Skip to main content

Docker Profiles

FOVEA uses Docker Compose profiles to support both CPU and GPU deployments from a single configuration file.

Introduction​

Docker Compose profiles enable explicit deployment choices without maintaining separate configuration files:

  • Default profile: CPU mode (works on any system)
  • GPU profile: Requires NVIDIA GPU and nvidia-docker2
  • Build modes: minimal (fast builds) vs full (all inference engines)

CPU vs GPU Comparison​

Performance Comparison​

OperationCPU MinimalCPU FullGPU MinimalGPU Full
Video Summarization (30s clip)45s40s8s5s
Object Detection (1 frame)1.2s1.0s0.2s0.15s
Object Tracking (100 frames)25s20s5s3s
Cold Start (model loading)10s15s20s30s
Memory Usage2GB3GB6GB10GB

These benchmarks use a sample video with 1920x1080 resolution at 30 FPS. Actual performance varies based on hardware, video complexity, and model configuration.

Build Modes​

Minimal Build​

The minimal build mode (BUILD_MODE=minimal) includes base dependencies only:

  • Faster build times (5 minutes vs 15 minutes)
  • Smaller image size (2GB vs 8GB)
  • Suitable for CI/CD pipelines and local development
  • Includes basic inference capabilities

Full Build​

The full build mode (BUILD_MODE=full) includes all inference engines:

  • SGLang inference engine (optimized for VLMs and LLMs)
  • vLLM fallback inference engine
  • All object detection models (YOLO variants)
  • All tracking models (SAMURAI, SAM2Long, ByteTrack)
  • Longer build times but better performance

When to Use Each Mode​

CPU + Minimal​

Use for:

  • CI/CD pipelines (fast builds, low resource usage)
  • Local development without GPU
  • Testing core functionality without AI features
  • Systems where model inference is not required

CPU + Full​

Use for:

  • Production CPU deployments (cloud servers without GPUs)
  • Testing inference engines before GPU deployment
  • Development with full AI features on CPU-only machines

GPU + Minimal​

Use for:

  • Quick GPU testing
  • Prototyping with minimal dependencies
  • Resource-constrained GPU environments

GPU + Full (Recommended for Production)​

Use for:

  • Production deployments with NVIDIA GPUs
  • Maximum AI inference performance
  • Complete feature set with all models available

Commands​

Start CPU Mode (Default)​

docker compose up

Starts all services in CPU mode. The model service uses CPU inference.

Start GPU Mode​

docker compose --profile gpu up

Starts services with GPU support. Requires NVIDIA GPU and nvidia-docker2 installed.

Build with Specific Mode​

# Minimal build (fast, smaller image)
docker compose build --build-arg BUILD_MODE=minimal

# Full build (slow, larger image, all engines)
docker compose build --build-arg BUILD_MODE=full

Verify GPU Access​

After starting GPU mode, verify GPU is accessible:

docker compose exec model-service nvidia-smi

This command should display NVIDIA GPU information. If it fails, check NVIDIA driver and nvidia-docker2 installation.

Next Steps​