memorylayer/memorylayer-embed-server/Dockerfile.cpu 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
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
# CPU build of the MemoryLayer Embed Server.
#
# Same server, same API, same entrypoint as the CUDA build — it just runs the
# providers that do not need a GPU:
#
#   single-vector : sentence-transformers (all-MiniLM-L6-v2, 384-d) — the server's
#                   own default provider
#   multi-vector  : ColPali / ModernVBERT, which runs on CPU (colpali-engine
#                   auto-selects cpu when CUDA is absent and loads in float32)
#
# Deliberately NOT included: `vllm` (GPU-only) and `ocr` (a second large
# transformers stack). Use the -cuda13 tag for those.
#
# This exists because the CUDA image cannot run on a laptop, which left a first-time
# user with no way to self-host real embeddings — only cloud providers, or the
# `hash` provider, which is lexical rather than semantic.
#
# Published as the UNSUFFIXED tag of the same repository the CUDA build uses:
#   scitrera/memorylayer-embed-server:<version>            <- this image
#   scitrera/memorylayer-embed-server:<version>-cuda13     <- GPU build

# ── Stage 0: pull the prebuilt proxy-sidecar binary from upstream ────
# Kept in parity with the CUDA build so the two images are drop-in swappable:
# the same entrypoint honours EMBED_SERVER_RUN_SIDECAR either way.
ARG AETHER_VERSION=latest
FROM ghcr.io/scitrera/aether:${AETHER_VERSION} AS aether-image

# ── Stage 1: build the venv ──────────────────────────────────────────
FROM python:3.12-slim AS builder

# Empty AETHER_CLIENT_VERSION installs the latest scitrera-aether-client from
# PyPI; supply a concrete version (e.g. ``0.2.0``) to pin.
ARG AETHER_CLIENT_VERSION=""

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

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

# uv, for its --torch-backend flag
RUN curl -LsSf https://astral.sh/uv/install.sh | sh
ENV PATH="/root/.local/bin:$PATH"

RUN python3 -m venv /opt/memorylayer-embed

# Source packages, relative to the oss/ build context.
COPY memorylayer-core-python /src/memorylayer-core-python
COPY memorylayer-embed-server /src/memorylayer-embed-server

# --torch-backend=cpu matters: torch's default PyPI wheels for linux/amd64 bundle
# CUDA runtime libraries, which would add gigabytes to an image that cannot use
# them. core-python[all] is pure Python (embeddings/llm/context/documents).
RUN set -eu; \
    if [ -n "${AETHER_CLIENT_VERSION}" ]; then \
        AETHER_PKG="scitrera-aether-client==${AETHER_CLIENT_VERSION}"; \
    else \
        AETHER_PKG="scitrera-aether-client"; \
    fi; \
    uv pip install --python /opt/memorylayer-embed/bin/python --torch-backend=cpu \
        "${AETHER_PKG}" \
        "/src/memorylayer-core-python[all]" \
        "/src/memorylayer-embed-server[local,colpali,google,observability]"

# ── Stage 2: runtime ─────────────────────────────────────────────────
FROM python:3.12-slim

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

RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        curl ca-certificates && \
    rm -rf /var/lib/apt/lists/*

RUN groupadd --gid 65532 memorylayer && \
    useradd --uid 65532 --gid memorylayer --create-home memorylayer && \
    # Pre-create the cache tree so a mounted HF cache volume lands inside an
    # existing memorylayer-owned .cache rather than a root-owned one synthesized
    # by docker — otherwise the model download hits EPERM on first run.
    mkdir -p /home/memorylayer/.cache/huggingface && \
    chown -R memorylayer:memorylayer /home/memorylayer && \
    chmod 0755 /home/memorylayer/.cache

COPY --from=builder /opt/memorylayer-embed /opt/memorylayer-embed
ENV PATH="/opt/memorylayer-embed/bin:$PATH" \
    VIRTUAL_ENV="/opt/memorylayer-embed"

COPY --from=aether-image /usr/local/bin/proxy-sidecar /usr/local/bin/proxy-sidecar

COPY --chown=memorylayer:memorylayer memorylayer-embed-server/configs/proxy-sidecar.default.yaml /etc/aether-sidecar.yaml
COPY --chown=memorylayer:memorylayer --chmod=0755 memorylayer-embed-server/scripts/entrypoint.sh /usr/local/bin/embed-server-entrypoint.sh

WORKDIR /app
USER memorylayer

# The multi-vector default is vllm_subprocess, which needs a GPU. Point it at the
# in-process ColPali provider and pin both model devices to CPU, so this image is
# correct out of the box rather than only after the operator finds three env vars.
ENV MEMORYLAYER_EMBED_MULTI_VECTOR_PROVIDER=colpali_inprocess \
    MEMORYLAYER_EMBEDDING_DEVICE=cpu \
    MEMORYLAYER_EMBEDDING_ST_DEVICE=cpu

EXPOSE 61051

# start-period is generous: the first run downloads model weights (~90 MB for
# MiniLM, ~250 MB for ColPali) before the server can answer.
HEALTHCHECK --interval=30s --timeout=10s --start-period=180s --retries=3 \
    CMD curl -f http://localhost:61051/health || exit 1

ENTRYPOINT ["/usr/local/bin/embed-server-entrypoint.sh"]