tensorflow/models · error

Failed to clean up TemporaryDirectory

Error message

Failed to clean up TemporaryDirectory

What it means

Error "Failed to clean up TemporaryDirectory" thrown in tensorflow/models.

Source

Thrown at official/nlp/tools/export_tfhub_lib.py:429

                         tokenize_with_offsets: bool,
                         default_seq_length: int,
                         experimental_disable_assert: bool = False) -> None:
  """Exports preprocessing to a SavedModel for TF Hub."""
  with tempfile.TemporaryDirectory() as tmpdir:
    # TODO(b/175369555): Remove experimental_disable_assert and its use.
    with _maybe_disable_assert(experimental_disable_assert):
      preprocessing = create_preprocessing(
          vocab_file=_move_to_tmpdir(vocab_file, tmpdir),
          sp_model_file=_move_to_tmpdir(sp_model_file, tmpdir),
          do_lower_case=do_lower_case,
          tokenize_with_offsets=tokenize_with_offsets,
          default_seq_length=default_seq_length)
      preprocessing.save(export_path, include_optimizer=False, save_format="tf")
    if experimental_disable_assert:
      _check_no_assert(export_path)
  # It helps the unit test to prevent stray copies of the vocab file.
  if tf.io.gfile.exists(tmpdir):
    raise IOError("Failed to clean up TemporaryDirectory")


# TODO(b/175369555): Remove all workarounds for this bug of TensorFlow 2.4
# when this bug is no longer a concern for publishing new models.
# TensorFlow 2.4 has a placement issue with Assert ops in tf.functions called
# from Dataset.map() on a TPU worker. They end up on the TPU coordinator,
# and invoking them from the TPU worker is either inefficient (when possible)
# or impossible (notably when using "headless" TPU workers on Cloud that do not
# have a channel to the coordinator). The bug has been fixed in time for TF 2.5.
# To work around this, the following code avoids Assert ops in the exported
# SavedModels. It monkey-patches calls to tf.Assert from inside TensorFlow and
# replaces them by a no-op while building the exported model. This is fragile,
# so _check_no_assert() validates the result. The resulting model should be fine
# to read on future versions of TF, even if this workaround at export time
# may break eventually. (Failing unit tests will tell.)


def _dont_assert(condition, data, summarize=None, name="Assert"):

View on GitHub (pinned to e006f5f0d5)

When it happens

Trigger: Thrown at official/nlp/tools/export_tfhub_lib.py:429 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of tensorflow/models@e006f5f0d5 (2026-08-24). Data as JSON: /api/errors/b07955bfd55d48c4. Report an issue: GitHub.