aTrain/Dockerfile.dev at develop · aTrainTranscription/aTrain · 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
FROM python:3.11-slim-bookworm

# git: install aTrain_core from its GitHub ref (over HTTPS)
# gcc/g++: compile native extensions from aTrain_core deps
# curl/ca-certificates: required to run the official uv installer below
RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        git \
        gcc \
        g++ \
        curl \
        ca-certificates \
    && rm -rf /var/lib/apt/lists/*

# uv: matches the workflow documented in CONTRIBUTING.md (uv sync / uv run).
# Use the official standalone installer; UV_INSTALL_DIR places the binaries
# on the system PATH so subsequent layers can call `uv` directly.
ENV UV_INSTALL_DIR=/usr/local/bin
RUN curl -LsSf https://astral.sh/uv/install.sh | sh

WORKDIR /workspace

# Copy only dependency metadata first to leverage Docker layer caching.
# The actual source is bind-mounted at runtime, but uv needs pyproject.toml
# and uv.lock here to resolve and install dependencies, plus version.py
# because the project version is read dynamically from it.
COPY pyproject.toml uv.lock ./
COPY aTrain/version.py aTrain/version.py

# uv settings tuned for container builds:
#   UV_LINK_MODE=copy        — bind-mounted /workspace is on a different
#                              filesystem than /workspace/.venv, so hardlinks
#                              would fail; copy is the safe default.
#   UV_COMPILE_BYTECODE=1    — pre-compile .pyc on install for faster startup.
#   UV_PYTHON_DOWNLOADS=never — use the system Python from this image, do not
#                              let uv fetch its own interpreter.
#   UV_HTTP_TIMEOUT=300      — the Linux lock pins torch==2.9.1+cu128, which
#                              pulls the large nvidia-*-cu12 CUDA wheels (cublas
#                              ~570 MiB, cusparselt ~270 MiB) from
#                              download.pytorch.org. The default per-request
#                              timeout aborts these mid-download on slower
#                              links; 300s lets them complete.
ENV UV_LINK_MODE=copy \
    UV_COMPILE_BYTECODE=1 \
    UV_PYTHON_DOWNLOADS=never \
    UV_HTTP_TIMEOUT=300

# Sync the locked dependency set. This image is built for linux/amd64 (pinned
# in docker-compose.yml), so the x86_64 wheels in uv.lock — CUDA torch,
# torchcodec, pyannote.audio, and aTrain_core@develop — all resolve directly.
# This matches CONTRIBUTING.md's documented `uv sync` and what CI installs.
#
# The only deviation is the excluded packages below, all build/UX concerns
# rather than platform concerns. The container runs aTrain with --no-native
# (browser mode), so neither pywebview GUI backend is ever loaded:
#   - pyqt5/pyqt6 chain + qtpy: need the Qt SDK to build from source.
#   - pygobject/pygobject-stubs: pywebview's GTK backend. pygobject pulls in
#     pycairo, which builds from an sdist and needs the cairo/glib native
#     libraries plus pkg-config — none of which the GTK backend needs here.
#   - dask: only used by full transcription, not by UI dev.
RUN --mount=type=cache,target=/root/.cache/uv \
    uv sync --frozen --no-install-project \
        --no-install-package pyqt5 \
        --no-install-package pyqtwebengine \
        --no-install-package pyqt5-qt5 \
        --no-install-package pyqt5-sip \
        --no-install-package pyqtwebengine-qt5 \
        --no-install-package pyqt6 \
        --no-install-package pyqt6-qt6 \
        --no-install-package pyqt6-sip \
        --no-install-package pyqt6-webengine \
        --no-install-package pyqt6-webengine-qt6 \
        --no-install-package qtpy \
        --no-install-package pygobject \
        --no-install-package pygobject-stubs \
        --no-install-package pycairo \
        --no-install-package dask

# Install the project itself. The source is bind-mounted at runtime, so this
# layer only provides the console-script entry point and metadata.
COPY . .
RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --no-deps .

# Put the uv-managed venv on PATH so the `aTrain` console script resolves
# without an explicit `uv run`.
ENV PATH="/workspace/.venv/bin:$PATH" \
    VIRTUAL_ENV=/workspace/.venv

# NiceGUI runs on 8080 by default
EXPOSE 8080

# Run in browser mode (no pywebview native window), with hot-reload.
# The source directory is bind-mounted by docker-compose, so edits on the
# host are reflected immediately inside the container; --reload restarts the
# NiceGUI server on file change so contributors don't need a manual restart.
CMD ["aTrain", "start", "--no-native", "--reload"]