tensorflow/models · error · ValueError

Distribution strategy has not been specified.

Error message

Distribution strategy has not been specified.

What it means

Guard in run_bert: no distribution strategy object was provided (strategy is None/falsy) before setting up distributed training. The strategy orchestrates replication across devices, so training cannot proceed; usually the strategy_type flag was not converted into a strategy via distribution_utils.get_distribution_strategy.

Solutions

1. Build a strategy first: strategy = distribution_utils.get_distribution_strategy(FLAGS.strategy_type) and pass it to run_bert
2. Check that the strategy_type/ tpu flag is set so get_distribution_strategy does not return None

When it happens

Trigger: Thrown at official/legacy/bert/run_classifier.py:385 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/86f8c48a623a6c8a. Report an issue: GitHub.

Appendix: source

Thrown at official/legacy/bert/run_classifier.py:385

             custom_metrics=None):
  """Run BERT training."""
  # Enables XLA in Session Config. Should not be set for TPU.
  keras_utils.set_session_config(FLAGS.enable_xla)
  performance.set_mixed_precision_policy(common_flags.dtype())

  epochs = FLAGS.num_train_epochs * FLAGS.num_eval_per_epoch
  train_data_size = (
      input_meta_data['train_data_size'] // FLAGS.num_eval_per_epoch)
  if FLAGS.train_data_size:
    train_data_size = min(train_data_size, FLAGS.train_data_size)
    logging.info('Updated train_data_size: %s', train_data_size)
  steps_per_epoch = int(train_data_size / FLAGS.train_batch_size)
  warmup_steps = int(epochs * train_data_size * 0.1 / FLAGS.train_batch_size)
  eval_steps = int(
      math.ceil(input_meta_data['eval_data_size'] / FLAGS.eval_batch_size))

  if not strategy:
    raise ValueError('Distribution strategy has not been specified.')

  if not custom_callbacks:
    custom_callbacks = []

  if FLAGS.log_steps:
    custom_callbacks.append(
        keras_utils.TimeHistory(
            batch_size=FLAGS.train_batch_size,
            log_steps=FLAGS.log_steps,
            logdir=FLAGS.model_dir))

  trained_model, _ = run_bert_classifier(
      strategy,
      model_config,
      input_meta_data,
      FLAGS.model_dir,
      epochs,
      steps_per_epoch,

View on GitHub (pinned to e006f5f0d5)