Skip to content
Murano

Steps — Patch

Patch step: cross-run activation patching (interchange intervention).

Activation patching runs one input while injecting another input’s activations at chosen sites, then reads the logits, isolating how much those sites carry the behaviour. This is the primitive under circuit discovery and interchange analysis.

The default direction is denoising: run the corrupt prompts and patch in the clean activations at the target sites, so the metric recovers toward the clean run. Compose with three metric runs and :class:~murano.steps.metrics.RecoveredMetricStep to read the recovered fraction, and sweep targets (per layer, per head, or per positions) to localize the behaviour. Swapping base_key and source_key flips the direction to noising (run clean, patch in corrupt).

Patch is a thin preset over :class:~murano.steps.ablate.Ablate’s resample engine: it captures the source batch’s activations at the targets and blends them into the base run, reusing the same per-head dispatch and positions selection. The base and source prompts must be token-length-matched per pair so positions align (the step raises if they are not), which a :class:~murano.dataset.CleanCorruptDataset of equal-length pairs satisfies.

Patch

Section titled “Patch”

Patch source-run activations into a base run and write the logits.

Runs the base prompts (base_key) through one forward pass while replacing each target component’s activation with the source prompts’ (source_key) activation at the same site, then stores the resulting logits. With the defaults (base = corrupt, source = clean) this is denoising activation patching; a metric step scores how far the patched run recovers toward clean.

Reads from results:

  • results[base_key]: PromptBatch (default corrupt_prompts), the run to
  • patch into.
  • results[source_key]: PromptBatch (default prompts), the run supplying
  • the replacement activations.

Writes to results:

  • results[logits_key]: Tensor [B, S, vocab] of patched logits.
  • results[mask_key]: Tensor [B, S] marking real (1) vs padding (0) for the
  • base batch, so a downstream metric step can locate the answer position.

Args:

  • model: Model backend to run.
  • targets: Sites to patch: a :class:~murano.nodes.NodeSet, a single
  • address, or an iterable of addresses. A whole-component target patches
  • the module output; a head target (Node(layer, "self_attn", head=h))
  • patches that head. One call is a single mode (all whole-component or
  • all per-head). Pass this or targets_key, not both.
  • targets_key: Results key holding a
  • “: class:~murano.artifacts.ComponentSelection a discovery step wrote,
  • read at run time so attribute-then-patch composes in one pipeline. Pass
  • this or targets, not both.
  • base_key: Results key of the batch to patch into (default
  • corrupt_prompts).
  • source_key: Results key of the batch supplying replacement activations
  • (default prompts, the clean side).
  • positions: Token position(s) to patch, in the
  • “: func:~murano.steps.metrics._answer_positions form (an int or
  • per-example sequence, negatives allowed); None patches every
  • position.
  • logits_key: Results key to write the patched logits under.
  • mask_key: Results key to write the base attention mask under.

Raises:

  • ValueError: If neither or both of targets and targets_key are
  • given, targets is empty or mixes modes, or the base and source
  • prompts are not token-length-matched per pair.

__init__

Section titled “__init__”
def __init__(self, model: ModelBackend, targets: NodeSet | AddressLike | Iterable[AddressLike] | None = None, targets_key: str | None = None, base_key: str = keys.CORRUPT_PROMPTS, source_key: str = keys.PROMPTS, positions: int | Sequence[int] | torch.Tensor | None = None, logits_key: str = keys.PATCHED_LOGITS, mask_key: str = keys.PATCHED_MASK):