Docker & Build¶
The ComfyUI Studio Docker image packages ComfyUI, 38 custom nodes, Python with PyTorch/CUDA, optional attention optimizations, and an optional llama-server binary into a single image ready for GPU cloud providers like RunPod.
The image is configurable via build arguments to target any NVIDIA GPU from Volta (V100) through Blackwell (B200).
Pages¶
| Page | Description |
|---|---|
| Build Configurator | Interactive shell script that walks you through GPU selection, LLM support, and node categories, then outputs a ready-to-run docker build command or a customized Dockerfile. |
| Build Arguments | Reference for all 7 --build-arg parameters: CUDA version, PyTorch index, Python version, SageAttention, FlashAttention, LLM support, and llama.cpp version. Includes build examples. |
| Custom Nodes | The 38 custom nodes bundled in the image, organized into 6 categories. How nodes.txt and install_nodes.sh work, how to add or remove nodes, and runtime installation via the UI. |
| llama-server | How the llama.cpp server binary is compiled with CUDA inside the Docker image. Build flags, CUDA architecture targets, the --allow-shlib-undefined linker workaround, and version pinning. |
| GPU Compatibility | Full compatibility table for 33 GPUs across 7 generations. Lists VRAM, compute capability, recommended CUDA version, PyTorch index, and supported SageAttention/FlashAttention versions. |
| CI/CD | GitHub Actions workflow that builds and pushes the image to GHCR. Trigger rules, registry caching strategy, image tagging, and configurable build variables. |
Quick Reference¶
Default build (Blackwell/Hopper, all features):
Guided build (interactive configurator):
Image registry:
Key Files¶
| File | Purpose |
|---|---|
docker/production/Dockerfile |
Multi-stage build, all build args, layer caching strategy |
docker/configure.sh |
Interactive build configurator script |
docker/production/nodes.txt |
Custom node list with section markers |
docker/production/install_nodes.sh |
Node installer (reads nodes.txt between section markers) |
docker/production/start.sh |
Container entrypoint (bootstrap + process supervision) |
docker/production/bootstrap.py |
First-boot setup (clone repo, copy to working dirs) |
.github/workflows/build.yml |
CI/CD workflow for automated builds |