Registry / ai-ml / transformer-lens

transformer-lens

library3.4.0pypypi✓ verified 81d ago

A library for training and analysing transformer models, focused on mechanistic interpretability. Provides tools to probe, edit, and visualise model internals. Current version 3.1.0, released with support for Python >=3.10, <4.0. Active development.

ai-ml llm-agents

pip install transformer-lens

package ↗homepage ↗

INSTALL

IMPORT

SIG · TRANSFORMER-LENS

transformer-lens

ai-mlpythonv3.4.0

Install

—

Import

—

Disk

—

Pass rate

0/ 10

Env Coverage0 / 10

glibc

3.9–3.13

musl

3.9–3.13

Install & Compatibility

Where this runs

tested against v? · pip install

Install × environment matrix

Each cell = how many times install + import succeeded across repeated harness runs. Partial = flaky.
glibc = Debian/Ubuntu slim · musl = Alpine Linux

musl

py 3.10–3.95 runs

build_error

glibc

py 3.10–3.95 runs

timeout

Code

Verified usage

Verified import paths — ran on the pinned version, not inferred.

HookedTransformer

✓ from transformer_lens import HookedTransformer

✗ from transformer_lens.components import HookPoint

HookedTransformerConfig

✓ from transformer_lens import HookedTransformerConfig

ActivationCache

✓ from transformer_lens import ActivationCache

Load a pretrained model and run a forward pass.

import torch from transformer_lens import HookedTransformer model = HookedTransformer.from_pretrained("tiny-stories-1L-21M") prompts = "The capital of France is" logits = model(prompts) token = logits.argmax(dim=-1).squeeze() print(model.tokenizer.decode(token))

Debug

Known issues

breakingIn version 3.0, the 'transformer_lens' package was renamed from 'transformer-lens' (hyphen) to 'transformer_lens' (underscore) for imports. Old imports using 'transformer_lens' (with hyphen) will fail.

fix

Change all imports to use underscores: 'import transformer_lens'.

affects: >=3.0.0

deprecatedThe old 'utils' module (transformer_lens.utils) has been deprecated in favour of individual submodules. Functions like 'to_numpy' have moved.

fix

Check the changelog for the new location of utility functions.

affects: >=3.0.0

gotchaWhen using 'model.run_with_cache', the returned cache is a dictionary keyed by layer names, but the tensor dimensions can be counterintuitive (batch, pos, d_model). Ensure you permute correctly for visualisation.

fix

Cache tensors are of shape (batch, pos, d_model). Use cache['blocks.0.hook_mlp'].shape to verify.

affects: all

Upgrade

Version history

3.4.0latest on PyPI · released Jun 9, 2026

Audit

Dependencies

No dependency data recorded yet.

Agent activity

40 hits · last 30 days

node

OpenAI (training)

Resources

packagehttps://pypi.org/project/transformer-lens/ ↗

homepagehttps://transformerlens.org ↗