FunSearch Implementation

This repository implements the following publication:

Romera-Paredes, B. et al. Mathematical discoveries from program search with large language models. Nature (2023)

Installation and Requirements

Please note that the Python version must be larger or equal to Python 3.9, or the 'ast' package used in the implementations will fail to work.

You can run FunSearch for online bin packing locally if enough GPU devices are available. Or you can try to use LLM interfaces to request responses online.

Please install the packages listed in requirements.txt.

Project Structure

There are some independent directories in this project:

• bin_packing contains an example jupyter notebook for the bin packing task. See here.
• implementation contains an implementation of the evolutionary algorithm, code manipulation routines, and a single-threaded implementation of the FunSearch pipeline.
• llm-server contains the implementations of an LLM server that gets the prompt by monitoring requests from FunSearch and response to the inference results to the FunSearch algorithm.

Files in funsearch/implementation

There are some files in funsearch/implementation. They are as follows:

• code_manipulatoin.py provides functions to modify the code in the specification.
• config.py includes configs of funsearch.
• evaluator.py trims the sample results from LLM, and evaluates the sampled functions.
• evaluator_accelerate.py accelerates the evaluation using the 'numba' library.
• funsearch.py implements funsearch pipeline.
• profile.py records the score of the sampled functions.
• programs_database.py evolves the sampled functions.
• sampler.py sends prompts to LLM and gets results.

Run FunSearch Demo on Colab

The jupyter notebook in bin_packing/bin_packing_funsearch.ipynb can be opened via . Please note that do not run jupyter notebook locally, as the jupyter notebook backend does not support multiprocess running.

Run FunSearch Demo Locally

Parameters and Settings

If you want to adjust the following parameters, you should modify the code in funsearch/implementation manually.

• timeout_seconds This parameter defines the maximum evaluation time for a single function. If the evaluation time exceeds this, the evaluation process will be killed. This strategy can prevent while True loop and reduce total evaluation costs but may discard potential outstanding functions. You can modify this in implementation/evaluator.py/class Evaluator.
• _reduce_score This function does reduction to the score of a sampled function in some instances. The reduction is implemented as mean by default. You can modify it in implementation/program_database.py, where you can find a '_reduce_score' function.

Use Local LLM

1. First, start the local LLM server.

# Suppose we are in funsearch directory (root dir of this project).
cd llm-server
# Start LLM server: python llm_server.py --port 8088 --path [model path] --d [GPU IDs]
python llm_server.py --port 8088 --path /LLms/CodeLlama-34b --d 0 1 2 3 4 5

2. Then, start FunSearch.

# Run FunSearch
python funsearch_bin_packing_local_llm.py

You can see logs via Tensorboard. Please check the log_dir variable defined in bin_packing_funsearch_my_template.py, and start the Tensorboard using the following instructions:

# Suppose we are in funsearch directory (root directory of this project)
cd logs
tensorboard --logdir funsearch_local_llm

Use LLM Interfaces

1. Set the API's IP address according to your API provider. The code is in funsearch_bin_packing_llm_api.py line 33.

conn = http.client.HTTPSConnection("api.chatanywhere.com.cn")

2. Set the API key in request headers, the code lies in funsearch_bin_packing_llm_api.py line 44-48. You should replace sk-ys... with your API key.

headers = {
  'Authorization': 'Bearer sk-ys02zx...(replace with your API key)...',
  'User-Agent': 'Apifox/1.0.0 (https://apifox.com)',
  'Content-Type': 'application/json'
}

3. Start FunSearch.

# Run FunSearch
python funsearch_bin_packing_llm_api.py

You can see logs via Tensorboard. Please check the log_dir variable defined in bin_packing_funsearch_my_template.py, and start the Tensorboard using the following instructions:

# Suppose we are in funsearch directory (root directory of this project).
cd logs
tensorboard --logdir funsearch_llm_api

Issue

If you encounter any difficulty using the code, please do not hesitate to submit an issue!