Which tool for this job?¶

What TorchLens and seven other tools do, job by job, including the jobs another tool does and TorchLens does not. Checked 2026-11-03 against the versions named in the table header; tool status and sources are on the landscape page.

When you need TorchLens, and when five lines of hooks are enough¶

Five lines of register_forward_hook are enough when the value you want is the output of one module, you know which module, and the model is yours to read. Hooks ship with torch, add no dependency, and cost almost nothing per call.

You need a capture when the value is not a module output (a functional call, a residual add, an in-place op inside forward), when you do not know in advance which site matters, when you want to check that what you recorded is what the forward computed, or when you will edit a site and compare runs. The fence below shows the first case: the hook sees the module outputs, and the ReLU that forward calls as a function is not one of them.

import torch
from torch import nn

import torchlens as tl


class Block(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.linear = nn.Linear(4, 4)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return torch.relu(self.linear(x)) + x


torch.manual_seed(0)
model = Block().eval()
x = torch.randn(2, 4)

hooked = []
handles = [m.register_forward_hook(lambda m, a, out: hooked.append(out)) for m in model.modules()]
model(x)
for handle in handles:
    handle.remove()

trace = tl.trace(model, x)
funcs = [op.func_name for op in trace.ops]
assert len(hooked) == 2  # the Linear and the Block itself
assert "relu" in funcs  # the functional ReLU is a recorded operation

The table¶

Cells read: yes, partial (with what is missing), no, or by hand (possible with code you write, not provided). Rows where another tool covers a job TorchLens does not are marked in the last column.

Job	TorchLens	Plain hooks (torch)	TransformerLens 4.3	nnsight 0.8	pyvene 0.1.8	Captum 0.9	thingsvision 1.4	torchvision feature extraction	Another tool covers more
Read operations no module returns (functional calls, residual adds)	yes	no: module boundaries only	partial: the hook points its model classes define	yes: `.source`	no: named components	no: layer hooks	no: module outputs	partial: fx graph nodes, symbolically traceable models only
Any PyTorch architecture, no per-family adapter	yes	yes	partial: transformers, through `TransformerBridge`	yes	partial: component names per model type	yes	partial: a model registry plus custom models	partial: symbolically traceable models
Data-dependent control flow in `forward`	yes: the executed path	yes	yes	yes	yes	yes	yes	no: symbolic trace
Check the recording against the real forward by replay	yes	no	no	no	no	no	no	no
Draw the executed graph	yes	no	no	no	no	no	no	partial: prints the fx graph
Edit an activation and rerun (ablate, patch, steer)	yes	by hand	yes	yes	yes	partial: ablation inside attribution methods	no	no
Edit applied live at forward-hook cost, including `generate()`	yes: `tl.when(...).bind(model)`	yes	yes	yes	yes	no	no	no
Batch feature extraction over a stimulus set, with provenance	yes: `tl.extract_dataset`	by hand	partial: `run_with_cache` in your loop	by hand	no	no	yes	partial: your loop
Save a recording, load it, diff two runs	yes	by hand	partial: the cache as tensors	partial: saved values	partial: saved interventions	no	partial: saves features	no
Gradients at internal sites	yes	yes: tensor hooks	yes	yes	partial	yes: layer gradients	no	partial: autograd on outputs
Trainable interventions (DAS, LoReFT)	no: reads pyreft artifacts, ships no trainer	by hand	no	by hand	yes	no	no	no	pyvene, pyreft
Sparse autoencoder training	no: use a trainer	no	no: SAELens builds on it	no	no	no	no	no	SAELens
Attribution method breadth (TCAV, influence functions, attacks, text DeepLift and LIME)	partial: a gradient and perturbation set in `tl.attribution`	no	no	no	no	yes	no	no	Captum, inseq
Serving-scale and remote capture (vLLM, remote GPUs, tensor parallel)	no: local eager forward	no	no	yes: vLLM back end, NDIF	no	no	no	no	nnsight
Model-specific LLM conveniences (canonical hook names across families, pretrained loader, folded LayerNorm)	partial: resolves TransformerLens names on Hugging Face models	no	yes	partial: `LanguageModel`	no	no	no	no	TransformerLens
One module's output, nothing to install	works, with a dependency and capture cost	yes	no	no	no	no	no	yes, with torchvision	plain hooks
Non-PyTorch models	preview: JAX, TensorFlow, MLX, Paddle, tinygrad	no	no	no	no	no	partial: TensorFlow and Keras models	no

What the table does not say¶

Reading module-less operations, diffusers pipelines, quantized models and vision-language models is shared ground with nnsight 0.8; the first row is not a claim TorchLens alone has.
"No" means the tool does not provide it on the check date, not that it cannot be built on top.
A tool's own documentation is the authority for its column; corrections are welcome as issues.

Where a row points to another tool, where to go instead says how to get there and how to bring the result back to TorchLens.