WhiteFox: White-box Compiler Fuzzing Empowered by Large Language Models

Important

We are keeping improving the documents and adding more implementation details. Please stay tuned at README-DEV.md for more information.

About

• 🦊WhiteFox is the first white-box compiler fuzzer using LLMs with source-code information to test compiler optimization.
• WhiteFox adopts a dual-model framework: (i) an analysis LLM examines the low-level optimization source code and produces requirements on the high-level test programs that can trigger the optimization; (ii) a generation LLM produces test programs based on the summarized requirements.

Important

• WhiteFox has detected 96 bugs 🐛 in the PyTorch Inductor, Tensorflow XLA, TensorFlow Lite and LLVM, with 80 confirmed as previously unknown and 61 of them are fixed.

We provide a list of confirmed bug reports in bugs.

🔨 Getting Started

Prerequisites

1. Python version >= 3.9.0 (It must support f-string.)
   • highly recommend to use Python 3.9
2. Check our dependent python libraries in requirements.txt and install with:
   • pip install -r requirements.txt
3. Install StarCoder
   • Please follow the instructions in StarCoder.

Running WhiteFox

Step 1: Request Summarization

The prompts for NL generation are in Prompts with the format Prompts/{compiler}/src2req/{name}.txt.

If you want to generate the prompt by you own, take the prompt for torch-inductor as an example:

bash scripts/whitefox-torch-prompt-gen-src2req.sh
# Or
bash scripts/whitefox-torch-prompt-gen-src2req.sh {generated-prompt-dir}

The generated prompts will be in Prompts-generated by default.

After getting the prompts, you can run the following command to generate the requirements:

python gpt4.py --prompt-dir=Prompts/torch-inductor/src2req \ 
    --outdir=Requirements/torch-inductor/req \
    --temperature=0.0 \
    --batch-size=1

Before running the command, please put your OpenAI API key in openai.key:

echo {api_key} > openai.key

Step 2: Test Generation

First, you need to generate the prompts for the test generation based on the requirements:

bash scripts/whitefox-torch-prompt-gen-req2test.sh 

# Or
bash scripts/whitefox-torch-prompt-gen-req2test.sh {req-dir} {generated-prompt-dir}

The generated prompts will be in Prompts-generated by default.

Or you can use the prompts we generated in Prompts with the format Prompts/{compiler}/req2test/{name}.txt.

We leverage StarCoder to generate the tests based on the prompts.

[Option 1]: Local Mode (Recommended!)

We recoomend to use the local mode to generate the tests, which utilizes vllm.

You can execute the following command to generate the tests:

python starcoder_gen.py --hf-home={path-to-huggingface} --hf-cache={path-to-huggingface-cache} --prompt-dir=Prompts/torch-inductor/req2test ----output-dir=starcoder-generated --num=10 

The generated tests will be in starcoder-generated.

[Option 2]: Server Mode

You can execute the following command to generate the tests:

1. Run the starcoder server:

python starcoder_service.py --hf-home={path-to-huggingface} --hf-cache={path-to-huggingface-cache} --prompt-dir=starcoder-prompts --outdir=starcoder-generated --device='cuda:0' --num=10 --batch_size=10

2. Put the prompts in starcoder-prompts and the generated tests will be in starcoder-generated.

mkdir starcoder-prompts/torch-inductor
cp -r Prompts/torch-inductor/req2test starcoder-prompts/torch-inductor/

Step 3: Test Execution

You can execute the following command to execute the tests:

cd torch-exec && python run_torch.py --input-dir=../starcoder-generated/torch-inductor-generated/step1 --res-dir=_results-torch

The output of the execution will be in torch-exec/_results-torch.