Skip to content

Add torch_audiomentations to parakeet Dockerfile #8085

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
beastoin merged 5 commits into from
Jun 22, 2026
+36 −9
Merged

Add torch_audiomentations to parakeet Dockerfile #8085

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
82ba2c9
Add torch_audiomentations to parakeet Dockerfile
beastoin Jun 21, 2026
366c755
Add julius and torch-pitch-shift to parakeet Dockerfile
beastoin Jun 21, 2026
4557e67
Stub torch_audiomentations instead of installing dep chain
beastoin Jun 21, 2026
00d4d12
Adopt verified Dockerfile from dev testing (v7)
beastoin Jun 21, 2026
12bb90b
Pin pyannote.audio <4.0 and expose torchaudio.functional
beastoin Jun 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
45 changes: 36 additions & 9 deletions backend/parakeet/Dockerfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,13 +17,36 @@ WORKDIR /app
RUN pip install --no-cache-dir --no-deps --force-reinstall \
"nemo_toolkit[asr] @ git+https://github.com/beastoin/NeMo.git@68d99a17944181452205dc60da21e82bf3647054"

# NGC's custom torch ABI is incompatible with ALL standard torchaudio
# wheels (PyPI, cu128 index). NeMo imports torchaudio at module level
# (via squim metrics) but parakeet never calls those functions.
# Create a lightweight stub that satisfies `import torchaudio`.
RUN mkdir -p /usr/local/lib/python3.12/dist-packages/torchaudio && \
printf '__version__ = "stub"\n' > \
/usr/local/lib/python3.12/dist-packages/torchaudio/__init__.py
# NGC torch ABI is incompatible with standard torchaudio C extensions.
# Install torchaudio --no-deps for the pure-Python compliance.kaldi module
# (wespeaker needs kaldi.fbank for mel filterbank features), then patch
# __init__.py to skip the C extension loader and only expose compliance.
RUN pip install --no-cache-dir --no-deps torchaudio && \
printf '__version__ = "2.11.0-ngc-compat"\nfrom . import compliance\nfrom . import functional\n' > \
/usr/local/lib/python3.12/dist-packages/torchaudio/__init__.py && \
printf '_IS_TORCHAUDIO_EXT_AVAILABLE = False\ndef fail_if_no_align(*a, **kw): pass\ndef fail_if_no_sox(*a, **kw): pass\ndef fail_if_no_ffmpeg(*a, **kw): pass\ndef fail_if_no_soundfile(*a, **kw): pass\ndef fail_if_no_kaldi(*a, **kw): pass\n' > \
/usr/local/lib/python3.12/dist-packages/torchaudio/_extension/__init__.py

# pyannote.audio.core.task imports torch_audiomentations for training-time
# data augmentation. We only use Model + Inference (embedding extraction),
# never the training pipeline. Stub the package with all symbols pyannote needs.
RUN mkdir -p /usr/local/lib/python3.12/dist-packages/torch_audiomentations/core \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/augmentations \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/utils && \
printf '__version__ = "stub"\nclass Identity:\n pass\n' > \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/__init__.py && \
printf '' > \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/core/__init__.py && \
printf 'class BaseWaveformTransform:\n pass\n' > \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/core/transforms_interface.py && \
printf '' > \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/augmentations/__init__.py && \
printf 'class Mix:\n pass\n' > \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/augmentations/mix.py && \
printf '' > \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/utils/__init__.py && \
printf 'def from_dict(*a, **kw):\n pass\n' > \
/usr/local/lib/python3.12/dist-packages/torch_audiomentations/utils/config.py

# Extra deps not in the NGC image.
# pyannote.audio and torch-dependent deps installed --no-deps to prevent
Expand All @@ -37,11 +60,15 @@ RUN pip install --no-cache-dir \
"prometheus-client>=0.21.0" \
"soundfile>=0.13.0"

RUN pip install --no-cache-dir --no-deps "pyannote.audio>=3.1.0" && \
# pyannote.audio + deps. Post-install: stub telemetry (needs opentelemetry
# OTLP exporter which we don't need for inference-only usage).
RUN pip install --no-cache-dir --no-deps "pyannote.audio>=3.1.0,<4.0" && \
pip install --no-cache-dir --no-deps \
"pyannote.core" "pyannote.database" "pyannote.pipeline" \
"speechbrain" "asteroid-filterbanks" "einops" "semver" \
"hf_transfer" "tensorboardX"
"hf_transfer" "tensorboardX" && \
printf 'def set_opentelemetry_log_level(*a, **kw): pass\ndef set_telemetry_metrics(*a, **kw): pass\ndef track_model_init(*a, **kw): pass\ndef track_pipeline_init(*a, **kw): pass\ndef track_pipeline_apply(*a, **kw): pass\n' > \
/usr/local/lib/python3.12/dist-packages/pyannote/audio/telemetry/__init__.py

COPY backend/parakeet/ .

Expand Down
Loading