DistributedFunSearch
DistributedFunSearch (disfun) is a multi node distributed implementation of FunSearch (Romera et al., 2024) that uses LLM guided evolutionary search to discover novel algorithms. It uses RabbitMQ for asynchronous message passing and supports local models (via vLLM) and API models (via LiteLLM) through a unified inference interface.
- Independent workers: ProgramsDatabase, Samplers, and Evaluators work independently and process tasks asynchronously to maximize throughput
- Multi node execution: Distributes across multiple nodes and allows adding Samplers or Evaluators from the same or different nodes to a running experiment (see Docker Setup, Cluster Setup)
- Dynamic scaling: Automatically spawns/terminates Samplers and Evaluators based on workload and available resources (see Scaling Guide)
In each iteration:
- A few-shot prompt is constructed by sampling high-scoring functions from the program database
- The LLM generates a new candidate function (optionally with metadata, e.g., a natural language summary)
- The function is evaluated in a sandboxed environment on user-defined problem instances
- If executable and producing unique outputs (not a duplicate), the new function is stored with its scores and metadata
Our implementation includes an example application for discovering large deletion correcting codes.
Quickstart
Installation
conda create -n env python=3.11 pip numpy==1.26.4 -y
conda activate env
# 2. Install PyTorch (skip if using API based LLM)
conda install pytorch==2.2.2 pytorch-cuda=12.1 -c pytorch -c nvidia -y
# 3. Install C compiler (required for vLLM/Triton)
# For local models only, skip if using API based LLM
conda install -c conda-forge gcc_linux-64 gxx_linux-64 -y
# 4. Start RabbitMQ
sudo systemctl start rabbitmq-server
# 5. Install DistributedFunSearch
pip install . # or pip install -e . for development mode
# 6. Build FastGraph C++ module (optional, for graph problems)
# This compiles for your active Python version (e.g., cpython-311 for Python 3.11).
# If nodes have different Python versions, build on each node separately.
# Requires: build-essential, liblmdb-dev
sudo apt-get install -y build-essential liblmdb-dev
./tools/build_fast_graph.sh
See Docker Setup for container based installation or Cluster Setup for cluster execution.
Run experiment
This runs the example specification for discovering deletion correcting codes using StarCoder2-15B (local model):
python -m disfun
# Resume from checkpoint (if needed)
python -m disfun --checkpoint path/to/checkpoint.pkl
Note: On first run, model weights (~30GB for StarCoder2-15B) are downloaded to ~/.cache/huggingface/. To change the model or cache location, edit model or cache_dir in src/experiments/experiment1/config.py.
You can monitor message load in real time at http://localhost:15672 (login: guest/guest). Enable the management plugin first with sudo rabbitmq-plugins enable rabbitmq_management.
Evolve your problem
Adapt DistributedFunSearch to your problem by creating a specification directory with evaluation scripts, prompt templates, and initial functions. See src/disfun/specifications/ECC/ and its README for a complete working example and documentation.