Separate GPU-accelerated steps to GPU the workload image.

.github/workflows/docker-build.yaml

@@ -15,9 +15,17 @@ jobs:
     - name: Set up Docker Buildx
       uses: docker/setup-buildx-action@v1
 
-    - name: Build
+    - name: Build EasyVideoTrans service
       uses: docker/build-push-action@v2
       with:
         context: .
         push: false
-        tags: hanfa/pytvzhen-web:${{github.event.pull_request.number}}
+        tags: hanfa/easyvideotrans:${{github.event.pull_request.number}}
+
+    - name: Build EasyVideoTrans workloads
+      uses: docker/build-push-action@v2
+      with:
+        context: .
+        file: Dockerfile-gpu-workload
+        push: false
+        tags: hanfa/easyvideotrans-workloads:${{github.event.pull_request.number}}

.github/workflows/docker-release.yaml

@@ -1,4 +1,4 @@
-name: Pytvzhen-web Docker Image Release
+name: EasyVideoTrans Service Docker Image Release
 
 on:
   workflow_run:
@@ -29,4 +29,4 @@ jobs:
       with:
         context: .
         push: true
-        tags: hanfa/pytvzhen-web:latest
+        tags: hanfa/easyvideotrans:latest

.github/workflows/docker-workload-release.yaml

@@ -0,0 +1,32 @@
+name: EasyVideoTrans Workloads Docker Image Release
+
+on:
+  workflow_run:
+    workflows: [ "Pytvzhen-web application test" ]
+    branches: [ "master" ]
+    types:
+      - completed
+
+jobs:
+  build:
+    runs-on: self-hosted
+
+    steps:
+    - name: Check out code
+      uses: actions/checkout@v2
+
+    - name: Set up Docker Buildx
+      uses: docker/setup-buildx-action@v1
+
+    - name: Login to DockerHub
+      uses: docker/login-action@v1
+      with:
+        username: ${{ secrets.DOCKERHUB_USERNAME }}
+        password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+    - name: Build and push
+      uses: docker/build-push-action@v2
+      with:
+        context: .
+        push: true
+        tags: hanfa/easyvideotrans-workloads:latest

.gitignore

@@ -27,3 +27,5 @@ output/
 !celery_results/*
 
 .DS_Store
+
+.pytest_cache

Dockerfile

@@ -1,5 +1,8 @@
 # Use an official NVIDIA runtime with CUDA and Miniconda as a parent image
-FROM pytorch/pytorch:2.1.0-cuda11.8-cudnn8-runtime AS base
+FROM python:3.9-slim AS base
+
+ENV PYTHONDONTWRITEBYTECODE 1
+ENV PYTHONUNBUFFERED 1
 
 # Disable interactive debian
 ENV TZ=America/New_York \
@@ -15,7 +18,7 @@ COPY requirements.txt .
 
 # Install dependencies
 RUN pip install --upgrade pip
-RUN pip install -r requirements.txt
+RUN pip install --default-timeout=200 -r requirements.txt
 
 
 FROM base AS final
@@ -27,11 +30,11 @@ COPY . /app
 COPY configs/supervisord.conf /etc/supervisor/conf.d/supervisord.conf
 
 # Set environment variables to configure Celery
-ENV CELERY_BROKER_DOMAIN=localhost
-ENV CELERY_BROKER_URL=pyamqp://guest@localhost:5672//
-ENV CELERY_RESULT_BACKEND=file:///app/celery_results
-ENV CELERY_WORKER_PREFETCH_MULTIPLIER=1
-ENV CELERY_TASK_ACKS_LATE=true
+ENV CELERY_BROKER_DOMAIN localhost
+ENV CELERY_BROKER_URL pyamqp://guest@localhost:5672//
+ENV CELERY_RESULT_BACKEND file:///app/celery_results
+ENV CELERY_WORKER_PREFETCH_MULTIPLIER 1
+ENV CELERY_TASK_ACKS_LATE true
 
 # Make port 8080 available to the world outside this container
 EXPOSE 8080
@@ -42,7 +45,7 @@ ENV FLASK_APP app.py
 ENV FLASK_DEBUG 0
 
 ARG PYTVZHEN_STAGE=beta
-ENV PYTVZHEN_STAGE=${PYTVZHEN_STAGE}
+ENV PYTVZHEN_STAGE ${PYTVZHEN_STAGE}
 
 # Run supervisord to start both Flask and Celery
 CMD ["/usr/bin/supervisord"]

Dockerfile-gpu-workload

@@ -0,0 +1,30 @@
+FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime AS base
+
+ENV PYTHONDONTWRITEBYTECODE 1
+ENV PYTHONUNBUFFERED 1
+
+# Disable interactive debian
+ENV TZ=America/New_York \
+    DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y \
+    ffmpeg \
+    git \
+    && rm -rf /var/lib/apt/lists/*
+
+
+COPY workloads/requirements.txt /app/
+
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY workloads /app/workloads/
+COPY src /app/src/
+COPY inference.py /app
+
+ENV LD_LIBRARY_PATH /opt/conda/lib/python3.11/site-packages/nvidia/cudnn/lib
+
+EXPOSE 8188
+
+CMD ["python", "inference.py"]

app.py

@@ -4,11 +4,10 @@
 import zipfile
 import shutil
 import uuid
-from src.service.audio_processing.audio_remove import audio_remove
-from src.service.audio_processing.transcribe_audio import transcribe_audio_en
-from src.service.audio_processing.voice_connect import connect_voice
-from src.service.translation import get_translator, srt_sentense_merge
+from src.service.video_synthesis.voice_connect import connect_voice
+from src.service.translation import get_translator
 from src.service.tts import get_tts_client
+from src.workload_client import EasyVideoTransWorkloadClient
 from src.task_manager.celery_tasks.tasks import video_preview_task
 from src.task_manager.celery_tasks.celery_utils import get_queue_length
 from werkzeug.utils import secure_filename
@@ -19,16 +18,23 @@
 from prometheus_flask_exporter import PrometheusMetrics
 
 app = Flask(__name__, template_folder="./appendix/templates", static_folder="./appendix/static")
-app.config.from_file("./configs/pytvzhen.json", load=json.load)
+app.config.from_file("./configs/easyvideotrans.json", load=json.load)
 metrics = PrometheusMetrics(app)
 metrics.info('pytvzhen_web', 'Pytvzhen backend API', version='1.0.0')
+
 PYTVZHEN_STAGE = 'PYTVZHEN_STAGE'
 pytvzhen_api_request_counter = metrics.counter(
     'pytvzhen_api_request_counter', 'Request count by request paths',
     labels={'base_url': lambda: url_rule_to_base(request.url_rule), 'stage': lambda: pytvzhen_stage(),
             'method': lambda: request.method, 'status': lambda r: r.status_code}
 )
 
+# Setup workloads client to submit any GPU workloads to EasyVideoTrans compute backend
+gpu_workload = EasyVideoTransWorkloadClient(
+    audio_separation_endpoint=app.config['VOICE_BACKGROUND_SEPARATION_ENDPOINT'],
+    audio_transcribe_endpoint=app.config['AUDIO_TRANSCRIBE_ENDPOINT'],
+)
+
 
 def pytvzhen_stage():
     return os.environ[PYTVZHEN_STAGE] if PYTVZHEN_STAGE in os.environ else 'default'
@@ -283,9 +289,7 @@ def remove_audio_bg(video_id):
             f'not found at {output_path}, please extract it first')}), 404
 
     try:
-        baseline_path = app.config['REMOVE_BACKGROUND_MUSIC_BASELINE_MODEL_PATH']
-        audio_remove(audio_path, audio_no_bg_path, audio_bg_fn_path, baseline_path,
-                     app.config['REMOVE_BACKGROUND_MUSIC_TORCH_DEVICE'])
+        audio_bg_fn_path, audio_no_bg_fn = gpu_workload.separate_audio(audio_fn)
         return jsonify({"message": log_info_return_str(
             f"Remove remove background music for {audio_fn} as {audio_no_bg_fn} and {audio_bg_fn_path} successfully."),
             "video_id": video_id}), 200
@@ -333,7 +337,6 @@ def audio_bg_serve(video_id):
 def transcribe(video_id):
     output_path = app.config['OUTPUT_PATH']
 
-    transcribe_model = "medium"
     en_srt_fn, en_srt_merged_fn, audio_no_bg_fn = f'{video_id}_en.srt', f'{video_id}_en_merged.srt', f'{video_id}_no_bg.wav'
 
     en_srt_path, en_srt_merged_path, audio_no_bg_path = (os.path.join(output_path, en_srt_fn),
@@ -351,10 +354,7 @@ def transcribe(video_id):
             f'not found at {audio_no_bg_path}, please extract it first')}), 404
 
     try:
-        transcribe_audio_en(app.logger, path=audio_no_bg_path, modelName=transcribe_model, language="en",
-                            srtFilePathAndName=en_srt_path)
-        srt_sentense_merge(app.logger, en_srt_path, en_srt_merged_path)
-
+        gpu_workload.transcribe_audio(audio_no_bg_fn, [en_srt_fn, en_srt_merged_fn])
         return jsonify({"message": log_info_return_str(
             f"Transcribed SRT from {audio_no_bg_fn} as {en_srt_fn} and {en_srt_merged_fn} successfully."),
             "video_id": video_id}), 200

configs/easyvideotrans.json

@@ -0,0 +1,6 @@
+{
+  "OUTPUT_PATH": "./output",
+  "VIDEO_MAX_DURATION": 3610,
+  "VOICE_BACKGROUND_SEPARATION_ENDPOINT": "http://localhost:8199/audio-sep",
+  "AUDIO_TRANSCRIBE_ENDPOINT": "http://localhost:8199/audio-transcribe"
+}