Skip to content

Add Blackwell Cuda GB10 Support #154

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

Open
snmabaur wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Add Blackwell Cuda GB10 Support #154

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
199 changes: 199 additions & 0 deletions backend/Dockerfile.blackwell
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,199 @@
# =============================================================================
# OpenTranscribe GPU Worker for NVIDIA DGX Spark / ARM64 / Blackwell
# - basiert auf NVIDIA PyTorch Container
# - ohne appuser, stattdessen USER 1000:1000
# - mit Blackwell/NVRTC Patches fuer WhisperX / Torchaudio
# - mit pyannote SemVer Workaround fuer NVIDIA Torch Dev-Versionen
# =============================================================================

FROM nvcr.io/nvidia/pytorch:25.01-py3

WORKDIR /app

# Blackwell / NVRTC Kompatibilitaet + Cache-Pfade
ENV TORCH_CUDA_ARCH_LIST="9.0" \
CUDA_FORCE_PTX_JIT=1 \
PYTHONUNBUFFERED=1 \
PYTHONDONTWRITEBYTECODE=1 \
HF_HOME=/home/user/.cache/huggingface \
TRANSFORMERS_CACHE=/home/user/.cache/huggingface/transformers \
TORCH_HOME=/home/user/.cache/torch \
NLTK_DATA=/home/user/.cache/nltk_data \
SENTENCE_TRANSFORMERS_HOME=/home/user/.cache/sentence-transformers \
XDG_CACHE_HOME=/home/user/.cache \
PATH=/usr/local/bin:$PATH

# System-Abhaengigkeiten
RUN apt-get update && apt-get install -y --no-install-recommends \
curl \
ffmpeg \
libsndfile1 \
libimage-exiftool-perl \
libgomp1 \
git \
cmake \
build-essential \
libopenblas-dev \
sox \
libsox-dev \
&& rm -rf /var/lib/apt/lists/*

# Verzeichnisse fuer Runtime und Caches
RUN mkdir -p \
/app \
/app/models \
/app/temp \
/home/user/.cache/huggingface \
/home/user/.cache/torch \
/home/user/.cache/nltk_data \
/home/user/.cache/sentence-transformers \
/home/user/.cache/yt-dlp \
&& chown -R 1000:1000 /app /home/user

# Requirements zuerst fuer Layer-Cache
COPY requirements.txt /app/requirements.txt

# Basis Python-Werkzeuge
RUN python -m pip install --no-cache-dir --upgrade pip setuptools wheel && \
python -m pip install --no-cache-dir pybind11 packaging ninja cmake

# -----------------------------------------------------------------------------
# Torchaudio gegen NVIDIA Torch bauen
# -----------------------------------------------------------------------------
RUN git clone --depth 1 --branch v2.6.0 --recursive https://github.com/pytorch/audio.git /tmp/torchaudio && \
cd /tmp/torchaudio && \
pip install --no-cache-dir --no-deps --no-build-isolation . && \
rm -rf /tmp/torchaudio

# -----------------------------------------------------------------------------
# CTranslate2 mit CUDA/cuDNN bauen
# -----------------------------------------------------------------------------
RUN git clone --recursive --depth 1 --branch v4.4.0 https://github.com/OpenNMT/CTranslate2.git /tmp/ctranslate2 && \
cd /tmp/ctranslate2 && \
mkdir build && cd build && \
cmake .. \
-DCMAKE_BUILD_TYPE=Release \
-DWITH_CUDA=ON \
-DWITH_CUDNN=ON \
-DCUDNN_ROOT=/usr \
-DCUDA_DYNAMIC_LOADING=ON \
-DWITH_MKL=OFF \
-DWITH_OPENBLAS=ON \
-DOPENMP_RUNTIME=COMP \
-DCMAKE_INSTALL_PREFIX=/usr/local && \
make -j"$(nproc)" && \
make install && \
ldconfig && \
cd /tmp/ctranslate2/python && \
pip install --no-cache-dir --no-build-isolation . && \
rm -rf /tmp/ctranslate2

# -----------------------------------------------------------------------------
# NVIDIA Torch-Stack sichern, bevor pip evtl. Dinge ueberschreibt
# -----------------------------------------------------------------------------
RUN cp -r /usr/local/lib/python3.12/dist-packages/torch /tmp/torch_nvidia && \
cp -r /usr/local/lib/python3.12/dist-packages/torchvision /tmp/torchvision_nvidia && \
cp -r /usr/local/lib/python3.12/dist-packages/torchaudio /tmp/torchaudio_custom && \
cp -r /usr/local/lib/python3.12/dist-packages/torio /tmp/torio_custom && \
cp -r /usr/local/lib/python3.12/dist-packages/numpy /tmp/numpy_nvidia && \
cp -r /usr/local/lib/python3.12/dist-packages/numpy.libs /tmp/numpy_libs_nvidia 2>/dev/null || true

# -----------------------------------------------------------------------------
# OpenTranscribe-Requirements ohne GPU-kritische Pakete installieren
# So bleibt der NVIDIA-Torch-Stack erhalten
# -----------------------------------------------------------------------------
RUN grep -vE '^(torch==|torch>=|torchaudio==|torchaudio>=|ctranslate2|whisperx==|whisperx>=|pyannote\.audio)' /app/requirements.txt > /tmp/requirements.safe.txt && \
pip install --no-cache-dir -r /tmp/requirements.safe.txt

# GPU-/WhisperX-relevante Pakete explizit setzen
RUN pip install --no-cache-dir \
huggingface_hub==0.23.5 \
whisperx==3.3.1 \
faster-whisper==1.1.0 \
pyannote.audio==3.3.2 \
python-multipart \
nltk \
matplotlib

# -----------------------------------------------------------------------------
# NVIDIA Torch-Stack wiederherstellen
# -----------------------------------------------------------------------------
RUN rm -rf /usr/local/lib/python3.12/dist-packages/torch && \
rm -rf /usr/local/lib/python3.12/dist-packages/torchvision && \
rm -rf /usr/local/lib/python3.12/dist-packages/torchaudio && \
rm -rf /usr/local/lib/python3.12/dist-packages/torio && \
rm -rf /usr/local/lib/python3.12/dist-packages/numpy && \
rm -rf /usr/local/lib/python3.12/dist-packages/numpy.libs && \
mv /tmp/torch_nvidia /usr/local/lib/python3.12/dist-packages/torch && \
mv /tmp/torchvision_nvidia /usr/local/lib/python3.12/dist-packages/torchvision && \
mv /tmp/torchaudio_custom /usr/local/lib/python3.12/dist-packages/torchaudio && \
mv /tmp/torio_custom /usr/local/lib/python3.12/dist-packages/torio && \
mv /tmp/numpy_nvidia /usr/local/lib/python3.12/dist-packages/numpy && \
mv /tmp/numpy_libs_nvidia /usr/local/lib/python3.12/dist-packages/numpy.libs 2>/dev/null || true

# =========================
# BLACKWELL PATCHES
# =========================

# Patch 1: get_device_capability fuer SM_121 auf SM_90 umbiegen
RUN sed -i 's/def get_device_capability/def _original_get_device_capability/g' \
/usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py && \
printf '\n# BLACKWELL PATCH: Spoof SM_121 as SM_90 for nvrtc compatibility\n' >> /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py && \
printf 'def get_device_capability(device=None):\n' >> /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py && \
printf ' major, minor = _original_get_device_capability(device)\n' >> /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py && \
printf ' if major == 12 and minor == 1:\n' >> /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py && \
printf ' return (9, 0)\n' >> /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py && \
printf ' return (major, minor)\n' >> /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py

# Patch 2: harte compute_121 / sm_121 Referenzen ersetzen
RUN sed -i 's/compute_121/compute_90/g' /usr/local/lib/python3.12/dist-packages/torch/utils/cpp_extension.py 2>/dev/null || true && \
sed -i 's/sm_121/sm_90/g' /usr/local/lib/python3.12/dist-packages/torch/utils/cpp_extension.py 2>/dev/null || true

# Patch 3: optional Inductor-Codecache patchen
RUN sed -i 's/compute_121/compute_90/g' /usr/local/lib/python3.12/dist-packages/torch/_inductor/codecache.py 2>/dev/null || true && \
sed -i 's/sm_121/sm_90/g' /usr/local/lib/python3.12/dist-packages/torch/_inductor/codecache.py 2>/dev/null || true

# Patch 4: Torchaudio fbank / jiterator Crash umgehen
RUN sed -i 's/spectrum = torch.fft.rfft(strided_input).abs()/# BLACKWELL PATCH: Avoid jiterator by computing abs manually\n fft_result = torch.fft.rfft(strided_input)\n spectrum = torch.sqrt(fft_result.real**2 + fft_result.imag**2)/' \
/usr/local/lib/python3.12/dist-packages/torchaudio/compliance/kaldi.py

# Patch 5: pyannote SemVer Check deaktivieren fuer NVIDIA Torch Dev-Versionen
RUN python - <<'PY'
from pathlib import Path
import re

p = Path("/usr/local/lib/python3.12/dist-packages/pyannote/audio/utils/version.py")
if not p.exists():
raise SystemExit("pyannote version.py not found")

text = p.read_text()

patched = re.sub(
r"def check_version\(.*?\n(?=def |\Z)",
"def check_version(*args, **kwargs):\n return\n\n",
text,
flags=re.S,
)

if patched == text:
print("No check_version patch applied; pattern not found exactly, leaving file unchanged")
else:
p.write_text(patched)
print("Disabled pyannote check_version")
PY

# App-Code kopieren
COPY . /app

# Besitzrechte fuer Runtime-User
RUN chown -R 1000:1000 /app /home/user

USER 1000:1000

EXPOSE 8080

HEALTHCHECK --interval=30s --timeout=10s --start-period=60s --retries=3 \
CMD curl -f http://localhost:8080/health || exit 1

# Default fuer Backend-Container
CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8080"]
Loading