|
pulsatrix 1.1.0
|
Attribution graphs with cross-layer transcoders (FEAT-10), after Ameisen, Lindsey et al., "Circuit Tracing: Revealing Computational Graphs in Language Models" (Transformer Circuits, 2025) and its open implementation, circuit-tracer (github.com/decoderesearch/circuit-tracer). Each MLP is replaced by a transcoder's features plus an error term, attention patterns and normalization scales are frozen, and what is left is linear: the graph's edges are exact attributions between features, errors, token embeddings and output logits. More...
#include <cstdint>#include <string>#include <vector>#include "pulsatrix/causal_lm.hpp"#include "pulsatrix/tensor.hpp"#include "pulsatrix/viz/attribution_graph.hpp"
Go to the source code of this file.
Classes | |
| class | pulsatrix::CrossLayerTranscoder |
| A cross-layer transcoder (CLT) over a model's L MLPs, or a set of per-layer transcoders (PLTs, a CLT whose features only write to their own layer). More... | |
| struct | pulsatrix::CircuitTraceOptions |
| struct | pulsatrix::CircuitTrace |
| An attribution graph before pruning: every active feature, error and token, the explained logits, and the full adjacency. More... | |
| struct | pulsatrix::CircuitTrace::Feature |
| struct | pulsatrix::CircuitScores |
| Circuit-tracer's scores of a graph. More... | |
Namespaces | |
| namespace | pulsatrix |
Functions | |
| CrossLayerTranscoder | pulsatrix::LoadTranscoders (const std::string &directory, int64_t num_layers, DeviceBackend *backend) |
| Loads transcoders from a directory in one of the layouts circuit-tracer reads, with bf16/fp16 weights upcast: | |
| CircuitTrace | pulsatrix::TraceCircuit (CausalLM &model, const CrossLayerTranscoder &transcoder, const std::vector< int64_t > &ids, const CircuitTraceOptions &options={}) |
Runs the model on ids (one sequence) with transcoder in place of its MLPs, and attributes the salient logits and every active feature through the frozen model. | |
| CircuitScores | pulsatrix::ScoreCircuit (const CircuitTrace &trace) |
| Circuit-tracer's replacement and completeness scores (graph.py). | |
| AttributionGraph | pulsatrix::ToAttributionGraph (const CircuitTrace &trace, const CircuitTraceOptions &options, const std::vector< std::string > &logit_labels={}) |
The graph for Neuronpedia's or circuit-tracer's viewer, pruned as circuit-tracer prunes: nodes by indirect influence on the logits (node_threshold), then edges by influence-weighted score (edge_threshold), then nodes left without edges. Node ids, types and labels follow circuit-tracer ("<layer>_<feature>_<position>", "cross layer transcoder", "mlp reconstruction error", "E_<token>_<position>", "<L+1>_<token>_<position>" for logits). | |
Attribution graphs with cross-layer transcoders (FEAT-10), after Ameisen, Lindsey et al., "Circuit Tracing: Revealing Computational Graphs in Language Models" (Transformer Circuits, 2025) and its open implementation, circuit-tracer (github.com/decoderesearch/circuit-tracer). Each MLP is replaced by a transcoder's features plus an error term, attention patterns and normalization scales are frozen, and what is left is linear: the graph's edges are exact attributions between features, errors, token embeddings and output logits.