Dockerfile

# ─── Stage 1: Build llama-server from source ─────────────────────────────────
FROM python:3.11-slim AS llama-builder

RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential \
    cmake \
    git \
    pkg-config \
    libopenblas-dev \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /llama-build
RUN git clone --depth 1 https://github.com/ggerganov/llama.cpp .

RUN cmake -B build \
    -DGGML_BLAS=ON \
    -DGGML_BLAS_VENDOR=OpenBLAS \
    -DCMAKE_BUILD_TYPE=Release \
    -DLLAMA_BUILD_TESTS=OFF \
    && cmake --build build --target llama-server -j$(nproc)

# ─── Stage 2: Runtime image ───────────────────────────────────────────────────
FROM python:3.11-slim

LABEL org.opencontainers.image.title="Surya OCR 2"
LABEL org.opencontainers.image.description="CPU-powered document OCR using Surya OCR 2"

RUN apt-get update && apt-get install -y --no-install-recommends \
    libopenblas0 \
    libgomp1 \
    curl \
    && rm -rf /var/lib/apt/lists/*

# Copy llama-server binary from builder
COPY --from=llama-builder /llama-build/build/bin/llama-server /usr/local/bin/llama-server
RUN chmod +x /usr/local/bin/llama-server

# Create non-root user (required for HF Spaces)
RUN useradd -m -u 1000 appuser
WORKDIR /app

# Install Python deps as root first (faster layer caching)
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

# Copy app files
COPY app.py .

# HF Spaces runs as uid 1000
RUN chown -R appuser:appuser /app
USER appuser

# Writable cache dirs for HF model downloads & llama.cpp
ENV HF_HOME=/tmp/hf_cache
ENV SURYA_INFERENCE_BACKEND=llamacpp
# Use all available threads for llama.cpp
ENV OMP_NUM_THREADS=4
ENV GRADIO_SERVER_PORT=7860
ENV GRADIO_SERVER_NAME=0.0.0.0

EXPOSE 7860

CMD ["python", "app.py"]