transformer_lens.model_bridge.generalized_components.conv1d module

Conv1D bridge component for wrapping Conv1D layers with hook points.

class transformer_lens.model_bridge.generalized_components.conv1d.Conv1DBridge(name: str | None, config: Any | None = None, submodules: Dict[str, GeneralizedComponent] | None = None, conversion_rule: BaseTensorConversion | None = None, hook_alias_overrides: Dict[str, str] | None = None, optional: bool = False)

Bases: GeneralizedComponent

Bridge component for Conv1D layers.

This component wraps a Conv1D layer (transformers.pytorch_utils.Conv1D) and provides hook points for intercepting the input and output activations.

Conv1D is used in GPT-2 style models and has shape [in_features, out_features] (transpose of nn.Linear which is [out_features, in_features]).

forward(input: Tensor, *args: Any, **kwargs: Any) → Tensor

Forward pass through the Conv1D layer with hooks.

Parameters:

  • input – Input tensor

  • *args – Additional positional arguments

  • **kwargs – Additional keyword arguments

Returns:

Output tensor after Conv1D transformation