Add CircuitLens feature-circuit attribution (#131)

* Add CircuitLens feature-circuit attribution

* Address CircuitLens review feedback

Author: Xmaster6y

pyproject.toml

@@ -13,6 +13,7 @@ dependencies = [
     "tensordict>=0.10",
     "torch>=2.7.1",
 ]
+
 classifiers = [
   "Development Status :: 3 - Alpha",
   "Intended Audience :: Science/Research",
@@ -29,6 +30,12 @@ source = "https://github.com/Xmaster6y/tdhook"
 issues = "https://github.com/Xmaster6y/tdhook/issues"
 releasenotes = "https://github.com/Xmaster6y/tdhook/releases"
 
+[project.optional-dependencies]
+circuit-lens = [
+    "circuit-tracer>=0.5.0",
+    "transformer-lens>=2.16.0",
+]
+
 [dependency-groups]
 dev = [
     "captum>=0.8.0",

src/tdhook/attribution/__init__.py

@@ -2,20 +2,47 @@
 Module for attribution methods.
 """
 
+# Import order is intentional: legacy modules import Saliency from this package.
+# ruff: noqa: I001
+
 from .lrp import LRP
 from .saliency import Saliency
 from .grad_cam import GradCAM
 from .guided_backpropagation import GuidedBackpropagation
 from .activation_maximisation import ActivationMaximisation
 from .integrated_gradients import IntegratedGradients
+from .circuit_lens import (
+    AttentionContributor,
+    AttentionSite,
+    AttributionConventions,
+    CircuitLensArtifact,
+    FeatureContributor,
+    FeatureSite,
+    LogitContributor,
+    attention_contributions,
+    attribute_feature_circuit,
+    feature_contributions,
+    logit_contributions,
+)
 
 __all__ = [
+    "LRP",
     "ActivationMaximisation",
+    "AttentionContributor",
+    "AttentionSite",
+    "AttributionConventions",
+    "CircuitLensArtifact",
+    "FeatureContributor",
+    "FeatureSite",
     "GradCAM",
     "GuidedBackpropagation",
     "IntegratedGradients",
+    "LogitContributor",
     "Saliency",
-    "LRP",
+    "attention_contributions",
+    "attribute_feature_circuit",
+    "feature_contributions",
+    "logit_contributions",
 ]
 
 # TODO: Implement Occlusion