uv sync produces a large (9.5GB) docker layer upon build

[fortminors]

I have the following pyproject.toml:

[project]
name = "cv_tests"
version = "0.1.0"
description = "cv_tests"
authors = [
    {name="me", email="[email protected]"},
]
readme = "README.md"
requires-python = ">=3.11"
dependencies = [
    "circle-fit>=0.2.1",
    "einops>=0.8.0",
    "fastapi>=0.115.4",
    "imutils>=0.5.4",
    "matplotlib>=3.9.2",
    "open-clip-torch>=2.29.0",
    "opencv-contrib-python>=4.10.0.84",
    "opencv-python>=4.10.0.84",
    "pyrealsense2>=2.55.1.6486",
    "safetensors>=0.4.5",
    "scikit-image>=0.24.0",
    "scikit-learn>=1.5.2",
    "scipy>=1.14.1",
    "shapely>=2.0.6",
    "transformers>=4.46.1",
    "ultralytics>=8.3.27",
    "uvicorn>=0.32.0",
    "xformers>=0.0.28.post3",
    "sam-2",
    "setuptools>=75.3.0",
    "torch-tensorrt>=2.5.0",
    "pre-commit>=4.0.1",
    "lakefs>=0.7.1",
]

[tool.uv.sources]
sam-2 = { git = "https://github.com/facebookresearch/segment-anything-2.git", rev = "c2ec8e14a185632b0a5d8b161928ceb50197eddc" }

and the following Dockerfile:

FROM nvcr.io/nvidia/cuda:12.6.2-cudnn-devel-ubuntu22.04

# https://docs.docker.com/reference/dockerfile/#shell-and-exec-form
# https://manpages.ubuntu.com/manpages/noble/en/man1/sh.1.html
SHELL ["/bin/sh", "-exc"]

ARG DEBIAN_FRONTEND=noninteractive
ARG python_version=3.11

COPY --link --from=ghcr.io/astral-sh/uv:0.5.2 /uv /usr/local/bin/uv

RUN apt-get update --quiet && \
    apt-get upgrade -y && \
    apt-get install --quiet --no-install-recommends -y build-essential git ca-certificates \
    libgl1 libglib2.0-0 libusb-1.0-0-dev && \
    # Forcing http 1.1 to fix https://stackoverflow.com/q/59282476
    git config --global http.version HTTP/1.1 && \
    uv python install $python_version

ENV UV_PYTHON="python$python_version" \
    UV_PYTHON_DOWNLOADS=never \
    UV_PROJECT_ENVIRONMENT=/app \
    UV_LINK_MODE=copy \
    PYTHONFAULTHANDLER=1 \
    PYTHONUNBUFFERED=1 \
    UV_COMPILE_BYTECODE=1 \
    PYTHONOPTIMIZE=1

WORKDIR /project
COPY pyproject.toml uv.lock README.md /project

# Building deps
RUN --mount=type=cache,target=/root/.cache/uv \
    uv sync --no-dev --no-install-project --frozen

It builds successfully, however when the last RUN command is executed, it creates a huge layer of size 9.5 GB.
I inspected it with dive, to find that it is all taken by /app/lib/python3.11/site-packages

I thought that whenever I do RUN --mount=type=cache,target=/root/.cache/uv, the packages are cached internally in a local docker instance and it is not a part of the current image that I am building. Is there anything I can do about that? I thought of building a base (reusable) image with thick packages like pytorch and then inheriting from it and installing the other packages, so to split the installation in several layers, but then I would need several .toml project configurations and lock files. So I am curious what is the best way to do that? I am not sure if multi-stage build would help me here, but I might be mistaken.

[zanieb]

I don't quite follow the issue. uv sync installs the packages. If the layer is big because of installed packages, I'm not sure what we could do differently.

Using a cache mount can speed up installation, but won't remove the installed packages from the image because then it'd be broken.

You could try --no-install-package <big-package> to separate that one out from the rest.

[fortminors]

Thank you for the suggestion. I thought that if I use RUN --mount=type=cache,target=/root/.cache/uv, it will mount the packages and cache to the image, but now that I think of it, RUN --mount invokes a temporary mount during build time only, so launching scripts won't work - there will be no packages any longer.