# Optional GGUF backend. Gemma keeps the 0.3.34 floor; Qwen adapters require
# 0.3.35 or newer. Use the CUDA wheel index matching the host CUDA runtime.
llama-cpp-python>=0.3.34,<0.4
