# Inference-only dependencies — no training, no data-prep, no evaluation metrics. # PyTorch CUDA 12.6 wheels; PyPI as fallback for everything else. --index-url https://download.pytorch.org/whl/cu126 --extra-index-url https://pypi.org/simple/ # Core ML (inference) — torch/torchaudio installed in a separate Dockerfile layer transformers>=4.40.0 accelerate>=0.27.0 peft>=0.10.0 # needed to load LoRA adapter layers if checkpoint uses them # Audio soundfile>=0.12.1 soxr>=0.3.7 silero-vad>=6.0.0 # Gemini post-processing google-genai>=1.73.1 tenacity>=8.2.0 httpx>=0.27.0 # API server fastapi>=0.111.0 uvicorn[standard]>=0.29.0 python-multipart>=0.0.9 # Validation & config pydantic>=2.13.2 pydantic-settings>=2.3.0 # Utilities numpy>=1.24.0 python-dotenv>=1.2.2 huggingface-hub>=0.22.0