memorylayer/memorylayer-embed-server/Dockerfile.real-test-full at main · scitrera/memorylayer · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
# MemoryLayer Embed Server — full real-model integration test image.
#
# Build context: repo root (scitrera-memorylayer-ai-cc/)
#   docker build -f oss/memorylayer-embed-server/Dockerfile.real-test-full \
#                -t memorylayer-embed-server:real-test-full .
#
# Same shape as Dockerfile.real-test but installs BOTH the [vllm] and
# [colpali] extras so:
#   * Single-vector requests go through vLLM with ``Qwen/Qwen3-VL-Embedding-2B``
#     (the real production single-vector provider).
#   * Multi-vector requests go through ColPali with
#     ``ModernVBERT/colmodernvbert``.
#
# Heaviest of the three test variants. Expect:
#   * ~5-10 minutes for ``pip install vllm`` (torch + CUDA kernels).
#   * ~3-5 minutes for first-request vLLM model warm-up.
#   * Several GB of disk for vLLM wheels + model weights.

FROM nvidia/cuda:13.1.1-runtime-ubuntu24.04 AS base

ENV PYTHONDONTWRITEBYTECODE=1 \
    PYTHONUNBUFFERED=1 \
    DEBIAN_FRONTEND=noninteractive

RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        python3.12 python3.12-venv python3.12-dev \
        build-essential curl ca-certificates && \
    ln -sf /usr/bin/python3.12 /usr/bin/python3 && \
    ln -sf /usr/bin/python3 /usr/bin/python && \
    rm -rf /var/lib/apt/lists/*

RUN curl -LsSf https://astral.sh/uv/install.sh | sh
ENV PATH="/root/.local/bin:$PATH"

RUN groupadd --gid 65532 memorylayer && \
    useradd --uid 65532 --gid memorylayer --create-home memorylayer

RUN python3 -m venv /opt/memorylayer-embed
ENV PATH="/opt/memorylayer-embed/bin:$PATH" \
    VIRTUAL_ENV="/opt/memorylayer-embed"

# Copy source packages
COPY oss/memorylayer-core-python /src/memorylayer-core-python
COPY oss/memorylayer-embed-server /src/memorylayer-embed-server

# Install OSS server + embed-server with BOTH vllm and colpali extras.
# --torch-backend=auto resolves the correct CUDA torch wheels.
RUN uv pip install --python /opt/memorylayer-embed/bin/python --torch-backend=auto \
        /src/memorylayer-core-python \
        "/src/memorylayer-embed-server[vllm,colpali]"

WORKDIR /app
USER memorylayer

# Plain-HTTP mode; both real providers wired separately.
ENV EMBED_SERVER_RUN_SIDECAR=false \
    MEMORYLAYER_EMBED_USE_MOCK_PROVIDERS=false \
    MEMORYLAYER_EMBED_USE_MULTI_FOR_SINGLE=false \
    MEMORYLAYER_EMBED_PRELOAD_MODELS=false \
    MEMORYLAYER_EMBEDDING_DEVICE="cuda"

EXPOSE 61051

# Long start_period — both vLLM and ColPali load on first request and
# vLLM in particular can take several minutes to warm.
HEALTHCHECK --interval=15s --timeout=5s --start-period=120s --retries=20 \
    CMD curl -fsS http://localhost:61051/health || exit 1

ENTRYPOINT ["memorylayer-embed", "serve"]