Passer au contenu principal
B

dsh-zcode-farm

bowluna/dsh-zcode-farm

Multi-instance ComfyUI orchestration: probes every configured ComfyUI endpoint for its GPU, free VRAM and queue depth, then dispatches a generation job to the idlest eligible instance. Ships three tools (comfyui_farm_status, comfyui_farm_pick, comfyui_farm_run). An instance is eligible only if it is reachable, has at least the requested free VRAM, and its queue is no deeper than the configured ceiling; eligible instances are ranked by free VRAM minus queue depth times a weight, so an idle shallow queue wins over a busy large one. Zero runtime dependencies. Optional retry on probe failure, because a single timeout across an SSH tunnel is not proof that an instance is down.

Installer

dsh plugin --profile web add github:bowluna/dsh-zcode-farm

README

dsh-zcode-farm

English | 中文

test license dsh node

Let the DeepSeek Harness agent see a whole farm of ComfyUI instances and dispatch each job to the idlest one.

dsh plugin --profile web add dsh-zcode-farm

How the farm picks an instance

dsh-comfyui (77★)this plugin
Endpoints it can addressonemany, probed concurrently
Queue depthread from /queue, but with one instance it informs the user rather than a choiceweighted into the ranking, across instances
When the job landson that one instanceon the instance with the most free VRAM and an empty queue
Writes to ComfyUIyesyes (comfyui_farm_run), but reading is enough

Zero runtime dependencies · Node ≥ 22 · MIT


Why this exists

ComfyUI's own docs are explicit: one ComfyUI process executes one workflow at a time. Real concurrency only comes from running one process per GPU and routing each job to the least-busy instance. The ecosystem already has pooling libraries such as comfyui-orchestrator.

But when you wire that up to an AI agent, one piece is missing: the agent can't see the farm.

Measured on a real machine (a single moment on 2026-09-21, taken with this plugin's own probe):

InstanceGPUFree VRAMQueueVerdict
18100 ← the only one dsh-comfyui can seeRTX 50900.8 / 33.7 GB1 runningtoo full
18300RTX 50900.8 / 33.7 GB1 runningtoo full
18301RTX 50900.9 / 33.7 GB1 runningtoo full
18302RTX 50900.5 / 33.7 GB1 runningtoo full
18303RTX PRO 6000 Blackwell99.4 / 102.0 GB0✅ idle

dsh-comfyui (77★, the most mature plugin in this space) talks to one endpoint, so every generation request hits the busiest machine — while 102 GB of VRAM sits idle.

This plugin fills exactly that gap: let the agent see, then let it dispatch.


What it does

Three tools:

ToolPurpose
comfyui_farm_statusFull farm view: per-instance GPU, free VRAM, queue depth, what's running, and who should get the next job
comfyui_farm_pickChoose without executing: filter by VRAM floor / queue ceiling / GPU-name hint / exclude list, and get the reason
comfyui_farm_runLoad-aware dispatch: auto-pick (or force an instance), accepting a raw API workflow or a one-shot "prompt + checkpoint" text-to-image

Install

dsh plugin --profile web add dsh-zcode-farm

Restart the app and the agent gets all three tools.


Configuration

- id: comfyui-farm
  config:
    primaryUrl: 'http://127.0.0.1:8188'
    instances:
      - { id: gpu-18301, baseUrl: 'http://127.0.0.1:18301', label: '5090 shard' }
      - { id: gpu-18303, baseUrl: 'http://127.0.0.1:18303', label: 'Blackwell 102G' }
    needVramGb: 8
    maxQueueDepth: 3
    queueWeight: 1000
    probeTimeoutMs: 5000
    probeRetries: 1

How an instance is chosen (two stages)

  1. Filter — unreachable, or free VRAM < needVramGb, or queue depth > maxQueueDepth → out.
  2. Rankscore = freeVramGb − queueDepth × queueWeight

The default queueWeight = 1000 is deliberate: queue depth dominates the ranking and free VRAM only breaks ties. That matches the intuition — an idle small GPU finishes sooner than a busy large one.

Prefer "biggest VRAM wins"? Set queueWeight: 0.


Relationship to dsh-comfyui: complementary, not competing

dsh-comfyui (77★)dsh-zcode-farmHow to check the dsh-comfyui side
Scopeone endpointa whole farmgrep -n baseUrl package/lib/config.js → one z.string()
Strengthworkflow library, skill packs, canvas, asset panelawareness, selection, dispatchls package/lib
How a job is routedwherever the single endpoint isranked by free VRAM − depth × weightgrep -n "constructor(baseUrl" package/lib/comfyui.js
Assembly row idcomfyuicomfyui-farmboth appear in this repo's cordis.patch.yml

The row ids differ, so both can be installed side by side. Neither depends on the other. dsh-comfyui's single-endpoint design is verified from its published source, not asserted — see "Reproducing the comparison" below for the three commands.


Relationship to ZCode

This repository is self-developed — there is no ZCode code lineage. ZCode (zai-org/ZCode, zai-org/GLM-skills) is credited here as a source of inspiration for the engineering conventions this project follows, not as a dependency: nothing in the install path, the runtime, or the acceptance checks requires it, and no third-party vendor key is involved anywhere in this repository.

Most cells below are therefore deliberately blank rather than filled with something invented.

What ZCode hasWhat this plugin tookWhat this plugin adds beyond itEvidence (run it yourself)
Nothing for ComfyUI farm scheduling——the whole probe → rank → dispatch loopgrep -rli comfyui <zcode-checkout> | wc -l0
A memory-extraction sub-agent with two skip conditions (apps/zcode-cli/packages/core/src/memory/extraction.ts:23)—— (different domain)——n/a — nothing was taken, so there is no claim to reproduce
Bilingual paired documentsadopted as the house conventionextended to cover figure files (*.svg), so a diagram can't drift from its translationnode tools/verify-translation-pairing.mjs
Guard discipline — runnable checks before every commitadopted as the house conventionextended to six guards, the sixth being a real-boot checknode test/run.mjs and node tools/boot-check.mjs --port 32041

A row reading —— is an honest blank, not an oversight.


Reproducing the comparison

Nothing above needs to be taken on trust. These are the commands behind it:

git clone https://github.com/BOWLUNA/dsh-zcode-farm && cd dsh-zcode-farm
node test/run.mjs                        # 23 checks in 2 suites
node tools/ordering-replay.mjs           # the ranking table behind the figure
node tools/boot-check.mjs --port 32041   # really installs and really boots (needs a harness)

The dsh-comfyui claims are checked against its published package, not against this repo:

npm view dsh-comfyui version                # 0.5.1
npm pack dsh-comfyui@0.5.1 && tar xzf dsh-comfyui-0.5.1.tgz
grep -n baseUrl package/lib/config.js       # a single string, not a list
grep -n "constructor(baseUrl" package/lib/comfyui.js
9:    baseUrl: z.string().default('http://127.0.0.1:8188'),
71:    constructor(baseUrl, apiKey, connectTimeoutMs, maxMediaBytes) {

Run on 2026-09-22 against dsh-comfyui@0.5.1.


Design notes

  • Zero runtime dependencies. Node's built-in fetch / AbortController only.
  • No workflow editing. That's dsh-comfyui's territory; this plugin only answers "who gets it".
  • No assumption of locality. Every member is just a URL — SSH tunnels, LAN and remote hosts are treated identically. The 5 tunnels on the test machine are plain ssh -L forwards.
  • Probe failures are retried. Tunnel flakiness is real: the same instance once timed out at >3s and answered in 106ms on the very next call. A single failure is not a verdict — we retry, then report attempts so the agent can tell "down" from "hiccup".
  • Failure reasons are never overwritten. When filtering by GPU hint, an unreachable instance keeps its "unreachable" reason instead of being relabelled "GPU mismatch", which would hide the actual problem.

Development

npm test                              # unit tests (pure functions, no network)
node probes/probe-farm.mjs            # standalone probe: farm status only
node probes/probe-tools.mjs           # load the entry and really execute the tools (read-only)
node probes/probe-tools.mjs --run     # actually dispatch one job (smallest payload: 512² / 12 steps)

During development the host-provided peer deps (@deepseek-ai/schemastery and its deps) must be present in a local node_modules/, otherwise the entry cannot be imported. This is not needed when installed into a real profile — the host supplies them.


Known limitations

  • The built-in template of comfyui_farm_run covers minimal text-to-image only (KSampler + CheckpointLoaderSimple + EmptyLatentImage + 2× CLIPTextEncode + VAEDecode + SaveImage). For anything richer, pass a workflow.
  • Aliases pointing at the same backend are not detected. If you configure both a switchable forwarder (e.g. 127.0.0.1:8188 → "currently focused instance") and its real target, you will see two members with identical readings. That is honest reporting, not a bug — probing cannot distinguish them.
  • number returned by /prompt is a server-side cumulative counter, not a queue position (measured: it returns 22 even when the queue is empty). Trust comfyui_farm_status for real queue depth.

License

MIT


Status

  • 2 suites, 23 checks — run node test/run.mjs
  • Declared compatibility: >=0.1.5-rc.2 <0.2.0-0 (see engines.dsh and the peer range)
  • Pin the version to bypass pnpm's release cooldown: dsh plugin --profile web add dsh-zcode-farm@1.0.1
  • Verified against five SSH-tunnelled ComfyUI instances plus one switchable forwarder

Plugins associés