# ComfyUI stack tunables. Copy to `.env` on irv-ml1 before deploying. # # cp .env.example .env # # edit as needed # docker compose up -d # Image tag. Format is `ubuntu_cuda-`; the floating # equivalent is `ubuntu_cuda-latest`. # # irv-ml1 runtime constraints: # - Host driver 570.x → max runtime CUDA 12.8 (per nvidia-smi). Tags # above 12.8 will fail at runtime on driver/CUDA mismatch even # though the manifest pulls fine. # - Host cuda-toolkit is 12.8 to match native Python venvs (ComfyUI, # SDNext, fluxgym, llama.cpp, etc.) — keep the container in sync. # # See https://hub.docker.com/r/mmartial/comfyui-nvidia-docker/tags. COMFYUI_VERSION=ubuntu24_cuda12.8-20260312 # Host port for the ComfyUI web UI (container listens on 8188) COMFYUI_PORT=8188 # Bind address. 0.0.0.0 exposes on all interfaces including the WG # tunnel IP (10.100.79.3) so homepage / the hub side can reach it. # Use 127.0.0.1 to restrict to local / SSH-forwarded access only. COMFYUI_BIND=0.0.0.0 # UID/GID the container user runs as. Must match the owner of # COMFYUI_BASEDIR so workflows, custom nodes, input, and output are # read/write from both the container and external host tooling. # On irv-ml1: lkraven = 1000:1000. COMFYUI_UID=1000 COMFYUI_GID=1000 # Host path for all user state (models, workflows, custom_nodes, # input, output). Must exist and be chowned to COMFYUI_UID:COMFYUI_GID # before the first `docker compose up`. COMFYUI_BASEDIR=/worktank/comfyui/basedir # Host path for container runtime (ComfyUI source + Python venv + # pip cache, ~5 GB after bootstrap). Disposable — safe to `rm -rf` # on version bumps to force a fresh re-bootstrap. Must be owned by # COMFYUI_UID:COMFYUI_GID. COMFYUI_RUNDIR=/worktank/comfyui/run # ComfyUI-Manager security level: # normal → default; installs custom nodes from vetted sources only. # weak → permits arbitrary pip installs via Manager. Use only if # you trust every workflow author you'll import. # strict → blocks pip installs via Manager entirely. COMFYUI_SECURITY_LEVEL=normal