Spaces:

Cavefishlabs
/

reflections

Running on CPU Upgrade

Reflections service: same code as the CPU fallback (f4aac7d), always-on CPU

5278313 verified about 21 hours ago

3.45 kB

# cuDNN 9 in the base image. faster-whisper (CTranslate2) and onnxruntime-gpu both need it, and the
# plain runtime image has none: Whisper on the GPU then fails to load, or everything quietly runs on
# the CPU while the GPU is billed.
FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04

# compute,utility is the default; video adds the GPU's H.264 encoder (NVENC) for the proxy and the
# re-encode of masters OpenCV cannot read (app.py nvenc_available, which falls back to libx264).
ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility,video
# Model caches inside the image, written at build time by prefetch_models.py and read at run time.
# HOME too, because HSEmotion, EasyOCR, DeepFace and funasr cache under ~ whatever they are told.
ENV HOME=/opt/home \
    HF_HOME=/opt/home/.cache/huggingface \
    DEEPFACE_HOME=/opt/home \
    EASYOCR_MODULE_PATH=/opt/home/.EasyOCR

# Install Python and system dependencies
RUN apt-get update && apt-get install -y \
    python3.10 \
    python3-pip \
    python3.10-venv \
    ffmpeg \
    curl \
    libsndfile1 \
    libgl1 \
    libglib2.0-0 \
    libegl1 \
    libgles2 \
    && rm -rf /var/lib/apt/lists/* \
    && ln -s /usr/bin/python3.10 /usr/bin/python

# Set working directory
WORKDIR /app

# Copy requirements first for caching
COPY requirements.txt .
COPY requirements-stage-read.txt .

# Install Python dependencies
RUN pip install --no-cache-dir --upgrade pip && \
    pip install --no-cache-dir -r requirements.txt

# Stage Read extras (docs/stage-read-spec.md). Non-fatal: if a package fails to install the
# service still starts and only the /stage-read routes are unavailable.
RUN pip install --no-cache-dir -r requirements-stage-read.txt || echo "stage-read extras failed to install"

# HSEmotion and the packages above pull the CPU build of onnxruntime. Swap it for the GPU build, after them all (same module, adds the
# CUDA provider; the same version the CPU build resolves to, built for CUDA 12 and cuDNN 9).
RUN pip uninstall -y onnxruntime && pip install --no-cache-dir onnxruntime-gpu==1.23.2

# Copy application code. cei_measures.py is the Creative Effectiveness Index pass; leaving it out
# of the image is silent, because app.py imports it inside a try and the report then says
# "no measurements" rather than failing.
COPY app.py cei_measures.py ffmpeg_util.py memorability.py memorability-model.json service_auth.py prefetch_models.py .
COPY stage_read ./stage_read

# Every model's weights, into the image (prefetch_models.py). Never fails the build: a model that
# does not download here downloads on first use instead. Readable and writable by the user the Space
# runs as, which is not the user that built it.
RUN mkdir -p $HOME && python prefetch_models.py && chmod -R a+rwX $HOME

# Stage Read model files (Apache-2.0), fetched at build time rather than committed (the Space rejects binaries outside LFS)
RUN mkdir -p stage_read/models && \
    curl -fsSL -o stage_read/models/face_landmarker.task \
      https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/1/face_landmarker.task && \
    curl -fsSL -o stage_read/models/face_detection_yunet_2023mar.onnx \
      https://github.com/opencv/opencv_zoo/raw/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx || \
    echo "stage-read model download failed"

# Expose port
EXPOSE 7860

# Run the FastAPI app
CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]

	# cuDNN 9 in the base image. faster-whisper (CTranslate2) and onnxruntime-gpu both need it, and the
	# plain runtime image has none: Whisper on the GPU then fails to load, or everything quietly runs on
	# the CPU while the GPU is billed.
	FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04

	# compute,utility is the default; video adds the GPU's H.264 encoder (NVENC) for the proxy and the
	# re-encode of masters OpenCV cannot read (app.py nvenc_available, which falls back to libx264).
	ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility,video
	# Model caches inside the image, written at build time by prefetch_models.py and read at run time.
	# HOME too, because HSEmotion, EasyOCR, DeepFace and funasr cache under ~ whatever they are told.
	ENV HOME=/opt/home \
	HF_HOME=/opt/home/.cache/huggingface \
	DEEPFACE_HOME=/opt/home \
	EASYOCR_MODULE_PATH=/opt/home/.EasyOCR

	# Install Python and system dependencies
	RUN apt-get update && apt-get install -y \
	python3.10 \
	python3-pip \
	python3.10-venv \
	ffmpeg \
	curl \
	libsndfile1 \
	libgl1 \
	libglib2.0-0 \
	libegl1 \
	libgles2 \
	&& rm -rf /var/lib/apt/lists/* \
	&& ln -s /usr/bin/python3.10 /usr/bin/python

	# Set working directory
	WORKDIR /app

	# Copy requirements first for caching
	COPY requirements.txt .
	COPY requirements-stage-read.txt .

	# Install Python dependencies
	RUN pip install --no-cache-dir --upgrade pip && \
	pip install --no-cache-dir -r requirements.txt

	# Stage Read extras (docs/stage-read-spec.md). Non-fatal: if a package fails to install the
	# service still starts and only the /stage-read routes are unavailable.
	RUN pip install --no-cache-dir -r requirements-stage-read.txt \|\| echo "stage-read extras failed to install"

	# HSEmotion and the packages above pull the CPU build of onnxruntime. Swap it for the GPU build, after them all (same module, adds the
	# CUDA provider; the same version the CPU build resolves to, built for CUDA 12 and cuDNN 9).
	RUN pip uninstall -y onnxruntime && pip install --no-cache-dir onnxruntime-gpu==1.23.2

	# Copy application code. cei_measures.py is the Creative Effectiveness Index pass; leaving it out
	# of the image is silent, because app.py imports it inside a try and the report then says
	# "no measurements" rather than failing.
	COPY app.py cei_measures.py ffmpeg_util.py memorability.py memorability-model.json service_auth.py prefetch_models.py .
	COPY stage_read ./stage_read

	# Every model's weights, into the image (prefetch_models.py). Never fails the build: a model that
	# does not download here downloads on first use instead. Readable and writable by the user the Space
	# runs as, which is not the user that built it.
	RUN mkdir -p $HOME && python prefetch_models.py && chmod -R a+rwX $HOME

	# Stage Read model files (Apache-2.0), fetched at build time rather than committed (the Space rejects binaries outside LFS)
	RUN mkdir -p stage_read/models && \
	curl -fsSL -o stage_read/models/face_landmarker.task \
	https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/1/face_landmarker.task && \
	curl -fsSL -o stage_read/models/face_detection_yunet_2023mar.onnx \
	https://github.com/opencv/opencv_zoo/raw/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx \|\| \
	echo "stage-read model download failed"

	# Expose port
	EXPOSE 7860

	# Run the FastAPI app
	CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]