Installing ComfyUI

Native, not containerised

Plain ComfyUI with SDXL-class models compiles nothing — it is prebuilt PyTorch wheels all the way down, so a native virtual environment is lighter and far easier to debug than a container.

Containers earn their keep when a custom node needs to build CUDA extensions against a specific nvcc. That is a real scenario, just not this one. Keep a container runtime available as the escape hatch and reach for it only when a node actually requires a compiler.

Checkout and dependencies

cd ~/projects
git clone --depth 1 https://github.com/comfyanonymous/ComfyUI.git comfyui
cd comfyui
uv pip install --python ~/projects/comfy-venv/bin/python -r requirements.txt

requirements.txt lists torch, torchvision, and torchaudio unpinned. Because the matched CUDA build from the toolchain page is already installed, those requirements are already satisfied and are left alone. Confirm it afterwards rather than assuming:

~/projects/comfy-venv/bin/python -c "import torch; print(torch.__version__)"
# -> 2.11.0+cu130   (a bare '2.11.0' means the generic PyPI build replaced yours)

Keep the weights out of the repository

Model files are tens to hundreds of gigabytes and must survive reinstalling, upgrading, or deleting ComfyUI. Put them in a sibling directory and point ComfyUI at it with extra_model_paths.yaml in the ComfyUI root:

nui_models:
  base_path: /home/fortega/models/
  is_default: true
  checkpoints: checkpoints/
  diffusion_models: diffusion_models/
  loras: loras/
  vae: vae/
  clip: clip/
  clip_vision: clip_vision/
  controlnet: controlnet/
  upscale_models: upscale_models/
  embeddings: embeddings/
  configs: configs/

is_default: true makes this the tree that new downloads land in. Now rm -rf ~/projects/comfyui costs a git clone, not a week of re-downloading.

Confirm it sees the card

Start it bound to loopback and ask it what hardware it found:

cd ~/projects/comfyui
~/projects/comfy-venv/bin/python main.py --listen 127.0.0.1 --port 8188
curl -s http://127.0.0.1:8188/system_stats | python3 -m json.tool

The devices array should name your GPU with a vram_total that matches it. On this box a clean result looks like ComfyUI 0.31.0, pytorch_version: 2.11.0+cu130, one CUDA device of 16652042240 bytes.

ss -lptnH 'sport = :8188'