Docker & Containerization
I containerize all my AI/ML workflows for reproducibility. If it works in my container, it works everywhere.
What I Containerize
ML Training Environments
PyTorch + CUDA + all dependencies locked. No more 'it worked on my machine' for training runs.
vLLM Inference Servers
Containerized model serving with GPU passthrough. Easy to deploy and scale.
ComfyUI Workflows
Full ComfyUI setup with custom nodes and models. Artists can spin up identical environments.
Development Environments
Consistent dev setups for Python/Node projects. New team members productive in minutes.
GPU Container Setup
I work extensively with NVIDIA GPU containers for ML workloads:
My Standard Practices
Multi-stage builds
Keep runtime images small. Build deps stay in build stage.
Layer caching
Requirements first, code last. Rebuilds are fast.
Non-root users
Security by default, especially for production.
Health checks
Containers that can tell you when they're unhealthy.
Technology Stack
Expertise by Sumit Chatterjee
Industrial Light & Magic, Sydney