}

// failed inference
type Tf0[A, B any] int
type Tf1[A any, B ~struct {
	a A
	c C
}, C any] int

func _() {
	var _ Tf0 /* ERROR "not enough type arguments for type Tf0: have 1, want 2" */ [int]
	var _ Tf1 /* ERROR "not enough type arguments for type Tf1: have 1, want 3" */ [int]

	// floating point values
	tf0 float32 = 0.
	tf1 float32 = 1.
	tf2 float64 = 4.2e1
	tf3 myfloat = 3.141592653589793238462643383279502884197169399375105820974944592307816406286
	tf4 myfloat = 1e-1

	tf5 = tf0 + tf1
	tf6 = tf1 - tf1
	tf7 = tf2 /* ERROR "mismatched types" */ * tf1
	tf8 = tf3 / tf3
	tf9 = tf3 /* ERROR "not defined" */ % tf3

)
```

When this option is enabled, exported SavedModels with proto size > 2GB will
automatically save with the new format (`.cpb` instead of `.pb`).

<!-- **Compatibility** -->

The official TF APIs (TF1/TF2 python or C++ loading) have already been
integrated to handle the new format, but some downstream converters may not

namespace tensorflow {
namespace tf2xla {
namespace v1 {

// Run all the passes involved in transforming the graph before execution so
// that it is suitable for targeting devices when called via the TF1 Session
// API.
// These transformations take as input a Tensorflow Graph as an MLIR Module
// and transforms the module in place to cluster the given ops for compilation

# Use /tmp to make debugging the tests easier (see README.md)
flags.DEFINE_string('save_model_path', '', 'Path to save the model to.')
FLAGS = flags.FLAGS


def set_tf_options():
  # Default TF1.x uses reference variables that are not supported by SavedModel
  # v1 Importer. To use SavedModel V1 Importer, resource variables should be
  # enabled.
  tf.enable_resource_variables()
  tf.compat.v1.disable_eager_execution()

// function will respect the request. If the user does not explicitly request
// enabled or disabled, it will decide whether or not to run the bridge.
//
// The config_proto param is a required input for all TF1 graphs but it is
// redundant for TF2 graphs.
// If getting rollout policy involves graph analysis, `record_stats` is used
// to decide whether to emit metrics on unsupported features of the graph.

def _save_function_alias(
    saved_model_dir: str,
    tags: Collection[str],
    function_aliases: Mapping[str, str],
) -> None:
  """Saves the function alias to the SavedModel.

  SavedModelBuilder (TF1 saved model saver) does not support saving function
  aliases, so this function loads the SavedModel proto and adds the
  `function_aliases` field.

  Args:
    saved_model_dir: Path to the saved model directory.

  # `LOG(ERROR)`, so this makes them print out by default.
  logging.set_stderrthreshold('error')

  # In true TF2 releases, v2 behavior is enabled as part of module __init__. In
  # TF1 builds, it must be enabled manually. If you get an error here,
  # it means that TF was used in V1 mode prior to calling this.
  tf.enable_v2_behavior()

  def app_main(argv):
    """Function passed to absl.app.run."""

    Raises:
      KeyError: If the set of input keys in the dataset samples doesn't match
      the set of expected input keys.
    """
    # When running in graph mode (TF1), tf.Tensor types should be converted to
    # numpy ndarray types to be compatible with `make_tensor_proto`.
    if not context.executing_eagerly():
      with session.Session() as sess:

    representative_dataset: rd.RepresentativeDataset,
) -> None:
  """Runs the representative dataset through a function for calibration.

  NOTE: This is intended to be run in graph mode (TF1).

  The function is identified by the SignatureDef.

  Args:
    sess: The Session object to run the function in.
    signature_def: A SignatureDef that identifies a function by specifying the