Dockerfile.judge

# nite-eval judge image
# Runs llama.cpp's llama-server directly. One image, two Deployments
# (Flow-Judge + RewardAnything), differentiated by JUDGE_MODEL_PATH and PORT env.
#
# Built for both Tesla P40 (sm_61) and RTX 3090 (sm_86) so the same image
# works on either GPU. Single-arch builds shave size/build time but lose
# portability across the homelab's mixed GPUs.

ARG LLAMACPP_REF=master
ARG CUDA_ARCHITECTURES="61;86"

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

ARG LLAMACPP_REF
ARG CUDA_ARCHITECTURES

ENV DEBIAN_FRONTEND=noninteractive

RUN apt-get update \
 && apt-get install -y --no-install-recommends \
        build-essential cmake git ca-certificates curl libcurl4-openssl-dev \
 && rm -rf /var/lib/apt/lists/*

WORKDIR /src
RUN git clone --depth 1 --branch ${LLAMACPP_REF} https://github.com/ggerganov/llama.cpp.git

WORKDIR /src/llama.cpp
RUN cmake -B build \
        -DGGML_CUDA=ON \
        -DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHITECTURES}" \
        -DLLAMA_CURL=ON \
        -DBUILD_SHARED_LIBS=OFF \
        -DCMAKE_BUILD_TYPE=Release \
 && cmake --build build --target llama-server --config Release -j


FROM nvidia/cuda:12.4.1-runtime-ubuntu22.04 AS runtime

ENV DEBIAN_FRONTEND=noninteractive
RUN apt-get update \
 && apt-get install -y --no-install-recommends curl ca-certificates tini libgomp1 \
 && rm -rf /var/lib/apt/lists/*

COPY --from=builder /src/llama.cpp/build/bin/llama-server /usr/local/bin/llama-server

# Defaults; overridden per-Deployment via env
ENV JUDGE_MODEL_PATH="" \
    PORT=9091 \
    NGL=999 \
    CTX_SIZE=4096 \
    NPARALLEL=1 \
    EXTRA_FLAGS=""

EXPOSE 9091
EXPOSE 9092

# Use shell form so env vars expand. tini reaps zombies + forwards SIGTERM.
ENTRYPOINT ["/usr/bin/tini", "--"]
CMD ["/bin/sh", "-c", "exec llama-server -m \"$JUDGE_MODEL_PATH\" --port \"$PORT\" -ngl \"$NGL\" --ctx-size \"$CTX_SIZE\" -np \"$NPARALLEL\" --no-webui $EXTRA_FLAGS"]