ComfyUI Engine and Queue

Engine & queue

Local Comfy installation is ready, but no instance is running. No remote update check was performed. Captured September 25, 2026.

Engine & queue is the ComfyUI runtime page: install status, start options, GPU pool, live queue, and custom nodes. Models stay unloaded until a workflow needs them. Starting the engine is not "loading SDXL."

Open ComfyUI → Engine & queue. The HTML id is comfy; the visible title is ComfyUI / Engine & queue depending on the nav label. The sidebar says Engine & queue.

Jump buttons: GPU & API guide, Workflows, Comfy models, Refresh.

Installation card

This block reports whether ComfyUI is on disk, the current git commit, how many workflow-template packages are present, and whether an update check has run.

• Check for updates queries the remote. It does not install.
• Update available shows the target commit. Core update is a separate, explicit operation from Manager update_all.
• Not checked means you only have local status. That is not the same as "up to date."

Do not update core in the middle of a long video graph. Stop the instance, update, confirm the commit, then start again.

Start ComfyUI

Do this for a first instance:

1. Device — pick the primary GPU (cuda:N - marketing name). This card becomes logical cuda:0 inside the Comfy process.
2. Leave Memory mode on Normal VRAM unless a documented graph requires low/none.
3. Leave Precision on auto precision.
4. If you have two or more NVIDIA GPUs, leave Enable component placement checked. The primary stays checked and disabled. Check every auxiliary GPU the workflow may use. Unchecked cards are invisible to placement for this instance's lifetime.
5. Click Start ComfyUI.

Then: the instance appears in Running instances with status starting, then ready. A first GPU start can take a few minutes while WSL initializes the shared CUDA driver. The yellow note is explicit: no model weights are being loaded during that wait.

Stop all / per-instance Stop tear the process down. Logs show that instance's output.

Opening the port link (when ready) opens native ComfyUI. Use it to edit graphs visually, then export API JSON back into Workflows. OmniBridge custom nodes appear in the Comfy node menu under OmniBridge.

GPU pool

The pool is primary-first. Placement cannot see a GPU you omitted. If Workflows later reports restart-required, you must stop this instance and start another with the extra UUID/device checked. There is no "add GPU" button on a live process.

Never assume the second card is visible because Runtime listed it. Check this instance's pool.

Memory modes

Mode	When to use
Normal VRAM	Default. Planner will not assume CPU offload will rescue an oversized assignment.
Force Normal VRAM	When Comfy would otherwise pick a more aggressive mode
High VRAM	Keep models loaded; needs headroom
GPU Only	Avoid CPU fallback
Low VRAM	Offload-heavy graphs; planner may accept a small GPU deficit if the cgroup CPU budget fits
No VRAM (CPU offload)	Last resort for huge graphs; slow
CPU Only	No CUDA; debugging

HuMo-class 17B graphs have been shown to OOM a 24 GB primary on low VRAM alone; those need documented none or low + large reserve_vram flags. Set those under advanced startup flags, then analyze-before-run. If the plan is valid: false, do not queue.

Advanced startup flags

Open the disclosure. Quick picks:

• Preview method — Automatic, None, Latent2RGB, TAESD
• Featured flags from the live catalog (attention, cache policy, dynamic VRAM) when the gateway advertises them
• Category groups from GET /api/comfy/start-options — only open what you need
• Troubleshooting → Disable pinned memory — useful on low-RAM or incompatible hosts (--disable-pinned-memory)

A badge counts how many flags differ from defaults. Do not invent raw CLI flags when a structured option exists.

reserve_vram is a Comfy-wide flag. It reduces usable memory on auxiliary GPUs too. Omni's per-device placement reserve is a different number, set on Workflows.

Execution (the queue)

Once an instance is ready, ComfyUI Execution shows:

• instance picker
• Refresh
• Clear Pending — drop queued-but-not-started prompts
• Cancel All — interrupt running work
• Unload Models — Comfy /free with unload; VRAM should drop if nothing is mid-graph

The panel lists running vs pending counts and recent jobs. A successful Workflows Run shows up here as a job with a prompt_id. Watch until finished, then open Media library. HTTP 200 on queue is not a PNG.

Interrupt/cancel does not always stop a CUDA kernel instantly. If a job is stuck after cancel, Stop the instance.

CLI equivalents: omni-cli comfy watch INSTANCE, comfy interrupt, comfy free --unload-models.

Custom nodes

The Custom Nodes card lists installed/disabled nodes. OmniBridge is auto-installed and exposes OmniChat, OmniDescribe, OmniTranscribe, OmniTTS that call back into this gateway. They still require a ready Omni worker for the matching modality, and they inherit the same feature compatibility rules (Qwen will not transcribe via a 501 path).

Install additional nodes from:

• this card's Manager/catalog actions when shown
• CLI omni-cli comfy nodes install <repo_url>
• Engine extensions APIs (/api/comfy/extensions/manage)

Enable/disable/update/remove are distinct. Core update ≠ Manager update_all. After a node install that loads only at process start, Stop and Start the instance, then Check again on Workflows.

Delete of an asset-managed node is immediate.

Assets (models on disk)

Engine & queue does not replace the ComfyUI models view. Use Model library's ComfyUI mode for template installs and the category filesystem scan. CLI:

omni-cli.bat comfy assets list checkpoints
omni-cli.bat comfy assets install checkpoints org/repo --file model.safetensors
omni-cli.bat comfy assets delete checkpoints exact-filename.safetensors

Deletion reports recoverable: false. Resolve category + relative filename from the scan; do not infer loras/ from a display name that says "LoRA."

Uploads go through the authenticated instance (upload/image, upload/mask) or comfy assets upload. Do not copy files into a guessed C:\...ComfyUI\input path on Windows.

Multiple instances

You can start more than one Comfy if you have the GPUs and RAM. Each has its own port, pool, and queue. Workflows' Run on must point at the instance whose pool contains the GPUs in the placement policy.

Stop unused instances. Two Comfy processes plus ACE-Step plus a 7B chat worker is how WSL runs out of host RAM even when VRAM looks fine.

Related pages

• Workflows
• GPU placement
• Runtime — compact instance list
• Command-line interface