Dockerfile.cuda

# Dockerfile.cuda — GPU build (CUDA + TensorRT + ASR)
#
# Target: NVIDIA Ada Lovelace (RTX 4060, L4, RTX 4090) — sm_89
# Override GPU_ARCH build arg for other architectures.
#
# OptiX (AR ray tracing) is not included — requires manual SDK download from
# developer.nvidia.com. To add OptiX support later, see docs/deployment.md.
#
# Build:
#   docker build -f Dockerfile.cuda -t switchframe:cuda .
#
# Run:
#   docker run -d --gpus all \
#     -p 8080:8080/udp -p 8080:8080/tcp \
#     -p 9090:9090/tcp -p 6464:6464/udp \
#     switchframe:cuda --demo --srt-listen :6464

# ---------- Stage 1: Build UI ----------
FROM node:22-bookworm-slim AS ui-builder
WORKDIR /app/ui
COPY ui/package.json ui/package-lock.json ./
RUN npm ci
COPY ui/ ./
RUN npm run build

# ---------- Stage 2: Build everything ----------
FROM nvidia/cuda:12.6.3-cudnn-devel-ubuntu24.04 AS builder

ARG GPU_ARCH=sm_89
ARG GO_VERSION=1.25.0
ARG WHISPER_CPP_VERSION=v1.7.5
ARG ONNXRUNTIME_VERSION=1.21.1

ENV DEBIAN_FRONTEND=noninteractive

# System packages + TensorRT dev
RUN apt-get update && apt-get install -y --no-install-recommends \
        build-essential pkg-config nasm cmake git curl wget python3 \
        autoconf automake libtool \
        libssl-dev zlib1g-dev \
        libopus-dev libopusfile-dev \
        libnvinfer-dev libnvinfer-plugin-dev libnvonnxparsers-dev \
    && rm -rf /var/lib/apt/lists/*

# Go
RUN curl -fsSL "https://go.dev/dl/go${GO_VERSION}.linux-amd64.tar.gz" \
    | tar -C /usr/local -xz
ENV PATH=/usr/local/go/bin:$PATH

# ---- From-source libraries (Debian ships versions too old) ----

# libsrt 1.5.4
RUN git clone --depth 1 --branch v1.5.4 https://github.com/Haivision/srt.git /tmp/srt \
    && cd /tmp/srt && mkdir build && cd build \
    && cmake .. -DCMAKE_INSTALL_PREFIX=/usr/local \
       -DENABLE_SHARED=ON -DENABLE_STATIC=OFF \
       -DENABLE_APPS=OFF -DENABLE_TESTING=OFF \
    && make -j"$(nproc)" && make install && ldconfig \
    && rm -rf /tmp/srt

# libfdk-aac 2.0.3
RUN git clone --depth 1 --branch v2.0.3 https://github.com/mstorsjo/fdk-aac.git /tmp/fdk-aac \
    && cd /tmp/fdk-aac && autoreconf -fiv \
    && ./configure --prefix=/usr/local --enable-shared --disable-static \
    && make -j"$(nproc)" && make install && ldconfig \
    && rm -rf /tmp/fdk-aac

# libx264
RUN git clone --depth 1 https://code.videolan.org/videolan/x264.git /tmp/x264 \
    && cd /tmp/x264 \
    && ./configure --prefix=/usr/local --enable-shared --disable-static \
       --enable-pic --disable-cli \
    && make -j"$(nproc)" && make install && ldconfig \
    && rm -rf /tmp/x264

# nv-codec-headers (NVENC API)
RUN git clone --depth 1 --branch n12.2.72.0 \
        https://github.com/FFmpeg/nv-codec-headers.git /tmp/nv-codec \
    && cd /tmp/nv-codec && make install PREFIX=/usr/local \
    && rm -rf /tmp/nv-codec

# FFmpeg 7.1 with NVENC + CUDA
RUN git clone --depth 1 --branch n7.1 https://github.com/FFmpeg/FFmpeg.git /tmp/FFmpeg \
    && cd /tmp/FFmpeg \
    && ./configure --prefix=/usr/local \
       --enable-gpl --enable-nonfree \
       --enable-libx264 --enable-libfdk-aac --enable-libopus --enable-libsrt \
       --enable-swresample --enable-swscale \
       --enable-shared --disable-static --disable-doc \
       --enable-nvenc --enable-cuda \
       --extra-cflags="-I/usr/local/cuda/include" \
       --extra-ldflags="-L/usr/local/cuda/lib64" \
    && make -j"$(nproc)" && make install && ldconfig \
    && rm -rf /tmp/FFmpeg

# ---- AI / ASR dependencies ----

# ONNX Runtime with CUDA EP (for Silero VAD)
RUN cd /opt \
    && wget -q "https://github.com/microsoft/onnxruntime/releases/download/v${ONNXRUNTIME_VERSION}/onnxruntime-linux-x64-gpu-${ONNXRUNTIME_VERSION}.tgz" \
    && tar xzf onnxruntime-linux-x64-gpu-${ONNXRUNTIME_VERSION}.tgz \
    && rm onnxruntime-linux-x64-gpu-${ONNXRUNTIME_VERSION}.tgz \
    && mv onnxruntime-linux-x64-gpu-${ONNXRUNTIME_VERSION} onnxruntime

# whisper.cpp with CUDA
# The CUDA driver stub is needed at link time (real libcuda.so comes from host at runtime)
RUN ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/libcuda.so \
    && git clone --depth 1 --branch "${WHISPER_CPP_VERSION}" \
        https://github.com/ggerganov/whisper.cpp.git /opt/whisper.cpp \
    && cd /opt/whisper.cpp && mkdir build && cd build \
    && cmake .. -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON \
       -DWHISPER_BUILD_EXAMPLES=OFF -DWHISPER_BUILD_TESTS=OFF \
    && make -j"$(nproc)"

# ---- Build SwitchFrame ----

WORKDIR /app

# Go module cache (changes less often than source)
COPY server/go.mod server/go.sum ./server/
RUN cd server && go mod download

# Copy source + embedded UI
COPY server/ ./server/
COPY --from=ui-builder /app/ui/build ./server/cmd/switchframe/ui

# CUDA kernels + TensorRT wrapper (no OptiX — SDK not available)
RUN cd server/gpu/cuda \
    && make ARCH="-arch=${GPU_ARCH}" clean all \
    && make tensorrt

# Main binary
RUN cd server \
    && CGO_CFLAGS="-I/usr/local/cuda/include \
       -I/opt/whisper.cpp/include -I/opt/whisper.cpp/ggml/include \
       -I/opt/onnxruntime/include" \
    CGO_LDFLAGS="-L/usr/local/lib -L/usr/local/cuda/lib64 \
       -L/opt/whisper.cpp/build/src \
       -L/opt/whisper.cpp/build/ggml/src \
       -L/opt/whisper.cpp/build/ggml/src/ggml-cuda \
       -L/opt/onnxruntime/lib" \
    go build -tags "embed_ui cuda tensorrt" \
       -o /switchframe ./cmd/switchframe

# ASR sidecar
RUN cd server \
    && CGO_CFLAGS="-I/usr/local/cuda/include \
       -I/opt/whisper.cpp/include -I/opt/whisper.cpp/ggml/include \
       -I/opt/onnxruntime/include" \
    CGO_LDFLAGS="-L/usr/local/lib -L/usr/local/cuda/lib64 \
       -L/opt/whisper.cpp/build/src \
       -L/opt/whisper.cpp/build/ggml/src \
       -L/opt/whisper.cpp/build/ggml/src/ggml-cuda \
       -L/opt/onnxruntime/lib" \
    go build -tags "cuda tensorrt" \
       -o /switchframe-asr ./cmd/switchframe-asr

# ---------- Stage 3: Runtime ----------
FROM nvidia/cuda:12.6.3-cudnn-runtime-ubuntu24.04

ENV DEBIAN_FRONTEND=noninteractive

RUN apt-get update && apt-get install -y --no-install-recommends \
        libnvinfer10 libnvinfer-plugin10 libnvonnxparsers10 \
        libcublas-12-6 \
        libnpp-12-6 \
        libopus0 libopusfile0 \
        libssl3t64 zlib1g libgomp1 \
        ca-certificates curl \
    && rm -rf /var/lib/apt/lists/*

# From-source shared libraries (FFmpeg, SRT, FDK-AAC, x264)
COPY --from=builder /usr/local/lib/libav*.so* /usr/local/lib/
COPY --from=builder /usr/local/lib/libsw*.so* /usr/local/lib/
COPY --from=builder /usr/local/lib/libpostproc*.so* /usr/local/lib/
COPY --from=builder /usr/local/lib/libx264*.so* /usr/local/lib/
COPY --from=builder /usr/local/lib/libfdk-aac*.so* /usr/local/lib/
COPY --from=builder /usr/local/lib/libsrt*.so* /usr/local/lib/

# ONNX Runtime
COPY --from=builder /opt/onnxruntime/lib/libonnxruntime*.so* /usr/local/lib/

# whisper.cpp + GGML
COPY --from=builder /opt/whisper.cpp/build/src/libwhisper.so* /usr/local/lib/
COPY --from=builder /opt/whisper.cpp/build/ggml/src/libggml*.so* /usr/local/lib/
COPY --from=builder /opt/whisper.cpp/build/ggml/src/ggml-cuda/libggml-cuda.so* /usr/local/lib/

RUN echo "/usr/local/lib" > /etc/ld.so.conf.d/local.conf && ldconfig

# Binaries
COPY --from=builder /switchframe /usr/local/bin/switchframe
COPY --from=builder /switchframe-asr /usr/local/bin/switchframe-asr

RUN useradd --system --create-home switchframe \
    && mkdir -p /home/switchframe/.switchframe \
    && chown -R switchframe:switchframe /home/switchframe/.switchframe
VOLUME /home/switchframe/.switchframe
USER switchframe

EXPOSE 8080 9090
EXPOSE 6464/udp 9000/udp

HEALTHCHECK --interval=30s --timeout=5s --retries=3 \
    CMD curl -f http://localhost:9090/health || exit 1

ENTRYPOINT ["switchframe", "--admin-addr=0.0.0.0:9090"]