requirements.txt

# Core ML / inference
torch>=2.4.0
torchaudio>=2.4.0
transformers>=4.56,<5.3,!=5.0.*,!=5.1.*   # MoonshineForConditionalGeneration support
accelerate>=0.30.0
bitsandbytes>=0.43.0                       # 4-bit quant for MiniCPM on CPU
sentencepiece
protobuf

# ASR audio helpers
soundfile
librosa
huggingface_hub

# Qwen2.5 intent parser (llama-cpp-python with CUDA offload)
# Install with CUDA support; do not use the plain pip version.
# Run this separately after pip install -r requirements.txt:
#   CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --no-cache-dir
# Windows (MSVC):
#   set CMAKE_ARGS=-DGGML_CUDA=on && pip install llama-cpp-python --no-cache-dir
llama-cpp-python>=0.2.85     # plain CPU fallback; override with CUDA build above

# VAD
numpy>=1.24.0                # Install early to avoid dependency conflicts
silero-vad                   # pulls torch; also needs torchaudio for audio I/O
onnxruntime>=1.16.0          # ONNX runtime for Silero VAD
pyaudio                      # microphone access (Linux: apt install python3-pyaudio)

# Gradio UI
gradio>=4.40.0

# Utilities
python-dotenv                # optional: load HF_TOKEN from .env file
pydantic>=2.0                # JSON schema validation for Qwen output