requirements.txt

ninja  # For faster builds.
psutil
ray >= 2.9
sentencepiece  # Required for LLaMA tokenizer.
numpy
torch == 2.1.2
transformers >= 4.38.0  # Required for Gemma.
xformers == 0.0.23.post1  # Required for CUDA 12.1.
fastapi
uvicorn[standard]
pydantic >= 2.0  # Required for OpenAI server.
prometheus_client
pynvml == 11.5.0
triton >= 2.1.0
cupy-cuda12x == 12.1.0  # Required for CUDA graphs. CUDA 11.8 users should install cupy-cuda11x instead.