tensorflow/models · error

Resetting the generator. %s: %s

Error message

Resetting the generator. %s: %s

What it means

Error "Resetting the generator. %s: %s" thrown in tensorflow/models.

Source

Thrown at official/projects/bigbird/recompute_grad.py:162

  ]


def _make_seed_if_none(seed: Optional[tf.Tensor]) -> tf.Tensor:
  """Uses the global generator to make a seed if necessary."""
  if seed is not None:
    return seed
  generator = tf.random.experimental.get_global_generator()
  # The two seeds for stateless random ops don't have individual semantics and
  # are scrambled together, so providing one seed is fine. This makes it easier
  # for users to provide a local seed without worrying about integer overflow.
  # See `make_seeds` in
  # https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/stateful_random_ops.py.
  try:
    return generator.uniform_full_int([], tf.int32, name='recompute_grad_seed')
  except (RuntimeError, TypeError, ValueError, tf.errors.NotFoundError) as e:
    # For a number of reasons, the above operation can fail like using multiple
    # graphs or toggling between eager and graph modes. Reset the generator.
    logging.warn('Resetting the generator. %s: %s', type(e), e)
    tf.random.experimental.set_global_generator(None)
    generator = tf.random.experimental.get_global_generator()
    return generator.uniform_full_int([], tf.int32, name='recompute_grad_seed')


def recompute_grad(f, seed=None):
  """An eager-compatible version of recompute_grad.

  For f(*args, **kwargs), this supports gradients with respect to args, or to
  gradients with respect to any variables residing in the kwarg 'variables'.
  Note that for keras layer and model objects, this is handled automatically.

  Warning: If `f` was originally a tf.keras Model or Layer object, `g` will not
  be able to access the member variables of that object, because `g` returns
  through the wrapper function `inner`.  When recomputing gradients through
  objects that inherit from keras, we suggest keeping a reference to the
  underlying object around for the purpose of accessing these variables.

View on GitHub (pinned to e006f5f0d5)

Solutions

1. Re-create the checkpoint reader/generator after this reset; the previous generator state is discarded.
2. Handle the logged exception (see the logged error) that triggered the reset.

When it happens

Trigger: Thrown at official/projects/bigbird/recompute_grad.py:162 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/73308c68ac84b288. Report an issue: GitHub.