Dockerfile

FROM nvidia/cuda:12.4.1-devel-ubuntu22.04

ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONUNBUFFERED=1
ENV HF_HOME=/app/model_cache
ENV TORCH_HOME=/app/torch_cache

# Install Python 3.12 + system deps
RUN apt-get update && apt-get install -y \
    software-properties-common \
    git \
    curl \
    ffmpeg \
    libsndfile1 \
    && add-apt-repository ppa:deadsnakes/ppa \
    && apt-get update \
    && apt-get install -y python3.12 python3.12-venv python3.12-dev \
    && curl -sS https://bootstrap.pypa.io/get-pip.py | python3.12 \
    && ln -sf /usr/bin/python3.12 /usr/bin/python \
    && ln -sf /usr/bin/python3.12 /usr/bin/python3 \
    && rm -rf /var/lib/apt/lists/*

# Remove conflicting system packages
RUN apt-get update && apt-get remove -y python3-blinker || true \
    && rm -rf /usr/lib/python3/dist-packages/blinker* \
    && rm -rf /var/lib/apt/lists/*

# Install Python dependencies
COPY requirements.txt /app/requirements.txt
WORKDIR /app
RUN pip install --no-cache-dir --break-system-packages -r requirements.txt

# flash-attn intentionally NOT built from source here: the compile needs ~10 GB
# RAM and gets OOM-killed on the HF Spaces builder, which fails the whole build
# (the `|| echo` fallback can't catch an OOM kill of the build process itself).
# vLLM runs fine on its default attention backend. For the ~1.5-2x speedup,
# install a prebuilt wheel matching torch+CUDA+python from
#   https://github.com/Dao-AILab/flash-attention/releases
# e.g.  RUN pip install <wheel-url>  --no-build-isolation

# Pre-download ASR model into the image (no persistent storage available on HF Spaces).
# NOTE: single-line python -c on purpose — ACR's Dockerfile dependency scanner
# chokes on backslash line-continuations inside a quoted python string.
RUN mkdir -p /app/model_cache /app/torch_cache && \
    python -c "from huggingface_hub import snapshot_download; snapshot_download('Qwen/Qwen3-ASR-1.7B', cache_dir='/app/model_cache')" && \
    echo "ASR model pre-downloaded"

# Pre-download Silero VAD model
RUN python -c "import torch; torch.hub.load('snakers4/silero-vad', 'silero_vad', trust_repo=True); print('Silero VAD cached')" || echo "VAD pre-download failed, will retry at runtime"

# Copy application code
COPY . /app

EXPOSE 7860

# Health check — long start period for model download + load
HEALTHCHECK --interval=30s --timeout=30s --start-period=600s --retries=3 \
    CMD curl -f http://localhost:7860/health || exit 1

CMD ["python", "main_async.py"]
# Rebuild trigger: Wed Apr 22 18:13:56 PKT 2026