Norwin YuAI usage research
中文 AI usage research

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.

192.168.1.10two services2 × 22 GBSnapshot / Nov 2026

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.

Workstation topology: two Xeon sockets, one 2080 Ti per socket, joined by NVLink, plus NVMe, HDD and a USB Wi-Fi adapter. NUMA 0 · SOCKET 0 Xeon E5-2696 v4 22C / 44T · 2.2–3.7 GHz DDR4 · 47 GB local GPU0 on PCIe 05:00.0 NUMA 1 · SOCKET 1 Xeon E5-2696 v4 22C / 44T · 2.2–3.7 GHz DDR4 · 47 GB local GPU1 on PCIe 83:00.0 RTX 2080 Ti 22,528 MiB · TU102 · sm_75 180 W cap · 616 GB/s RTX 2080 Ti 22,528 MiB · TU102 · sm_75 180 W cap · 616 GB/s NVLink NV2 · 2 links × 25.78 GB/s different NUMA nodes — the bridge does not move memory between them STORAGE · NETWORK NVMe · 1 TB WD Green SN350 · / 317 GB · /data 323 GB HDD · 3 TB WD30EFRX · /mnt/hdd 886 GB USB Wi-Fi · 480 Mbps RTL8852CU · LAN address · wired NICs down

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.

88threads · 2 sockets
94GiB DDR4
44GB VRAM · 2 cards
180W cap per card

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.

Local network: the workstation at 192.168.1.10 serves ComfyUI, an agent API and a photo drop to a MacBook, a phone and other agents on the same Wi-Fi. Router 192.168.1.1 · DHCP Wi-Fi · USB 2.0 · 480 Mbps WORKSTATION · 192.168.1.10 Ubuntu 24.04.5 · 2 × RTX 2080 Ti 22 GB · 44 GB VRAM · 94 GB DDR4 ComfyUI · browser browser · running Agent API · HTTP HTTP · running production UI film pipeline · stopped photo drop phone upload · on demand vLLM · local 27B text · on demand DSH GUI · loopback local only MacBook Pro Safari · UI · uploads iPhone / iPad Safari · photo upload Another agent HTTP · /capabilities → /generate

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.

Services exposed on the workstation
ServiceBindingAudienceStateWhat it does
ComfyUILANHumanRunningComfyUI 0.45 — H3 image and video graphs, node editor, queue
Agent APILANProgram / agentRunningAgent API — submit jobs, poll status, list and download outputs
production ComfyUILANInternalStoppedProduction ComfyUI 0.36 for the film pipeline; not meant to be public
photo dropLANPhone / tabletOn demandStreaming file drop — uploads from iOS/Android land in a local folder
vLLMloopbackLocal toolsOn demandvLLM — the 27B text model, OpenAI-compatible
DSH GUIloopbackThis desk onlyRunningDeepSeek 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.

01

Start

ComfyUI and the agent API come up; both load the H3 nodes and the TP2 shard hooks.

Human
02

Describe

Browser drags a graph, or an agent reads /capabilities to learn kinds, size presets and model names.

Both
03

Submit

POST /generate with kind image or video. The queue is serial: one job owns both cards at a time.

Program
04

Plan memory

32B text encoder (27 GB) is sharded across both cards, released, then the DiT (12 GB) takes the same space.

Model
05

Sample

Image: 6 steps at cfg 1.0. Video: up to 73 frames, 8 steps, 4-step turbo LoRA. Output lands in ~/ComfyUI-img/output.

Model
06

Collect

Poll /jobs/<id> for output_url, then GET /files/<name>. No upload, no cloud round trip.

Program

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
Models and defaults reported by the capability endpoint
KindModelDefaults
Imageminimax_h3_image_pruned_w4a86 steps · cfg 1.0 · 1k–4k presets up to 6272 × 2688
Videominimax_h3_ref2va_pruned_w4a8_mixed8 steps · 1344 × 768 · 73 frames max · 4-step turbo LoRA
Audiominimax_music3_dit_fp1630 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.

01

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.

02

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.

03

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)