Edge Pruning For Circuit Discovery

This repository contains the code and data for the paper "Finding Transformer Circuits with Edge Pruning".

Quick Links

• Environment
• Overview of Edge Pruning
• Repository Structure
• Running the Code
  • GPT-2 Experiments
  • Tracr Experiments
  • CodeLlama Experiments
  • Other Models
• Bugs or Questions?
• Citation

Environment

After installing PyTorch, run

pip install -r requirements.txt

to install the required packages.

Overview of Edge Pruning

The method itself is described in depth in our paper. Here we outline the implementation.

We expand the hidden state tensor from (batch_size, seq_len, hidden_dim) to (node, batch_size, seq_len, hidden_dim). This array acts as a sort of outbox for every node. Then, a downstream node can read its input via muxing these outboxes with a binary mask. The masks and their optimization are implemented with the L0 trick (described further in our paper)---this part of the code is in src/modeling/l0.py. We also provide some additional implementation comments here: since the "target" sparsities specified to the pruning scripts are just numbers, we can (and sometimes do) specify values above 1 to strongly push the model to be as sparse as possible. Also note that at the end of the training, we discretize the masks to 0 or 1 from their floating point values by rounding w.r.t a threshold. Instead of fixing the threshold to 0.5, we perform a binary search during evaluating so that the final threshold causes a sparsity of exactly 1 - <average mask value>.

Repository Structure

Note: Please unzip the data.zip file to the data/ directory before running the code!

The repository is structured as follows.

• The data/ folder contains everything related to datasets and their generation. Specifically, data/datasets/ has the datasets for the various tasks used in our evaluation, while the scripts in data/scripts/ use the seed data in data/helper_files to generate these datasets programmatically. data/tracr_models contains the precompiled weights of the tracr models we use, while data/runs provides the final checkpoints of the pruned versions.
• src/ houses the code related to actually performing the pruning, and evaluating the pruned circuits. In particular, src/modeling defines the model classes we use. src/prune provides fine-tuning scripts and src/eval/ contains scripts that evaluate the produced circuits (or the whole model).
• run_scripts/ contain helper scripts that both demonstrate and help with launching pruning and evaluation runs.
• tracrx/ has a slightly modified Tracr implementation. The modifications allow us to return the embedding matrix entries when calling the model, so that we can save all weights for future use in our equivalent Tracr class in PyTorch.
• assets/ contains images used in this README.

Running the Code

GPT-2 Experiments

The datasets for IOI-t1 / IOI (Indirect Object Identification), GT (Greater Than) and GP (Gendered Pronoun) are under data/datasets/{ioi-t1/ioi/gt/gp}. If you wish to re-generate these, use the scripts in data/scripts/prepare_{ioi/gt/gp}.py. The modeling file for GPT-2 is src/modeling/modeling_fpt2.py, and the pruning scripts are found at src/prune/fpt2_{ioi/gt/gp}.py. The IOI pruning script can also be used for IOI-t1.

A demonstration of how to use these scripts, along with the hyperparameters we used, is provided in run_scripts/{ioi/gt/gp}_sweep.sh. There are further explanations in these files, and flags for enabling a node loss (along with an edge loss) or to disallow removing edges of the form embedding -> node. Please note that---in contrast to some prior methods---we disallow removing edges of the form head.Q/K/V -> head since it is equivalent to removing all incoming edges of the former node.

Evaluation scripts are found in src/eval/{ioi/gt/gp}.py. To evaluate a circuit found with the default settings above, you can simply run, for example,

python src/eval/ioi.py -m /path/to/pruned_model -w

The -w flag signifies that embedding nodes were allowed to be pruned in the circuit. You can also use this script with the original GPT-2 model (or any GPT-2 checkpoint).

An example circuit

Visualizing a GPT-2 circuit: The following two steps will let you generate a drawing of a circuit. First, save the circuit edges using src/modeling/vis_fpt2.py as follows:

python src/modeling/vis_fpt2.py -i /path/to/checkpoint/dir/ -w

This will save the edges (by default to /path/to/checkpoint/dir/edges.json), and -w tells the script that you model masks over embedding edges as well. Then, you can draw the circuit with

python src/modeling/draw_fpt2.py -i /path/to/checkpoint/dir/edges.json

The default output path is /path/to/checkpoint/dir/edges.pdf; please look at the scripts for other arguments. An example circuit is shown above.

Tracr Experiments

The Tracr data and model preparation involves two steps (or you can use the datasets and models provided under data/datasets/ and data/tracr_models/). We will illustrate them for the case of the task reverse (the files/steps for xproportion are the same, with the change reverse -> xproportion). You can call data/scripts/prepare_reverse.py to generate the dataset for the task. The model itself can be compiled by calling data/scripts/prepare_reverse_tracr-model.py. This will compile the tracr models using the tracrx/ code and then save the weights as pickle files (by default under data/tracr_models/).

These weights are now compatible with our PyTorch model, defined in src/modeling/modeling_erazr.py. The pruning code in src/prune/erazr_{reverse/xproportion}.py is called upon by the helper scripts in run_scripts/tracr_{reverse/xproportion}.sh. The evaluation is performed at the end of the run by the pruning code itself. Please refer to the helper scripts for an example of how to launch a pruning run.

Extension to other tracr models should be straightforward by modifying the pruning scripts above. However, note that Tracr seems to map the BOS token to a random index in its vocabulary for each task, and figuring out the specific index for may need an inspection of a few outputs.

CodeLlama Experiments

The data preparation scripts for Boolean Expressions are provided under data/datasets/, and the dataset itself in data/datasets/boolean_expressions/. The modeling code is in src/modeling/modeling_fllama.py and the pruning code is in src/prune/fllama_boolean_expressions_{fs/ip}.py for instruction prompting and few-shot (in-context learning) settings. To directly evaluate the obtained checkpoints, use src/eval/boolean_expressions.py just like the GPT-2 evaluation scripts. This script has the following differences from the GPT-2 evaluation scripts:

• Use the flag -m or --mode to specify the mode of evaluation (fewshot or instruction).
• The -e/--edges flag allows you to specify a JSON file with a list of edges that are loaded into the model first: this lets you evaluate, e.g., the intersection of two models found with the file src/modeling/vis_fllama.py (use the -m1, -m2 and -o flags to specify model paths and the output JSON path).
• You will probably need to launch this script with 4 (or more) GPUs: run_scripts/launch_fllama_eval.sh shows how to do this. For the pruning itself, the helper scripts in run_scripts/launch_fllama_{instr/fs}_prune.sh use PyTorch FSDP and call upon src/prune/fllama_boolean_expressions_{ip/fs}.py. Note that these are sbatch scripts (despite having the .sh extension). Please refer to them for the hyperparameters we used (and find the FSDP config files under run_scripts/fsdp_configs/). The runs are quite resource-intensive and required us to use multi-node training with 32 GPUs.

Note 1: We are aware of a bug in the CodeLlama pruning code due to which the loss starts out as a large negative number sometimes (when using multi-node training and initializing from a Llama checkpoint instead of our class). Terminating the run and re-launching it fixes the issue. Additionally, initializing from an Fllama class (load Llama into our class, then save to disk with save_pretrained and load that checkpoint instead) seems to help. We are looking into this bug and will fix it soon!

Note 2: To facilitate use with other models, we plan on releasing modeling files for a few popular architectures soon. Stay tuned!

Other Models

Extension to other models should be straightforward by mimicking modeling_{fpt2/fllama}.py in modiyfing the corresponding HuggingFace model files (e.g., modeling_gpt2.py).

Custom Dataset

Using a custom dataset with, e.g., the GPT-2 pruning script is as straightforward as writing into a JSONL file (data/datasets/example_custom.jsonl)

{
    "clean": "1 + 2 = <predict>3</predict>",
    "corrupted": "1 + 4 = <predict>5</predict>"
}
{
    "clean": "5 + 1 = <predict>6</predict>",
    "corrupted": "1 + 1 = <predict>2</predict>"
}
{
    "clean": "2 + 2 = <predict>4</predict>",
    "corrupted": "2 + 4 = <predict>6</predict>"
}

and calling the src/prune/fpt2_custom.py as in run_scripts/custom_example.sh. The clean and corrupted text must be filled into the entries clean and corrupted. The part meant for the model to predict must be enclosed in <predict></predict>. You can also add a split key in your examples with value train/validation. If this key is missing, the entire dataset is used for both training and validation.

Bugs or Questions?

Please reach out to Adithya <[email protected]> with any questions or bug reports.

Citation

If you use our work or found it useful, please consider citing us:

@misc{bhaskar2024finding,
    title={{F}inding Transformer Circuits with {E}dge {P}runing}, 
    author={Adithya Bhaskar and Alexander Wettig and Dan Friedman and Danqi Chen},
    year={2024},
    eprint={2406.16778},
    archivePrefix={arXiv},
}