# vLLM and its matching torch are installed separately, without uv.lock or
# the development environment's CUDA index overrides.
# Hugging Face reference fixtures load models with device_map="cuda".
# Keep the reference implementation fixed across vLLM candidates.
transformers==5.18.0
accelerate>=1.1.0
# FlashInfer builds sampling kernels at runtime.
ninja>=1.11
pytest>=9.0.2
pytest-asyncio>=1.3.0
python-dotenv>=1.2.1
ray>=2.54.0
syrupy>=5.1.0
