainable=True)
  r = tf.matmul(x, y)

  tensor_info_x = tf.compat.v1.saved_model.utils.build_tensor_info(x)
  tensor_info_r = tf.compat.v1.saved_model.utils.build_tensor_info(r)

  return {

(TODO)

## Authoring Op Composition in Python

The composable TF provides a single API to define a new op with its composition
at the same time. For example, the following code defines a new
`FusedFullyConnected` op, which have `MatMul`, `Add` and some
`activation function` (specified by an op attribute) fused.


```python
import tensorflow as tf

@Composite(
    'FusedFullyConnected',
    inputs=['input_: T', 'filter_: T', 'bias: T'],

///     int idx = 3;
///     auto b = Variable(linear.WithOpName("b_", idx),
///                       {2}, DT_FLOAT);
///     auto x = Const(linear, {...});  // name: "linear/Const"
///     auto m = MatMul(linear, x, W);  // name: "linear/MatMul"
///     auto r = BiasAdd(linear, m, b); // name: "linear/BiasAdd"
///
/// Scope lifetime:
///
/// A new scope is created by calling Scope::NewRootScope. This creates some

                   absl::Span<AbstractTensorHandle* const> inputs,
                   absl::Span<AbstractTensorHandle*> outputs) {
  return ops::MatMul(ctx, inputs[0], inputs[1], &outputs[0],
                     /*transpose_a=*/false,
                     /*transpose_b=*/false, "MatMul");
}

Status MulModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,

// specs.
absl::flat_hash_set<int> GetQuantizableOperands(Operation* op) {
  absl::flat_hash_set<int> quantizable_operands;
  if (isa<TF::DepthwiseConv2dNativeOp, TF::Conv2DOp, TF::Conv3DOp, TF::MatMulOp,
          TF::BatchMatMulOp>(op)) {
    quantizable_operands.insert(1);
  } else if (isa<TF::GatherOp>(op)) {
    quantizable_operands.insert(0);
  } else if (auto einsum_op = dyn_cast<TF::EinsumOp>(op)) {

// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<-131072> : tensor<1x3xi32>}> : () -> tensor<1x3xi32>
// CHECK: %[[MATMUL:.*]] = "tf.XlaDotV2"({{.*}}, %[[WEIGHT]])
// CHECK-SAME: (tensor<1x1024xi8>, tensor<1024x3xi8>) -> tensor<1x3xi32>
// CHECK: %[[SUB:.*]] = "tf.Sub"(%[[MATMUL]], %[[CONST]]) : (tensor<1x3xi32>, tensor<1x3xi32>) -> tensor<1x3xi32>
}

// -----

module attributes {} {

      ]
  )
  def matmul(self, matmul_input: core.Tensor) -> Mapping[str, core.Tensor]:
    """Performs a matrix multiplication.

    Args:
      matmul_input: Input tensor to matmul with the filter.

    Returns:
      A map of: output key -> output result.
    """
    out = math_ops.matmul(matmul_input, self.matmul_filters)

    return {'output': out}

#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/util/equal_graph_def.h"

namespace tensorflow {
namespace {

using ops::Complex;
using ops::Const;
using ops::Div;
using ops::MatMul;
using ops::Placeholder;
using ops::Real;
using ops::Split;
using ops::Square;
using ops::Stack;
using ops::Sub;
using ops::Unstack;

TEST(GradientCheckerTest, BasicFloat) {

  %ch1 = tfrt.merge.chains %ch, %ch0 : !tfrt.chain, !tfrt.chain

  %ch2 = tfrt_fallback_async.createop(%ch1) key(0) device("/CPU:0") "tf.MatMul"() {T = i32} num_args(2)

  %ch3, %result = tfrt_fallback_async.executeop.seq(%ch2) key(0) cost(100) device("/CPU:0") "tf.MatMul"(%a, %b) {T = i32}  : 1

  %s = "tfrt_test.get_string"() { value = "Running @matmul_cpu" } : () -> !tfrt.string

def Test():

  x = tf.constant([[1.0], [1.0], [1.0]])
  y = tf.get_variable(
      name='y',
      shape=(1, 3),
      initializer=tf.random_normal_initializer(),
      trainable=True)
  r = tf.matmul(x, y)

  tensor_info_x = tf.saved_model.utils.build_tensor_info(x)
  tensor_info_r = tf.saved_model.utils.build_tensor_info(r)

  signature_def = tf.saved_model.signature_def_utils.build_signature_def(