Dockerfile.libs

# Dockerfile.libs — build linux-amd64 static libraries for llama.cpp and whisper.cpp
#
# Usage (CPU, default):
#   docker build -f Dockerfile.libs -o ./out .
#
# Usage (CUDA):
#   docker build -f Dockerfile.libs --build-arg GPU_BACKEND=cuda -o ./out .
#
# Usage (Vulkan):
#   docker build -f Dockerfile.libs --build-arg GPU_BACKEND=vulkan -o ./out .
#
# Build + link test (ensures .a files link correctly):
#   docker build -f Dockerfile.libs --target build-test .
#
# This extracts prebuilt .a files + headers into ./out/ on the host.

ARG GPU_BACKEND=cpu

# ============================================================================
# Stage: Download sources (shared by all backends)
# ============================================================================
FROM golang:1.24-bookworm AS sources

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

WORKDIR /src

# Copy version files so we can read versions from Go
COPY go.mod ./
COPY version.go ./
COPY cmd/versioncmd/ ./cmd/versioncmd/

# Download llama.cpp
RUN LLAMA_VERSION=$(go run ./cmd/versioncmd llama.cpp) && \
    echo "Downloading llama.cpp ${LLAMA_VERSION}..." && \
    wget -qO llama.cpp.tar.gz "https://github.com/ggerganov/llama.cpp/archive/refs/tags/${LLAMA_VERSION}.tar.gz" && \
    mkdir -p llama-src && \
    tar xzf llama.cpp.tar.gz --strip-components=1 -C llama-src && \
    rm llama.cpp.tar.gz

# Download whisper.cpp
RUN WHISPER_VERSION=$(go run ./cmd/versioncmd whisper.cpp) && \
    echo "Downloading whisper.cpp ${WHISPER_VERSION}..." && \
    wget -qO whisper.cpp.tar.gz "https://github.com/ggerganov/whisper.cpp/archive/refs/tags/${WHISPER_VERSION}.tar.gz" && \
    mkdir -p whisper-src && \
    tar xzf whisper.cpp.tar.gz --strip-components=1 -C whisper-src && \
    rm whisper.cpp.tar.gz

# ============================================================================
# Builder: CPU (default)
# ============================================================================
FROM golang:1.24-bookworm AS builder-cpu

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

WORKDIR /src
COPY --from=sources /src/llama-src llama-src
COPY --from=sources /src/whisper-src whisper-src

# Build llama.cpp (CPU)
RUN cd llama-src && \
    cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF && \
    cmake --build build --config Release -j$(nproc)

# Build whisper.cpp (CPU)
RUN cd whisper-src && \
    cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF && \
    cmake --build build --config Release -j$(nproc)

# Collect llama.cpp artifacts and strip debug symbols
RUN mkdir -p /out/llama.cpp/linux-amd64 /out/llama.cpp/include /out/llama.cpp/ggml/include /out/llama.cpp/common && \
    find llama-src/build -name "*.a" -exec cp {} /out/llama.cpp/linux-amd64/ \; && \
    find /out/llama.cpp/linux-amd64 -name "*.a" -exec strip --strip-debug {} \; && \
    cp llama-src/include/*.h /out/llama.cpp/include/ && \
    cp llama-src/ggml/include/*.h /out/llama.cpp/ggml/include/ && \
    cp llama-src/common/common.h /out/llama.cpp/common/ && \
    cp llama-src/common/sampling.h /out/llama.cpp/common/

# Collect whisper.cpp artifacts and strip debug symbols
RUN mkdir -p /out/whisper.cpp/linux-amd64 /out/whisper.cpp/include /out/whisper.cpp/ggml/include && \
    find whisper-src/build -name "*.a" -exec cp {} /out/whisper.cpp/linux-amd64/ \; && \
    find /out/whisper.cpp/linux-amd64 -name "*.a" -exec strip --strip-debug {} \; && \
    cp whisper-src/include/*.h /out/whisper.cpp/include/ && \
    cp whisper-src/ggml/include/*.h /out/whisper.cpp/ggml/include/

# ============================================================================
# Builder: CUDA
# ============================================================================
FROM nvidia/cuda:12.8.0-devel-ubuntu24.04 AS builder-cuda

RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential cmake wget && \
    rm -rf /var/lib/apt/lists/*

WORKDIR /src
COPY --from=sources /src/llama-src llama-src
COPY --from=sources /src/whisper-src whisper-src

# Build llama.cpp (CUDA)
RUN cd llama-src && \
    cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON && \
    cmake --build build --config Release -j$(nproc)

# Build whisper.cpp (CUDA)
RUN cd whisper-src && \
    cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON && \
    cmake --build build --config Release -j$(nproc)

# Collect llama.cpp artifacts (CUDA variant) and strip debug symbols
RUN mkdir -p /out/llama.cpp/linux-amd64-cuda /out/llama.cpp/include /out/llama.cpp/ggml/include /out/llama.cpp/common && \
    find llama-src/build -name "*.a" -exec cp {} /out/llama.cpp/linux-amd64-cuda/ \; && \
    find /out/llama.cpp/linux-amd64-cuda -name "*.a" -exec strip --strip-debug {} \; && \
    cp llama-src/include/*.h /out/llama.cpp/include/ && \
    cp llama-src/ggml/include/*.h /out/llama.cpp/ggml/include/ && \
    cp llama-src/common/common.h /out/llama.cpp/common/ && \
    cp llama-src/common/sampling.h /out/llama.cpp/common/

# Collect whisper.cpp artifacts (CUDA variant) and strip debug symbols
RUN mkdir -p /out/whisper.cpp/linux-amd64-cuda /out/whisper.cpp/include /out/whisper.cpp/ggml/include && \
    find whisper-src/build -name "*.a" -exec cp {} /out/whisper.cpp/linux-amd64-cuda/ \; && \
    find /out/whisper.cpp/linux-amd64-cuda -name "*.a" -exec strip --strip-debug {} \; && \
    cp whisper-src/include/*.h /out/whisper.cpp/include/ && \
    cp whisper-src/ggml/include/*.h /out/whisper.cpp/ggml/include/

# ============================================================================
# Builder: Vulkan
# ============================================================================
# Use Ubuntu 24.04 for Vulkan — bookworm's Vulkan 1.3.239 is too old
# (llama.cpp b8220+ needs VK_EXT_layer_settings from Vulkan 1.3.261+)
FROM ubuntu:24.04 AS builder-vulkan

RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential cmake wget ca-certificates libvulkan-dev glslang-tools glslc && \
    rm -rf /var/lib/apt/lists/*

WORKDIR /src
COPY --from=sources /src/llama-src llama-src
COPY --from=sources /src/whisper-src whisper-src

# Build llama.cpp (Vulkan)
RUN cd llama-src && \
    cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DGGML_VULKAN=ON && \
    cmake --build build --config Release -j$(nproc)

# Build whisper.cpp (Vulkan)
RUN cd whisper-src && \
    cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DGGML_VULKAN=ON && \
    cmake --build build --config Release -j$(nproc)

# Collect llama.cpp artifacts (Vulkan variant) and strip debug symbols
RUN mkdir -p /out/llama.cpp/linux-amd64-vulkan /out/llama.cpp/include /out/llama.cpp/ggml/include /out/llama.cpp/common && \
    find llama-src/build -name "*.a" -exec cp {} /out/llama.cpp/linux-amd64-vulkan/ \; && \
    find /out/llama.cpp/linux-amd64-vulkan -name "*.a" -exec strip --strip-debug {} \; && \
    cp llama-src/include/*.h /out/llama.cpp/include/ && \
    cp llama-src/ggml/include/*.h /out/llama.cpp/ggml/include/ && \
    cp llama-src/common/common.h /out/llama.cpp/common/ && \
    cp llama-src/common/sampling.h /out/llama.cpp/common/

# Collect whisper.cpp artifacts (Vulkan variant) and strip debug symbols
RUN mkdir -p /out/whisper.cpp/linux-amd64-vulkan /out/whisper.cpp/include /out/whisper.cpp/ggml/include && \
    find whisper-src/build -name "*.a" -exec cp {} /out/whisper.cpp/linux-amd64-vulkan/ \; && \
    find /out/whisper.cpp/linux-amd64-vulkan -name "*.a" -exec strip --strip-debug {} \; && \
    cp whisper-src/include/*.h /out/whisper.cpp/include/ && \
    cp whisper-src/ggml/include/*.h /out/whisper.cpp/ggml/include/

# ============================================================================
# Dynamic backend selection — picks the right builder stage
# ============================================================================
FROM builder-${GPU_BACKEND} AS builder

# ============================================================================
# Build test bases — provide the right link libraries per backend
# ============================================================================
FROM golang:1.24-bookworm AS build-test-base-cpu
RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential libgomp1 && \
    rm -rf /var/lib/apt/lists/*

FROM nvidia/cuda:12.8.0-devel-ubuntu24.04 AS build-test-base-cuda
RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential libgomp1 wget && \
    rm -rf /var/lib/apt/lists/* && \
    wget -qO go.tar.gz https://go.dev/dl/go1.24.6.linux-amd64.tar.gz && \
    tar -C /usr/local -xzf go.tar.gz && rm go.tar.gz
ENV PATH="/usr/local/go/bin:${PATH}"

FROM ubuntu:24.04 AS build-test-base-vulkan
RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential libgomp1 libvulkan-dev wget ca-certificates && \
    rm -rf /var/lib/apt/lists/* && \
    wget -qO go.tar.gz https://go.dev/dl/go1.24.6.linux-amd64.tar.gz && \
    tar -C /usr/local -xzf go.tar.gz && rm go.tar.gz
ENV PATH="/usr/local/go/bin:${PATH}"

# ============================================================================
# Build test — verifies the .a files link correctly with Go CGO
# ============================================================================
FROM build-test-base-${GPU_BACKEND} AS build-test

ARG GPU_BACKEND=cpu

WORKDIR /src
COPY . .

# Determine prebuilt directory suffix based on GPU_BACKEND
RUN SUFFIX="" && \
    if [ "$GPU_BACKEND" = "cuda" ]; then SUFFIX="-cuda"; fi && \
    if [ "$GPU_BACKEND" = "vulkan" ]; then SUFFIX="-vulkan"; fi && \
    echo "Using prebuilt suffix: linux-amd64${SUFFIX}"

# Copy freshly built .a files into the source tree
COPY --from=builder /out/llama.cpp/linux-amd64*/ /tmp/llama-libs/
COPY --from=builder /out/llama.cpp/include/ /src/ggml/llamacpp/third_party/include/
COPY --from=builder /out/llama.cpp/ggml/include/ /src/ggml/llamacpp/third_party/ggml/include/
COPY --from=builder /out/llama.cpp/common/ /src/ggml/llamacpp/third_party/common/
COPY --from=builder /out/whisper.cpp/linux-amd64*/ /tmp/whisper-libs/
COPY --from=builder /out/whisper.cpp/include/ /src/ggml/whispercpp/third_party/include/
COPY --from=builder /out/whisper.cpp/ggml/include/ /src/ggml/whispercpp/third_party/ggml/include/

# Copy .a files to correct prebuilt directory based on GPU_BACKEND
RUN SUFFIX="" && \
    if [ "$GPU_BACKEND" = "cuda" ]; then SUFFIX="-cuda"; fi && \
    if [ "$GPU_BACKEND" = "vulkan" ]; then SUFFIX="-vulkan"; fi && \
    mkdir -p /src/ggml/llamacpp/third_party/prebuilt/linux-amd64${SUFFIX} && \
    mkdir -p /src/ggml/whispercpp/third_party/prebuilt/linux-amd64${SUFFIX} && \
    cp /tmp/llama-libs/*.a /src/ggml/llamacpp/third_party/prebuilt/linux-amd64${SUFFIX}/ && \
    cp /tmp/whisper-libs/*.a /src/ggml/whispercpp/third_party/prebuilt/linux-amd64${SUFFIX}/

# Determine build tags based on GPU_BACKEND
RUN LLAMA_TAGS="llamacpp" && \
    WHISPER_TAGS="whispercpp" && \
    if [ "$GPU_BACKEND" = "cuda" ]; then LLAMA_TAGS="llamacpp,cuda"; WHISPER_TAGS="whispercpp,cuda"; fi && \
    if [ "$GPU_BACKEND" = "vulkan" ]; then LLAMA_TAGS="llamacpp,vulkan"; WHISPER_TAGS="whispercpp,vulkan"; fi && \
    echo "Build tags: llama=${LLAMA_TAGS} whisper=${WHISPER_TAGS}" && \
    CGO_ENABLED=0 go build ./ggml/llamacpp/... && \
    CGO_ENABLED=0 go build ./ggml/whispercpp/... && \
    echo "stub builds OK" && \
    CGO_ENABLED=1 go build -tags "${LLAMA_TAGS}" ./ggml/llamacpp/... && \
    echo "llamacpp CGO build OK (${LLAMA_TAGS})" && \
    CGO_ENABLED=1 go build -tags "${WHISPER_TAGS}" ./ggml/whispercpp/... && \
    echo "whispercpp CGO build OK (${WHISPER_TAGS})"

# Run stub tests
RUN CGO_ENABLED=0 go test ./ggml/llamacpp/... && \
    CGO_ENABLED=0 go test ./ggml/whispercpp/... && \
    echo "all tests passed"

# Output stage — docker build -o extracts from here
FROM scratch
COPY --from=builder /out/ /