Skip to content

Where to go instead

Jobs TorchLens does not do, the tool that does each one, and how a result comes back to TorchLens when you want to check it on an eager model. Tool status is as of 2026-11-03; see the landscape page for versions and sources.

Capture at serving scale, on remote GPUs, or across devices

TorchLens records a local eager forward. Serving engines run fused kernels, paged KV caches and CUDA graphs outside eager PyTorch, so there is nothing for it to record there, and tl.report.doctor says so in its serving_engine row.

  • Capture or steering inside vLLM: nnsight 0.8's vLLM back end, Neuronpedia's interp-engine, vllm-lens (UK AI Security Institute), EasySteer for steering.
  • Remote GPUs: nnsight with NDIF.
  • Tensor-parallel and multi-GPU models: nnsight's tensor-parallel back end. For data-parallel feature extraction with TorchLens, see multi-GPU extraction.

Coming back: load the same checkpoint as a Hugging Face model, trace it, and check the site or steering vector on the eager twin (deployment envelope). To check an engine's hooks site by site against TorchLens on a small model, by TransformerLens or nnterp site name, see cross-engine checks.

Train a sparse autoencoder or a probe

  • Sparse autoencoders: SAELens and the trainers on the SAE trainers recipe. TorchLens writes the activations they train on (tl.extract_dataset) and splices a trained SAE back in (torchlens.bridge.sae.splice, pretrained SAEs).
  • Probes: the training loop is yours (scikit-learn or plain PyTorch); tl.extract_dataset writes the features with their provenance (non-goals).

Train an intervention (DAS, LoReFT)

pyvene and pyreft train interventions. TorchLens ships no trainer; it reads a saved pyreft LoReFT artifact with torchlens.bridge.pyreft.read and applies it as a subspace edit (Coming from pyvene).

Attribution beyond the TorchLens set

tl.attribution covers gradients, integrated gradients, occlusion, Grad-CAM, SHAP-style and perturbation curves.

  • Influence functions and adversarial attacks: use Captum.
  • TCAV: Captum, or the TCAV recipe, which runs it on TorchLens sites with a causal check.
  • DeepLift and LIME for text: Captum or inseq. TorchLens provides no DeepLift or LIME for text.

Coming back: torchlens.bridge.captum and torchlens.bridge.inseq point those tools at a TorchLens site (Coming from Captum, Coming from inseq).

Language-model conveniences across model families

TransformerLens 4 gives canonical hook names across model families, its pretrained loader and LayerNorm folding. TorchLens resolves TransformerLens hook names on a stock Hugging Face model and refuses to trace a TransformerBridge (Coming from TransformerLens).

Read one module's output

A forward hook is enough: five lines, no dependency. See when five lines of hooks are enough.

Compiled, scripted or exported models

TorchLens does not trace torch.compile, TorchScript or torch.export artifacts. Trace the eager module they came from.

Representational similarity and brain scores

rsatoolbox and Brain-Score are where those analyses run. TorchLens hands them features with provenance (extract features, the brain-benchmarking seam).