Local / workflow
One machine, two audiences.
How the workstation is wired, what it actually serves on the LAN, and the path a request takes from a browser or an agent to a file in an output directory.
01 / Wiring
What is inside the box.
A 2014-era X99 board with two 22-core Xeons, two modified 22 GB cards and one NVLink bridge. Everything below is read off the workstation, not off a spec sheet.
Read it asTwo sockets, one card per socket, one bridge between the cards. The 50 GB/s link is real; the NUMA boundary above it is the part that costs context switches, not bandwidth.
02 / On the LAN
What the network can reach.
The workstation is a generation server for the flat network, over Wi-Fi rather than the two wired NICs. Six services exist; two are running right now.
Read it asTwo audiences on one host: a human with a browser, and a program with an HTTP client. The API exists because an agent cannot drag nodes in ComfyUI.
| Service | Binding | Audience | State | What it does |
|---|---|---|---|---|
| ComfyUI | LAN | Human | Running | ComfyUI 0.45 — H3 image and video graphs, node editor, queue |
| Agent API | LAN | Program / agent | Running | Agent API — submit jobs, poll status, list and download outputs |
| production ComfyUI | LAN | Internal | Stopped | Production ComfyUI 0.36 for the film pipeline; not meant to be public |
| photo drop | LAN | Phone / tablet | On demand | Streaming file drop — uploads from iOS/Android land in a local folder |
| vLLM | loopback | Local tools | On demand | vLLM — the 27B text model, OpenAI-compatible |
| DSH GUI | loopback | This desk only | Running | DeepSeek Harness web GUI — the agent that edits this site |
NoteThe image/video services are started by hand right now. Unit files exist in h3-setup/systemd/ (h3-comfy, h3-agent) but are not installed, so a reboot leaves them down.
03 / The path
From request to file.
Six steps, and only two of them involve a language model. The rest is queueing, memory planning and copying bytes.
Start
ComfyUI and the agent API come up; both load the H3 nodes and the TP2 shard hooks.
HumanDescribe
Browser drags a graph, or an agent reads /capabilities to learn kinds, size presets and model names.
Submit
POST /generate with kind image or video. The queue is serial: one job owns both cards at a time.
Plan memory
32B text encoder (27 GB) is sharded across both cards, released, then the DiT (12 GB) takes the same space.
ModelSample
Image: 6 steps at cfg 1.0. Video: up to 73 frames, 8 steps, 4-step turbo LoRA. Output lands in ~/ComfyUI-img/output.
Collect
Poll /jobs/<id> for output_url, then GET /files/<name>. No upload, no cloud round trip.
04 / Interface
What the API will admit.
The capability document is the contract: an agent should read it before guessing a size or a step count.
GET / human entry point
GET /openapi machine-readable endpoint list
GET /capabilities models, size presets, limits
GET /health service and VRAM state
POST /generate {"kind":"image"|"video","prompt":"...", ...}
GET /jobs recent jobs
GET /jobs/<id> one job, with output_url
GET /files produced files
GET /files/<name> download a file
| Kind | Model | Defaults |
|---|---|---|
| Image | minimax_h3_image_pruned_w4a8 | 6 steps · cfg 1.0 · 1k–4k presets up to 6272 × 2688 |
| Video | minimax_h3_ref2va_pruned_w4a8_mixed | 8 steps · 1344 × 768 · 73 frames max · 4-step turbo LoRA |
| Audio | minimax_music3_dit_fp16 | 30 steps · cfg 1.7 · caption + lyrics with [verse]/[chorus] |
Serial by designThe queue is one job wide. Two clients submitting at once do not share the cards; the second waits.
05 / Exposure
What this is not.
It is a lab service on a home network, not a hosted product. Three consequences worth stating plainly.
No authentication
The UI and the API bind the LAN interface with no token. Anyone on the Wi-Fi can queue a job and read the output folder. That is the trade for zero-friction access from a phone.
LAN only
No port forwarding, no tunnel and no public DNS. The router's NAT is the boundary; the workstation's address is a DHCP lease.
One job at a time
The queue is serial and the cards are the bottleneck. This is a workstation that answers when asked, not a service with a queue you can burst.
06 / Files
Where things land.
Everything stays on the workstation: references in, renders out, uploads beside them.
~/ComfyUI-img/input/ reference images, dropped in by hand or over the API
~/ComfyUI-img/output/ renders — currently ~709 MB of images, video and audio
~/received/ phone uploads via the drop page
/data large datasets and model work (separate ext4 partition)
/mnt/hdd cold archive (3 TB NTFS)