# RLVR Floating-Point Precision Experiment Dependencies # Core ML frameworks torch>=2.0.0 transformers>=4.36.0 accelerate>=0.25.0 # RL framework (VeRL) # Install from source: pip install git+https://github.com/volcengine/verl.git # verl # Inference vllm>=0.2.0 # Numerical computation numpy>=1.24.0 scipy>=1.10.0 # Visualization matplotlib>=3.7.0 # Progress tracking tqdm>=4.65.0 # Data handling datasets>=2.14.0 # Utilities pyyaml>=6.0 jsonlines>=3.1.0 # Distributed training (optional, usually comes with torch) # torch-distributed # Flash attention (optional, for faster inference) # flash-attn>=2.3.0 # Evaluation utilities # eval-chemy # Math verifier (install from RLVR repo)