elizaos--eliza/docker/certification/Dockerfile.gpu

# syntax=docker/dockerfile:1
# GPU certification image for vast.ai runs (#14548, epic #14541).
#
# scripts/vast/run-certification.mjs rents a vast.ai RTX 4090 and boots this
# image; the onstart script it injects only clones the repo at one sha and
# runs the packages/evidence certification chain. vast caps onstart-cmd at
# 16 KB, so EVERYTHING heavy is baked here instead: CUDA-enabled llama-server
# (>= build b8525, the minimum scripts/gpu-vision accepts for the
# DeepSeek-OCR mmproj), the sha256-pinned OCR + VLM GGUFs from
# scripts/gpu-vision/models.lock.json, Node 24 + Bun (the workspace
# toolchain), Playwright Chromium with OS deps, tesseract, and ffmpeg.
#
# Secrets NEVER bake into this image — the Ed25519 signing key and any
# storage push command are injected as instance env at create time and die
# with the instance. Build/publish: .github/workflows/certification-image.yml
# pushes ghcr.io/elizaos/certification-gpu:{latest,sha-<short>}.

ARG CUDA_VERSION=12.4.1
ARG UBUNTU_VERSION=22.04

# --- Stage 1: build llama-server with the CUDA backend --------------------
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} AS llama-builder

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

# b8525 is the exact minimum scripts/gpu-vision/lib.mjs (MIN_LLAMA_BUILD)
# demands; every llama.cpp CI build is tagged bNNNN so the ref is stable.
ARG LLAMA_CPP_REF=b8525
RUN git clone --depth 1 --branch "${LLAMA_CPP_REF}" \
      https://github.com/ggml-org/llama.cpp /opt/llama.cpp

# Static-ish build (no shared ggml libs to carry over) for the two GPU
# generations the offer filter can realistically land on: sm_86 (RTX 3090)
# and sm_89 (RTX 4090, the default --gpu-name).
RUN cmake -S /opt/llama.cpp -B /opt/llama.cpp/build \
      -DCMAKE_BUILD_TYPE=Release \
      -DGGML_CUDA=ON \
      -DCMAKE_CUDA_ARCHITECTURES="86;89" \
      -DBUILD_SHARED_LIBS=OFF \
      -DLLAMA_CURL=ON \
    && cmake --build /opt/llama.cpp/build --target llama-server -j "$(nproc)"

# --- Stage 2: runtime ------------------------------------------------------
FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}

LABEL org.opencontainers.image.source="https://github.com/elizaOS/eliza" \
      org.opencontainers.image.description="elizaOS GPU certification runner (vast.ai) — CUDA llama-server + pinned vision models + node24/bun/playwright/tesseract/ffmpeg" \
      org.opencontainers.image.licenses="MIT"

ENV DEBIAN_FRONTEND=noninteractive

# openssh-server: vast's `runtype ssh` instances start sshd inside the
# container; without it the instance never leaves the loading state.
RUN apt-get update && apt-get install -y --no-install-recommends \
      ca-certificates curl git unzip xz-utils jq \
      openssh-server \
      tesseract-ocr \
      ffmpeg \
      libcurl4 libgomp1 \
    && rm -rf /var/lib/apt/lists/*

# Node 24 — the engines.node pin of the workspace (plain-node scripts like
# scripts/gpu-vision run under it directly).
RUN curl -fsSL https://deb.nodesource.com/setup_24.x | bash - \
    && apt-get install -y --no-install-recommends nodejs \
    && rm -rf /var/lib/apt/lists/* \
    && node --version

# Bun, pinned to the same version CI pins (floating canary rewrites bun.lock).
ARG BUN_VERSION=1.3.14
ENV BUN_INSTALL=/root/.bun
ENV PATH="${BUN_INSTALL}/bin:${PATH}"
RUN curl -fsSL https://bun.sh/install | bash -s "bun-v${BUN_VERSION}" \
    && bun --version

# Playwright Chromium + OS deps, pinned to the packages/evidence devDep line.
ARG PLAYWRIGHT_VERSION=1.61.0
ENV PLAYWRIGHT_BROWSERS_PATH=/opt/playwright-browsers
RUN npx --yes "playwright@${PLAYWRIGHT_VERSION}" install --with-deps chromium \
    && rm -rf /var/lib/apt/lists/*

# CUDA llama-server from the builder.
COPY --from=llama-builder /opt/llama.cpp/build/bin/llama-server /usr/local/bin/llama-server
# Build gate: a broken binary (missing lib, bad link) fails here, not on a
# billed vast instance. --version needs no GPU.
RUN llama-server --version

# Pinned GPU-vision models, fetched + sha256-verified by the repo's own
# setup script (lockfile resolves relative to lib.mjs, so the trio is copied
# together). ELIZA_GPU_VISION_CACHE points the cloned repo's serve.mjs at
# this baked cache at runtime — no model download on the billed instance.
ENV ELIZA_GPU_VISION_CACHE=/opt/eliza/gpu-vision-cache
COPY scripts/gpu-vision/setup.mjs scripts/gpu-vision/lib.mjs scripts/gpu-vision/models.lock.json /opt/eliza/gpu-vision/
RUN node /opt/eliza/gpu-vision/setup.mjs --with-vlm

WORKDIR /workspace
CMD ["/bin/bash"]