// Prepare function arguments from ready Chain and input Tensors.
  llvm::SmallVector<tfrt::AsyncValue*> exec_arguments;
  exec_arguments.reserve(compute->num_arguments());
  exec_arguments.push_back(tfrt::GetReadyChain().release());
  for (const Tensor& input_tensor : arguments) {
    auto av = MakeAvailableAsyncValueRef<FallbackTensor>(input_tensor);

      %replicated_output:2 = "tf.TPUReplicatedOutput(%identity) : (tensor<i32>) -> (tensor<i32>, tensor<i32>)
      return %replicated_output#0, %replicated_output#1 : tensor<i32>, tensor<i32>
    }
    ```

    will be transformed into:

    ```mlir
    func @tpu_computation(%arg0: tensor<i32>, %arg1: tensor<i32>) -> (tensor<i32>, tensor<i32>) {

  // pointer. We can return an empty object and ignore num_bytes here because we
  // have control over all of the uses of this device tensor, and can lazily
  // allocate memory when used. This allows us to also know the shape of the
  // allocated Tensor, which is useful if the device's tensor representation
  // differs from the host.
  return XlaTensor::ToOpaquePointer(new XlaTensor());
}

  // TF2 models.
  map<string, string> function_aliases = 6;

  // Holds information about the asset files used for the model. It essentially
  // associates asset file names with the tensors to which the asset file names
  // should be fed.
  repeated AssetFileDef asset_file_defs = 8;

  // SaverDef including the information required for saving and restoring

}

LogicalResult VerifyWhitespaceTokenizer(func::FuncOp func) {
  // In the case of input tensor with 0 rank.
  // Whitespace tokenizer generates 1 output:
  // * String tensor for tokens.
  //
  // In the case of 1-D input tensor,
  // Whitespace tokenizer generates 2 outputs to make up a ragged tensor:
  // * 1st output is the value of ragged tensor;
  // * 2nd output is the offset.
  //
  // In the case of batched input tesnor,

// TF_DataType holds the type for a scalar value.  E.g., one slot in a tensor.
// The enum values here are identical to corresponding values in types.proto.
typedef enum TF_DataType {
  TF_FLOAT = 1,
  TF_DOUBLE = 2,
  TF_INT32 = 3,  // Int32 tensors are always in 'host' memory.
  TF_UINT8 = 4,
  TF_INT16 = 5,
  TF_INT8 = 6,
  TF_STRING = 7,

    }
  }

  // Verify the argument order: tensors, tensor list, attributes; and also
  // verify there is at most one tensor list argument.
  if (first_attr != -1 &&
      (first_attr < last_tensor_list || first_attr < last_tensor)) {
    func.emitError(
        "tfr.tensor/tfr.tensor_list argument should be before non tensor "
        "arguments.");
    return failure();
  }

namespace tensorflow {

namespace {

// Gets tensor names from tensor_info and inserts them into the set of tensor
// names.
void GetTensorNamesFromTensorInfo(const TensorInfo& tensor_info,
                                  std::unordered_set<string>* tensor_names) {
  if (tensor_info.has_coo_sparse()) {
    // If the tensor is sparse we have to add all three tensors of the sparse
    // representations.

  }
  return %graph#0, %graph#1, %graph#2, %graph#3 : !tf_res, !tf_res, tensor<f32>, tensor<f32>
}
```

After:

```mlir
func @while_body(%arg0: !tf_res, %arg1: !tf_res, %arg2: tensor<f32>, %arg3: tensor<f32>, %chain_0: tensor<i32>, %chain_1: tensor<i32>) -> (!tf_res, !tf_res, tensor<f32>, tensor<f32>, tensor<i32>, tensor<i32>) {
  %graph:6 = tf_executor.graph {

      }
    }

    // Special type fixes for scalar tensor types.
    // TFLite flatbuffer schema doesn't distinguish scalar tensor shapes
    // and unranked tensor shapes (i.e. they are both represented as an empty
    // INT32 list), see b/138865275. MLIR importer conservatively treats them as
    // unranked tensor types. Here we set them to scalar tensor types when it is
    // safe.