Dockerfile.brain

# Stage 1: Build - Pulling from NVIDIA's NGC registry (nvcr.io)
FROM nvcr.io/nvidia/cuda:12.9.0-devel-ubuntu22.04 AS build

# Install build dependencies
RUN apt-get update && apt-get install -y build-essential cmake git libcurl4-openssl-dev

# Create the missing .so.1 symlink for the CUDA driver stub and add it to the library path
RUN ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1
ENV LD_LIBRARY_PATH="/usr/local/cuda/lib64/stubs:${LD_LIBRARY_PATH}"

# Clone the latest llama.cpp master branch (critical for Gemma 4 support)
RUN git clone https://github.com/ggml-org/llama.cpp.git /app
WORKDIR /app

# Build with CUDA enabled.
# Multi-arch fat binary: covers Turing (RTX 20xx) through Blackwell (RTX 50xx).
# The NVIDIA driver picks the right cubin at load time, so this ONE image
# runs on sm_75 / 80 / 86 / 89 / 90 / 100 / 120.
# (Add 60;61;70 below if you must support Pascal/Volta — CUDA 12.9 still
#  compiles them with deprecation warnings; CUDA 13 removes them entirely.)
# NOTE: do NOT use CMAKE_CUDA_ARCHITECTURES=native here — that would bake in
# only the build machine's GPU and break everyone else.
RUN cmake -B build \
    -DGGML_CUDA=ON \
    -DCMAKE_CUDA_ARCHITECTURES="75;80;86;89;90;100;120" \
    -DCMAKE_BUILD_TYPE=Release
RUN cmake --build build --config Release -j$(nproc)

# Stage 2: Run - Match the 12.9.0 runtime from NGC
FROM nvcr.io/nvidia/cuda:12.9.0-runtime-ubuntu22.04

# Install runtime dependencies
RUN apt-get update && apt-get install -y libcurl4-openssl-dev libgomp1 && rm -rf /var/lib/apt/lists/*

# Keep everything in one directory
WORKDIR /app
COPY --from=build /app/build/bin/ /app/

# Tell the OS where to look for the dynamic backends
ENV LD_LIBRARY_PATH="/app:${LD_LIBRARY_PATH}"

ENTRYPOINT ["/app/llama-server"]