tensorflow/models · error · ValueError

Variable name is not supported on CustomLayerQuantize( )…

Error message

Variable name {} is not supported on CustomLayerQuantize({}) transform.

What it means

Guard in the nlp CustomLayerQuantize transform's _is_quantization_weight_name: during pattern replacement every TransformerEncoderBlock weight must be either a QAT weight or a whitelisted model weight; an unrecognized variable name cannot be mapped to the quantized layer, so this ValueError names the variable and the transform.

Solutions

  1. Add the new weight name to the supported weight-name sets in the scheme
  2. Check for a renamed variable in the updated TransformerEncoderBlock
  3. Update the quantized layer class to consume the unknown weight
Defensive patterns

Strategy: fallback

When it happens

Trigger: Thrown at official/projects/qat/nlp/quantization/schemes.py:82 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/0b5c7b41afdc8e8a. Report an issue: GitHub.

Appendix: source

Thrown at official/projects/qat/nlp/quantization/schemes.py:82

      'embeddings'
  })

  def __init__(self):
    super().__init__()
    self._original_layer_pattern = 'modeling>TransformerEncoderBlock'
    self._quantized_layer_class = transformer_encoder_block.TransformerEncoderBlockQuantized

  def pattern(self) -> LayerPattern:
    """See base class."""
    return LayerPattern(self._original_layer_pattern)

  def _is_quantization_weight_name(self, name):
    simple_name = name.split('/')[-1].split(':')[0]
    if simple_name in self._QUANTIZATION_AWARE_TRAINING_WEIGHT_NAMES:
      return True
    if simple_name in self._SUPPOTRED_MODEL_WEIGHT_NAMES:
      return False
    raise ValueError('Variable name {} is not supported on '
                     'CustomLayerQuantize({}) transform.'.format(
                         simple_name,
                         self._original_layer_pattern))

  def replacement(self, match_layer: LayerNode) -> LayerNode:
    """See base class."""
    bottleneck_layer = match_layer.layer
    bottleneck_config = bottleneck_layer['config']
    bottleneck_names_and_weights = list(match_layer.names_and_weights)
    quantized_layer = self._quantized_layer_class(
        **bottleneck_config)

    quantized_layer_config = quantized_layer.get_config()
    if 'hidden_size' in quantized_layer_config:
      dummy_input_shape = [
          1, 1, quantized_layer_config['hidden_size']]
      quantized_layer.compute_output_shape(dummy_input_shape)
    elif 'num_attention_heads' in quantized_layer_config:

View on GitHub (pinned to e006f5f0d5)