Dockerfile.cuda-all

# syntax=docker/dockerfile:1
#
# Thin CUDA runtime image. The prebuilt CUDA package (the `mistralrs` binary plus its bundled
# `lib/` and, for cutile builds, `bin/tileiras`) is staged into `dist/` by the release workflow
# and copied in - not compiled here. BASE_TAG selects the CUDA runtime matching the binary's
# toolkit. Build via .github/workflows/release.yml.

ARG BASE_TAG=13.1.2-cudnn
FROM nvidia/cuda:${BASE_TAG}-runtime-ubuntu22.04 AS runtime
SHELL ["/bin/bash", "-e", "-o", "pipefail", "-c"]

ARG DEBIAN_FRONTEND=noninteractive
RUN <<HEREDOC
    apt-get update
    apt-get install -y --no-install-recommends \
        libomp-dev \
        ca-certificates \
        libssl-dev \
        curl

    rm -rf /var/lib/apt/lists/*
HEREDOC

# The binary carries rpath $ORIGIN/lib, so it finds its bundled CUDA libs beside it.
# TARGETARCH (amd64/arm64) is set by buildx; the release workflow stages dist/<arch>/.
ARG TARGETARCH
COPY dist/${TARGETARCH}/ /opt/mistralrs/
RUN ln -sf /opt/mistralrs/mistralrs /usr/local/bin/mistralrs && \
    if [ -f /opt/mistralrs/bin/tileiras ]; then ln -sf /opt/mistralrs/bin/tileiras /usr/local/bin/tileiras; fi
# Chat templates for models that ship without one
COPY chat_templates /chat_templates

# hf-hub reads HF_HOME; mount a volume at /data to persist downloaded models
ENV HF_HOME=/data

# Default port of `mistralrs serve`
EXPOSE 1234

ENTRYPOINT ["mistralrs"]
CMD ["--help"]