ComfyUI CUDA 13 Docker

ComfyUI CUDA 13 Docker is the reproducible GPU runtime underneath my local AI media workflow. It runs separate ComfyUI workers for an RTX 4090 and RTX 3090, keeps models and generated media outside disposable containers, and provides an optional sharded worker for workflows that need both cards.

This page records the current system boundary. The repository is the canonical place for exact versions, startup commands, performance flags, smoke tests, and rollback notes. The related writing below explains how this runtime participates in the larger generation-management workflow.

At a glance

• Separate workers and user state for each physical GPU.
• Pinned PyTorch CUDA 13 wheels and audited ComfyUI extensions.
• Persistent model, input, output, user, and temporary directories.
• Per-GPU performance settings instead of one unsafe shared configuration.
• Health checks plus CUDA, API, and parallel-generation smoke tests.
• An opt-in sharded profile for unusually large workflows.

Current technical shape

persistent models + inputs
            |
     Docker Compose
       /         \
      v           v
ComfyUI GPU 0  ComfyUI GPU 1
 RTX 4090       RTX 3090
      \           /
       v         v
 persistent outputs

optional: one two-GPU sharded worker

Each default worker sees only its assigned device. Container images hold the runtime and pinned dependencies; host-mounted directories hold the state that must survive rebuilds.

Project boundary

This is local GPU infrastructure, so there is no public demo. GitHub is the useful public artifact: it contains the reproducible environment and the checks required to know that both workers can actually execute workloads.