Zero Shot MBPP Reproduction Using Code Llama

This project aims to reproduce the Zero Shot MBPP (Massive Bank of Problems in Programming) results as described in the "Code Llama: Open Foundation Models for Code" paper. We utilize the powerful capabilities of the Code Llama model to explore innovative approaches in code generation and comprehension.

Prerequisites

Before running the code, ensure you have a Lorax Predibase server operational. Follow the steps below to set up the server environment suitable for your needs.

Setting Up Lorax Predibase Server

For Bit-Sand Quantization:

docker run --runtime=nvidia -e PORT="8080" -p 8080:8080 -v $volume:/data ghcr.io/predibase/lorax:latest --model-id codellama/CodeLlama-7b-Instruct-hf --max-input-length 512 --max-batch-prefill-tokens 1024 --quantize bitsandbytes-nf4

For No Quantization (13B Model):

docker run --runtime=nvidia -e PORT="8080" -p 8080:8080 -v $volume:/data ghcr.io/predibase/lorax:latest --model-id codellama/CodeLlama-13b-Instruct-hf --max-input-length 512 --max-batch-prefill-tokens 1024

Running the Tests

To validate the setup and ensure everything is working as expected, run the unit tests using the following command:

pytest

Usage

Running Pass1 Script To execute the Pass1 script, use the following command. This will also log the output to a file for further analysis.

python3 async_main_pass1.py | tee 7B_np4_pass1.log

Running Pass10 Script Similarly, to run the Pass10 script, use the following command:

python3 async_main_pass10.py | tee 7B_np4_pass10.log

Maintainer

This project is maintained by Michael Drob. For inquiries or contributions, please reach out through .