Installation

Requirements

• Python 3.10 or higher
• PyTorch 2.7+
• A CUDA-capable GPU is recommended for running models

Install

pip install murano-interp

The base install is deliberately lean: it carries only what every workflow needs (recording, steering, intervention, the causal substrate). Feature-specific libraries ship as extras, so you install for what you actually do:

Extra	Use case	Pulls in
(base)	recording, steering, intervention, logits, ablation, metrics, paired datasets	nnsight, nnterp, torch, transformers
probe	linear probing	scikit-learn
data	loading datasets by name from the Hub	datasets
plot	figures and visualizations	matplotlib, seaborn, plotly
sae	sparse autoencoder features	sae-lens
all	everything above	all of the above

Combine extras as needed:

pip install "murano-interp[probe,plot]"   # probing with figures
pip install "murano-interp[all]"          # everything

If you call a feature whose extra is not installed, Murano raises a clear error naming the extra to install.

The PyPI distribution is murano-interp (the bare name murano is held by an unrelated OpenStack project). The Python module name is unchanged: import murano.

Development install

If you want to contribute or run from source:

git clone https://github.com/UKPLab/murano
cd murano
uv sync --all-extras --dev

Or with pip:

pip install -e ".[plot]"

HuggingFace setup

Murano loads models from HuggingFace Hub. On first use, models are downloaded to the local cache (~/.cache/huggingface/hub). Subsequent runs load from cache without network access.

If your models require authentication (e.g. gated Llama weights), log in first:

huggingface-cli login

Verify the install

import murano
model = murano.Model("meta-llama/Llama-3.2-1B-Instruct")
print(model)  # MuranoModel('meta-llama/Llama-3.2-1B-Instruct', layers=16, d=2048)