Skip to content

Latest commit

 

History

History
198 lines (154 loc) · 7.42 KB

README.md

File metadata and controls

198 lines (154 loc) · 7.42 KB

RepoQA: Evaluating Long-Context Code Understanding

🏠 Homepage: https://evalplus.github.io/repoqa.html

🚀 Installation

# without vLLM (can run openai, anthropic, and huggingface backends)
pip install --upgrade repoqa
# To enable vLLM
pip install --upgrade "repoqa[vllm]"
⏬ Install nightly version :: click to expand ::
pip install --upgrade "git+https://github.com/evalplus/repoqa.git"                 # without vLLM
pip install --upgrade "repoqa[vllm] @ git+https://github.com/evalplus/repoqa@main" # with vLLM
⏬ Using RepoQA as a local repo? :: click to expand ::
git clone https://github.com/evalplus/repoqa.git
cd repoqa
export PYTHONPATH=$PYTHONPATH:$(pwd)
pip install -r requirements.txt

🏁 Search Needle Function (SNF)

Search Needle Function is the first and base RepoQA task which aims to practice LLMs' ability of long-context code understanding and retrieval. Its corresponding real-life scenario is to perform precise code search from function description.

🔎 More dataset details :: click to expand ::

[!Note]

SNF includes 500 tests (5 programming languages x 10 repos x 10 needle functions) where an LLM is given:

  1. A large code context sorted in file dependency
  2. A NL description of the needle function without revealing keywords like function names
  3. An instruction to retrieve the described function

The evaluator passes a test if the searched function is syntactically closest to the ground-truth compared against other functions (systematically parsed by treesitter) and the similarity is greater than a user defined threshold (by default 0.8).

You can run the SNF evaluation using various backends:

OpenAI Compatible Servers

repoqa.search_needle_function --model "gpt4-turbo" --backend openai
# 💡 If you use openai API compatible server such as vLLM servers:
# repoqa.search_needle_function --base-url "http://localhost:[PORT]/v1" \
#                               --model "Qwen/CodeQwen1.5-7B-Chat" --backend openai

Anthropic Compatible Servers

repoqa.search_needle_function --model "claude-3-haiku-20240307" --backend anthropic

vLLM

repoqa.search_needle_function --model "Qwen/CodeQwen1.5-7B-Chat" --backend vllm
🔎 Context extension for small-ctx models :: click to expand ::

There are two ways to unlock a model's context at inference time:

  1. Direct Extension: Edit max_positional_embedding of the model's config.json (e.g., hub/models--meta-llama--Meta-Llama-3-8B-Instruct/snapshots/[hash]/config.json) to something like 22528.
  2. Dynamic RoPE Scaling: To extend Meta-Llama-3-8B-Instruct from 8k to 32k (4x), edit the config.json:

"rope_scaling": {"type": "dynamic", "factor": 4.0}

Note: This works for vLLM <0.4.3 and HuggingFace transformers. RepoQA will automatically configure dynamic RoPE for vLLM >= 0.4.3

Note

Reference evaluation time:

  • Llama3-8B-Instruct: 45 minutes on 2xA6000 (PCIe NVLink)
  • Llama3-70B-Instruct: 100 minutes on 4xA100 (PCIe NVLink)

HuggingFace transformers

repoqa.search_needle_function --model "Qwen/CodeQwen1.5-7B-Chat" --backend hf --trust-remote-code

Tip

Installing flash-attn and additionally set --attn-implementation "flash_attention_2" can largely lower the memory requirement.

🔨 Having trouble installing `flash-attn`? :: click to expand ::

If you have trouble with pip install flash-attn --no-build-isolation, you can try to directly use pre-built wheels:

export FLASH_ATTN_VER=2.5.8 # check latest version at https://github.com/Dao-AILab/flash-attention/releases
export CUDA_VER="cu122"     # check available ones at https://github.com/Dao-AILab/flash-attention/releases
export TORCH_VER=$(python -c "import torch; print('.'.join(torch.__version__.split('.')[:2]))")
export PY_VER=$(python -c "import platform; print(''.join(platform.python_version().split('.')[:2]))")
export OS_ARCH=$(python -c "import platform; print(f'{platform.system().lower()}_{platform.machine()}')")

export WHEEL=flash_attn-${FLASH_ATTN_VER}+${CUDA_VER}torch${TORCH_VER}cxx11abiFALSE-cp${PY_VER}-cp${PY_VER}-${OS_ARCH}.whl
wget https://github.com/Dao-AILab/flash-attention/releases/download/v${FLASH_ATTN_VER}/${WHEEL}
pip install ${WHEEL}

Google Generative AI API (Gemini)

repoqa.search_needle_function --model "gemini-1.5-pro-latest" --backend google

CLI Usage

  • Input:
    • --model: Hugging-Face model ID, such as ise-uiuc/Magicoder-S-DS-6.7B
    • --backend: vllm (default) or openai
    • --base-url: OpenAI API base URL
    • --code-context-size (default: 16384): #tokens (by DeepSeekCoder tokenizer) of repository context
    • --caching (default: True): accelerate subsequent runs by caching preprocessing; --nocaching to disable
    • --max-new-tokens (default: 1024): Maximum #new tokens to generate
    • --system-message (default: None): system message (note it's not supported by some models)
    • --tensor-parallel-size: #GPUS for doing tensor parallelism (only for vLLM)
    • --languages (default: None): List of languages to evaluate (None means all)
    • --result-dir (default: "results"): Directory to save the model outputs and evaluation results
    • --clean-ctx-comments (default: "none"): Clean context comments with padding ("positional_padding") or no padding ("no_padding")
    • --eval-ignore-comments (default: False): During evaluation, ignore groundtruth and model comments
    • --trust-remote-code (default: False): allow remote code (for HuggingFace transformers and vLLM)
    • --attn-implementation (default: None): Use "flash_attention_2" if your HF hits OOM
  • Output:
    • results/ntoken_{code-context-size}/{model}.jsonl: Model generated outputs
    • results/ntoken_{code-context-size}/{model}-SCORE.json: Evaluation results

Compute Scores

By default, the repoqa.search_needle_function command will evaluate model outputs and compute scores after text generation. However, you can also separately compute scores using the following command:

repoqa.compute_score --model-output-path={model-output}.jsonl

Tip

  • Input: Path to the model generated outputs.
  • Output: The evaluation scores would be stored in {model-output}-SCORES.json

📚 Read More

Citation

@article{repoqa,
  title = {RepoQA: Evaluating Long Context Code Understanding},
  author = {Liu, Jiawei and Tian, Jia Le and Daita, Vijay and Wei, Yuxiang and Ding, Yifeng and Wang, Yuhan Katherine and Yang, Jun and Zhang, Lingming},
  year = {2024},
  journal = {arXiv preprint arXiv:2406.06025},
}