Using TorchLens with other tools¶
The bridges that ship in torchlens.bridge, sorted by the direction the data moves. A bridge is
imported on its own (from torchlens.bridge import captum) and needs the other tool installed;
nothing here is imported by import torchlens. Checked 2026-11-03.
Into TorchLens: another tool's artifact runs inside a trace¶
| From | Bridge | What comes in | Page |
|---|---|---|---|
| pyreft | torchlens.bridge.pyreft.read |
A saved LoReFT intervention, applied as a TorchLens subspace edit | Coming from pyvene |
| SAELens | torchlens.bridge.sae_lens.encode, reconstruct |
A published sparse autoencoder, run on a captured site | pretrained SAEs |
| any SAE library | torchlens.bridge.sae.splice |
An SAE's reconstruction spliced back into the forward, with what the splice did | pretrained SAEs |
| Hugging Face Hub | torchlens.bridge.huggingface_hub.from_hub |
A .tlspec bundle someone published |
glossary |
| Hugging Face transformers | torchlens.bridge.transformers.trace_text, trace_image, trace_multimodal |
A model and its processor, traced on raw text or images | quickstart |
Out of TorchLens: a trace or its features feed another tool¶
| To | Bridge | What goes out | Page |
|---|---|---|---|
| rsatoolbox | torchlens.bridge.rsatoolbox.dataset |
Per-site features as an rsatoolbox Dataset, rows tied to stimulus ids |
extract features |
| xarray and Brain-Score assemblies | torchlens.bridge.xarray.data_array, neuroid_assembly |
Per-site presentation-by-neuroid arrays | NeuroAI journey |
| Brain-Score | torchlens.bridge.brain_score.activations_extractor, per_layer |
An activations extractor over TorchLens sites | brain-benchmarking seam |
| Captum | torchlens.bridge.captum.attribute, layer |
A TorchLens site as a Captum layer target | Coming from Captum |
| inseq | torchlens.bridge.inseq.attribute |
Text attribution on a TorchLens-addressed site | Coming from inseq |
| SHAP | torchlens.bridge.shap.explain |
A site as the explained output | Coming from Captum |
| pytorch-grad-cam | torchlens.bridge.pytorch_grad_cam.cam, layer |
A TorchLens site as the CAM target layer | Grad-CAM with torch-cam |
| steering-vectors, repeng, dialz | torchlens.bridge.steering_vectors.from_direction, torchlens.bridge.repeng.from_direction, torchlens.bridge.dialz.from_direction |
A TorchLens Direction as that library's steering object |
Coming from RepE |
| Hugging Face Hub | torchlens.bridge.huggingface_hub.push_to_hub |
A .tlspec bundle, published |
glossary |
| treescope | torchlens.bridge.treescope.register |
TorchLens objects rendered in treescope | treescope cards |
| torch.profiler | torchlens.bridge.profiler.join |
Profiler events joined to TorchLens operations | profilers alongside |
| depyf | torchlens.bridge.depyf.dump |
Decompiled source next to the trace | glossary |
| agents over MCP | torchlens.bridge.mcp.main |
The agent registry served over stdio | TorchLens for AI coding agents |
Both directions¶
| With | How |
|---|---|
| Hugging Face attention | torchlens.bridge.hf_attention.register_attention registers TorchLens as an attention implementation, so attention internals are recorded ops in a model that otherwise calls a fused kernel |
| Plain tensors and files | Every saved activation is a torch.Tensor; trace.save(path) and tl.load(path) move a whole recording, and any tool that reads tensors reads the rest |