tensorflow/models · error · NotImplementedError

Router is an abstract class that should be subclassed.

Error message

Router is an abstract class that should be subclassed.

What it means

Error "Router is an abstract class that should be subclassed." thrown in tensorflow/models.

Source

Thrown at official/nlp/modeling/layers/moe.py:204

        each token and expert. Used for routing tokens to experts.
      - <float32>[num_groups, tokens_per_group, num_experts] raw router logits.
        Used for computing router z-loss.
    """
    if apply_jitter and self.jitter_noise > 0:
      inputs *= tf.random.uniform(
          tf.shape(inputs),
          minval=1.0 - self.jitter_noise,
          maxval=1.0 + self.jitter_noise,
          dtype=inputs.dtype)
    # inputs <float>, router_logits <float32>
    router_logits = self.router_weights(inputs)
    router_probs = tf_keras.activations.softmax(router_logits, axis=-1)
    return router_probs, router_logits

  def _compute_routing_instructions(self, router_probs: tf.Tensor,
                                    expert_capacity: int) -> RouterOutput:
    """Computes instructions for routing inputs to experts."""
    raise NotImplementedError(
        "Router is an abstract class that should be subclassed.")


class MaskedRouter(Router):
  """Abstract base router class for masked matmul dispatch routers.

  MaskedRouter(s) return RouterMask(s) containing a dispatch mask and combine
  array for sending and receiving (via masked matmuls) inputs and outputs to and
  from experts.

  Routing using masked matmuls is generally faster than scatter-based routing on
  TPUs.

  Uses Keras add_loss() and add_metric() APIs.
  """

  def _compute_routing_instructions(self, router_probs: tf.Tensor,
                                    expert_capacity: int) -> RouterMask:

View on GitHub (pinned to e006f5f0d5)

When it happens

Trigger: Thrown at official/nlp/modeling/layers/moe.py:204 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/876c52634c200862. Report an issue: GitHub.