Files
karaoke-gen/backend/Dockerfile.gpu-base
Andrew Beveridge ☄️andClaude Opus 5.5 acb02e138a fix: un-pin audio-separator to 0.47.1 with RoFormer overlap=2 (v0.278.16) (#1196)
@coderabbitai ignore

## Why
BACKLOG "audio-separator 0.45–0.47 regression: mel-band RoFormer karaoke
models 3–5× slower on L4". The root cause is **python-audio-separator
#299 (0.47.0)**, not #298. #299 made RoFormer `overlap` a *window count*
that defaults to each model YAML's `num_overlap`. Before that, `overlap`
was effectively an 8-second step, which meant almost no overlap. The
karaoke models' YAMLs set `num_overlap` 4/8, so a 3-minute track went
from 23 forward passes to 87 (aufr33) and 124 (gabox_v2, becruily). That
matches the measured 25→75s, 26→149s and 27→157s. 0.45.0 and 0.46.0
still had the old step formula.

python-audio-separator #313 (released as **0.47.1**) keeps #299. It adds
a per-model forward-pass budget test on every PR and a GPU per-pass
timing test (T4: 1.21 s/pass), plus a README note. Andrew chose
overlap=2 for gen: smoother chunk seams than 0.44's no-overlap, at about
1.5× the 0.44 compute instead of 4–6×.

## What
- `karaoke_gen/audio_processor.py`: `SEPARATOR_MDXC_PARAMS` (overlap=2)
goes to all four `Separator(...)` constructions (stage 1, stage 2
karaoke ensemble, single-model paths) **only when audio-separator ≥
0.47**. On older versions `overlap=2` meant a 2-second step, which is 4×
slower, so they keep their defaults. Ensembles load each model through
`load_model`, so the param reaches every model in the preset.
- `backend/Dockerfile.gpu-base`: `AUDIO_SEPARATOR_VERSION=0.47.1`, with
the comment rewritten to explain the root cause. 0.47.1 needs
`torch>=2.3` on Linux, so the pre-installed `torch==2.6.0+cu126`
(required by the L4 driver) is kept.
- The poetry lock stays on 0.44.1. The version-aware params keep CI and
CPU images on the old semantics, and there's no torch churn.

## Expected cost per ~3-min track (forward passes)
| model | 0.44.2 | 0.47 YAML default | 0.47 + overlap=2 |
|---|---|---|---|
| karaoke_aufr33 | 23 | 87 | 44 |
| karaoke_gabox_v2 | 23 | 124 | 32 |
| karaoke_becruily | 23 | 124 | 32 |
| instrumental fv7z / resurrection | 23 | 32 / 21 | 32 / 21 |

## Verification plan (before /backlog-finish)
After the deploy, time the separation of the daily E2E or a real job on
the audio-separation-job logs. I expect roughly 1.5× the 0.44.2
karaoke-stage times (~40–50s per karaoke model rather than 75–157s). If
it's worse, revert the pin to 0.44.2. That's one line.

## Testing
- `tests/unit/test_audio_processor_local_gpu.py`: version gating
(0.47.x/0.48/1.0 → overlap 2; 0.44/0.46 → defaults; unknown → defaults;
fresh copy), and both stages' constructors receive `mdxc_params.overlap
== 2`.
- `make test`: the backend unit suite passes. The only failures are 8
`test_emulator_integration.py` tests, caused by my local environment:
another session's emulator setup is on :8080/:4443, and they fail
identically on untouched main. CI is green on main.
- Local CodeRabbit: 0 findings.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01Jw21syo4HizJJjYw2mg2Vp

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 15:09:41 -04:00

145 lines
6.5 KiB
Docker

# GPU base image with CUDA, PyTorch, audio-separator, and pre-downloaded models
# This image is rebuilt only when deps or models change (infrequently).
#
# Build with: gcloud builds submit --config=backend/cloudbuild-gpu-base.yaml --project=nomadkaraoke --timeout=3600s
# Or manually: docker build -f backend/Dockerfile.gpu-base -t karaoke-backend-gpu-base .
#
# Layer caching strategy:
# 1. System deps + fonts + ffmpeg (rarely change, always cached)
# 2. CUDA library path configuration
# 3. pyproject.toml + package stubs (changes with dep updates)
# 4. pip install karaoke-gen deps + audio-separator[gpu] (slow, cached unless pyproject.toml changes)
# 5. Source code copy (changes every rebuild, but AFTER deps are installed)
# 6. spaCy model download (cached unless pip install layer changes)
# 7. Pre-downloaded ensemble preset models (~1-1.5 GB, cached unless pip layer changes)
FROM nvidia/cuda:12.6.3-runtime-ubuntu22.04
# Prevent interactive prompts during package installation
ENV DEBIAN_FRONTEND=noninteractive
# Install Python 3.11 from deadsnakes PPA and system dependencies
RUN apt-get update && apt-get install -y --no-install-recommends \
software-properties-common \
&& add-apt-repository -y ppa:deadsnakes/ppa \
&& apt-get update && apt-get install -y --no-install-recommends \
# Python 3.11
python3.11 \
python3.11-dev \
python3.11-venv \
# Audio libraries
libsndfile1 \
libsox-dev \
sox \
# Build tools
build-essential \
# Utilities
curl \
xz-utils \
&& rm -rf /var/lib/apt/lists/* \
&& python3.11 --version
# Install comprehensive font support for international lyrics and symbols
# Noto fonts provide full Unicode coverage including:
# - Latin, Greek, Cyrillic (core)
# - CJK: Chinese, Japanese, Korean (~150MB)
# - Arabic, Hebrew, Thai, Hindi, and other scripts (extra)
# - Color emoji (limited support in libass, but available)
# - Musical symbols (♪ ♫ etc.)
# Note: This adds ~200-300MB to the image but ensures all languages render correctly
RUN apt-get update && apt-get install -y --no-install-recommends \
fonts-noto-core \
fonts-noto-cjk \
fonts-noto-extra \
fonts-noto-color-emoji \
fontconfig \
&& fc-cache -f \
&& rm -rf /var/lib/apt/lists/*
# Install FFmpeg from Ubuntu package
# The audio worker only uses FFmpeg for format conversion and padding, not
# encoding-intensive tasks, so the Debian package is sufficient. The static
# build from johnvansickle.com is used for the encoding worker instead.
RUN apt-get update && apt-get install -y ffmpeg && rm -rf /var/lib/apt/lists/*
# Set Python 3.11 as default and install pip
RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.11 1 \
&& update-alternatives --install /usr/bin/python python /usr/bin/python3.11 1 \
&& curl -sS https://bootstrap.pypa.io/get-pip.py | python3.11 \
&& python3 -m pip install --no-cache-dir --upgrade pip setuptools wheel
# Set up CUDA library paths so GPU-accelerated packages can find CUDA libs at runtime
RUN echo '/usr/local/cuda/lib64' >> /etc/ld.so.conf.d/cuda.conf && ldconfig
# Environment configuration for CUDA and Python
ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH \
PATH=/usr/local/cuda/bin:$PATH \
PYTHONUNBUFFERED=1 \
MODEL_DIR=/models
# Set working directory
WORKDIR /app
# Copy ONLY dependency specification files first (changes less often than source code)
COPY pyproject.toml README.md LICENSE /app/
# Create minimal package stubs so pip can resolve the editable install
# pip install -e . needs package directories to exist but doesn't validate content
RUN mkdir -p /app/karaoke_gen /app/lyrics_transcriber_temp /app/backend && \
touch /app/karaoke_gen/__init__.py /app/lyrics_transcriber_temp/__init__.py /app/backend/__init__.py
# Install all karaoke-gen Python dependencies (the slow part)
# BuildKit cache mount preserves downloaded wheels across builds
RUN --mount=type=cache,target=/root/.cache/pip \
pip install --upgrade pip && \
pip install -e .
# Install audio-separator with GPU support (PyTorch + CUDA + onnxruntime-gpu)
# This is separate from pip install -e . to keep layer caching granular
#
# IMPORTANT: Install PyTorch cu126 FIRST, then audio-separator[gpu].
# The default audio-separator[gpu] pulls PyTorch cu130, but Cloud Run L4 GPUs
# have NVIDIA driver 570 which only supports up to CUDA 12.8. PyTorch cu130
# fails with "NVIDIA driver is too old". PyTorch cu126 works with 12.8 drivers.
# Installing torch first means audio-separator[gpu] sees it as already satisfied.
#
# PINNED to an exact version: an unpinned install silently picked up 0.47.0 on the
# 2026-10-17 rebuild and separation went from ~5 to ~10-16 min. Cause: 0.47 (#299) made
# RoFormer `overlap` a window count defaulting to each model YAML's num_overlap (4-8 for
# the karaoke models = 4-6x the forward passes of 0.44's fixed 8 s step). karaoke_gen's
# audio_processor now passes overlap=2 on >= 0.47 (SEPARATOR_MDXC_PARAMS). Time a real
# job before bumping this, and read the audio-separator release notes for compute changes.
ARG AUDIO_SEPARATOR_VERSION=0.47.1
RUN --mount=type=cache,target=/root/.cache/pip \
pip install torch==2.6.0+cu126 torchvision==0.21.0+cu126 --index-url https://download.pytorch.org/whl/cu126 && \
pip install "audio-separator[gpu]==${AUDIO_SEPARATOR_VERSION}"
# Now copy actual source code (AFTER deps installed, so code changes don't
# invalidate the expensive pip install layer in registry cache)
COPY karaoke_gen /app/karaoke_gen
COPY lyrics_transcriber_temp /app/lyrics_transcriber_temp
COPY backend /app/backend
# Download spaCy language model for lyrics transcription
# Must be done at build time, not runtime (Cloud Run network restrictions)
RUN python -m spacy download en_core_web_sm
# Create models directory and pre-download ensemble preset models
# Baking models into the image eliminates cold-start model download latency.
# Total model size: ~1-1.5 GB for instrumental_clean + karaoke presets.
RUN mkdir -p /models
COPY backend/scripts/download_models.py /tmp/download_models.py
RUN python /tmp/download_models.py /models && \
rm /tmp/download_models.py && \
ls -lh /models/
# Label with build info
ARG BUILD_DATE
ARG BUILD_ID
ARG BASE_CONTENT_HASH
LABEL org.opencontainers.image.created="${BUILD_DATE}"
LABEL org.opencontainers.image.revision="${BUILD_ID}"
LABEL org.opencontainers.image.title="karaoke-backend-gpu-base"
LABEL org.opencontainers.image.description="GPU base image with CUDA, PyTorch, audio-separator, and pre-downloaded models for karaoke-backend audio worker"
LABEL base.content.hash="${BASE_CONTENT_HASH}"