PRO-V: An Efficient Program Generation Multi-Agent System for Automatic RTL Verification

📋 Table of Contents

• Background
• Experimental Results
• Quick Start
• Environment Setup
• Usage
• Resources

Background

PRO-V is a comprehensive framework for automated Register Transfer Level (RTL) code generation using reinforcement learning and multi-agent systems. The framework addresses the challenge of generating high-quality, functionally correct Verilog code through an iterative verification and refinement process.

Our approach combines:

• Multi-Agent System: Coordinated agents for code generation, testbench creation, simulation, and debugging
• Reinforcement Learning: Policy optimization using verification outcomes as rewards
• Iterative Refinement: Automated debugging loop to fix compilation and simulation errors

Experimental Results

Main Results on VerilogEval-v2

Note: Bold indicates best performance, italic indicates second best.

Ablation Study on VerilogEval-v2

Progressive improvements by adding PRO-V Flow, SFT, and RL to the base model:

Model / Method	Eval0	Eval1	Eval2-80	Eval2-90	Eval2-100
Qwen3-8B	0.6	0.6	0.6	0.6	0.0
Qwen3-8B w/ CorrectBench	47.4	25.6	23.7	23.1	21.8
Qwen3-8B w/ PRO-V sys	64.7	39.1	35.9	32.1	28.2
PRO-V-DS-8B w/ PRO-V sys	93.6	50.6	35.9	34.0	27.6
PRO-V-R1-8B w/ PRO-V sys	94.9	57.7	44.2	37.8	34.0

🚀 Quick Start

Step 1: Environment Setup

# Clone the repository
git clone https://github.com/YourRepo/PRO-V.git
cd PRO-V

# Create conda environment
conda create -n pro-v python=3.11
conda activate pro-v

# Install the package
pip install -e .

Step 2: Download Model

Download the PRO-V-R1-8B model from HuggingFace:

# Using huggingface-cli
huggingface-cli download YujieZhao/PRO-V-R1-8B --local-dir ./models/PRO-V-R1-8B

# Or using git
git lfs install
git clone https://huggingface.co/YujieZhao/PRO-V-R1-8B ./models/PRO-V-R1-8B

Step 3: Configure Evaluation Scripts

Edit the model path in the evaluation scripts:

For scripts/run_evaluation_think_simple.sh:

MODEL_PATH="./models/PRO-V-R1-8B"
SERVED_MODEL_NAME="PRO-V-R1-8B"

For scripts/run_evaluation_think.sh:

MODEL_PATH="./models/PRO-V-R1-8B"
SERVED_MODEL_NAME="PRO-V-R1-8B"
EXPERIMENT_NAME="pro_v_r1_8b_eval"

Step 4: Run Evaluation

cd scripts

# Simple evaluation (single process)
bash run_evaluation_think_simple.sh

# Or advanced evaluation (Ray-based parallel processing)
bash run_evaluation_think.sh

📦 Environment Setup

1.> Install the Repository

git clone https://github.com/YourRepo/PRO-V.git
cd PRO-V

# Get submodules (benchmark datasets)
git submodule update --init --recursive

# Create conda environment
conda create -n pro-v python=3.11
conda activate pro-v

# Install the package
pip install -e .

2.> Install Icarus Verilog (v12)

Ubuntu (Local Compilation)

apt install -y autoconf gperf make gcc g++ bison flex

git clone https://github.com/steveicarus/iverilog.git && cd iverilog
git checkout v12-branch
sh ./autoconf.sh && ./configure && make -j4
sudo make install

MacOS

brew install icarus-verilog

Version Confirmation

Verify the installation:

iverilog -v

Expected output (first line):

Icarus Verilog version 12.0 (stable) (v12_0)

3.> Install Verilator

# By apt (Ubuntu/Debian)
sudo apt install verilator

# By Homebrew (MacOS)
brew install verilator

# By Compilation
git clone https://github.com/verilator/verilator
cd verilator
autoconf
export VERILATOR_ROOT=`pwd`
./configure
make -j4

📖 Usage

Running Evaluations

PRO-V provides two evaluation scripts:

1. Simple Evaluation (Single Process)

Edit scripts/run_evaluation_think_simple.sh:

# Configure GPU and model
GPU_IDS="0,1"  # GPUs to use
MODEL_PATH="./models/PRO-V-R1-8B"
SERVED_MODEL_NAME="PRO-V-R1-8B"
EXPERIMENT_NAME="pro_v_eval"

# Parallelism settings
TP_SIZE=2  # Tensor parallel size
DP_SIZE=1  # Data parallel size

Run the evaluation:

cd scripts
bash run_evaluation_think_simple.sh

2. Advanced Evaluation (Ray-based Parallel)

Edit scripts/run_evaluation_think.sh:

# Configure GPU and model
GPU_IDS="0,1,2,3"  # GPUs to use
MODEL_PATH="./models/PRO-V-R1-8B"
SERVED_MODEL_NAME="PRO-V-R1-8B"
EXPERIMENT_NAME="pro_v_eval_parallel"

# Parallelism settings
TP_SIZE=4  # Tensor parallel size
DP_SIZE=1  # Number of replicas
MAX_CONCURRENCY=50  # Concurrent tasks

Run the evaluation:

cd scripts
bash run_evaluation_think.sh

Customizing Evaluation Parameters

Both scripts support various parameters:

# Sampling parameters
TEMPERATURE=0          # Temperature for generation
TOP_P=0.95            # Top-p sampling
MAX_TOKEN=20000       # Maximum tokens per generation

# Task filtering
TASK_NUMBERS="1,2,3"  # Specific tasks to run
FILTER_INSTANCE="^(Prob011_norgate)$"  # RegEx filter

# Benchmark selection
FOLDER_PATH="./verilog-eval/HDLBits/test_benchmark_new.json"

Output Structure

Results are saved to outputs/<EXPERIMENT_NAME>/:

outputs/
└── pro_v_eval/
    ├── all_results.json      # Aggregated results
    ├── overall_stats.json    # Statistics
    ├── task_1/              # Individual task outputs
    ├── task_2/
    └── ...