auto deallocator = [](void* data, size_t len, void* arg) {};
  unique_tensor_ptr t_in(TF_NewTensor(TF_STRING, nullptr, 0, &tstr[0],
                                      sizeof(tstr), deallocator, nullptr),
                         TF_DeleteTensor);

  // Create a TF_Operation with the attribute t_in
  auto desc = init("tensor");
  TF_SetAttrTensor(desc, "v", t_in.get(), s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

namespace odml {

using mhlo::CompositeOp;

// Struct for holding composite attrs for torch average pool as CC types.
struct TorchAvgPoolData {
  int n;  // Batch.
  int c;  // Channels.

  int h_in;  // Input height.
  int w_in;  // Input width.

  int h_out;  // Output height.
  int w_out;  // Output width.

  int kh;  // Kernel height.
  int kw;  // Kernel width.

	if firstTime {
		delta /= damp
	} else {
		delta /= 2
	}
	delta += delta / numPoints
	k := int32(0)
	for delta > ((base-tmin)*tmax)/2 {
		delta /= base - tmin
		k += base
	}
	return k + (base-tmin+1)*delta/(delta+skew)
}

// Strictly speaking, the remaining code below deals with IDNA (RFC 5890 and
// friends) and not Punycode (RFC 3492) per se.

	if firstTime {
		delta /= damp
	} else {
		delta /= 2
	}
	delta += delta / numPoints
	k := int32(0)
	for delta > ((base-tmin)*tmax)/2 {
		delta /= base - tmin
		k += base
	}
	return k + (base-tmin+1)*delta/(delta+skew)

  font-style: normal;
  font-weight: 100;
  src: local('Roboto Thin'), local('Roboto-Thin'), url(https://fonts.gstatic.com/s/roboto/v19/KFOkCnqEu92Fr1MmgVxEIzIXKMnyrYk.woff2) format('woff2');
  unicode-range: U+1F00-1FFF;
}
/* greek */
@font-face {
  font-family: 'Roboto';
  font-style: normal;
  font-weight: 100;

      when {
        first -> delta / DAMP
        else -> delta / 2
      }
    delta += (delta / numpoints)
    var k = 0
    while (delta > ((BASE - TMIN) * TMAX) / 2) {
      delta /= (BASE - TMIN)
      k += BASE
    }
    return k + (((BASE - TMIN + 1) * delta) / (delta + SKEW))
  }

  private fun String.requiresEncode(
    pos: Int,
    limit: Int,
  ): Boolean {
    for (i in pos until limit) {

      batch_timeout_micros = 1000000,
      container = "container",
      shared_name = "shared_name",
      batching_queue = "batching_queue",
      enable_large_batch_splitting = false,
      Tin = [i32],
      Tcaptured = [i32],
      Tout = [i32]} : 1

  %result_2 = tfrt_fallback_async.batch_function device("/device:CPU:0") @matmul_cpu (%a2, %b) {
      num_batch_threads = 1,
      max_batch_size = 4,

    value {
      shape {
      }
    }
  }
  attr {
    key: "shared_name"
    value {
      s: ""
    }
  }
}
node {
  name: "Call"
  op: "PartitionedCall"
  attr {
    key: "Tin"
    value {
      list {
      }
    }
  }
  attr {
    key: "Tout"
    value {
      list {
        type: DT_RESOURCE
      }
    }
  }
  attr {
    key: "f"
    value {

    for (int i = 0; i < TF_TensorElementCount(safe_input_ptr.get()); ++i) {
      const tensorflow::tstring& s_in = tags_array[i];
      tensorflow::Summary summary_in;
      if (!tensorflow::ParseProtoUnlimited(&summary_in, s_in)) {
        TF_SetStatus(status.get(), TF_INVALID_ARGUMENT,
                     "Could not parse one of the summary inputs");

    // CHECK-NEXT: op: "PartitionedCall"
    // CHECK:   func {
    // CHECK:     name: "foo"
    // CHECK:   }
    // CHECK: }
    %0 = tf_executor.island wraps "tf.PartitionedCall"() {Tin = [], Tout = [], config = "", config_proto = "", device = "", executor_type = "", f = @foo, name = "Call_foo"} : () -> ()
    tf_executor.fetch
  }
  func.return
}

// CHECK:      library {
// CHECK:        function {