Part 10 of a series — a retrospective appendix. The narrative ran across Parts 1–9; this one lays the evolution out stage by stage.
No part of the setup was chosen for its own sake. Every change — a new machine, a different coordinator, a smaller model — was forced by a specific failure in the stage before it. Below is each stage as a snapshot, with the arrow between snapshots showing what broke and what was added or removed to fix it.
The timeline, stage by stage
Stage 0 — Initial
| Machine | What it runs |
|---|---|
| Main — 5950X · ~80 GB · RX 6900XT | Coordinator and all workers — everything on gemma4:26b, local |
⬇ Broke:
gemma4:26bcan't hold a coordination loop beyond ~1–2 delegations.
Added: a search for a dedicated coordinator model (workers stay on gemma).
Stage 1 — Coordinator search (still one machine)
| Machine | What it runs |
|---|---|
| Main — ~80 GB · RX 6900XT | Coordinator: trialing gpt-oss:120b (and groq-tool-use:8b, which fabricated sessions) · Workers: gemma4:26b
|
⬇ Broke: the 120B (~77 GB) and the gemma workers (~20 GB) can't both be resident in 80 GB RAM → ~100 GB swapped per hand-off; 291 s to say hello.
Removed: the workers from the main machine · Added: a second machine (5600G + GTX 1060).
Stage 2 — Distributed (two machines)
| Machine | What it runs |
|---|---|
| Main — ~80 GB · RX 6900XT | Coordinator: gpt-oss:120b (resident; runs on CPU + RAM) |
| Worker (.114) — 32 GB · GTX 1060 | Workers: gemma4:26b
|
⬇ Noticed: with the 120B running on CPU and RAM, the main machine's GPU sits idle.
Added: image generation (ComfyUI + SDXL) on the main 6900XT.
Stage 3 — Image generation added
| Machine | What it runs |
|---|---|
| Main — ~80 GB · RX 6900XT | Coordinator gpt-oss:120b (~77 GB RAM) + SDXL image gen (~15 GB VRAM) |
| Worker (.114) |
gemma4:26b workers |
⬇ Broke: the 120B leaves only ~3 GB RAM free → SDXL generations time out.
Removed:gpt-oss:120b· Added:gpt-oss:20b(~14 GB) in its place.
Stage 4 — Right-sized coordinator
| Machine | What it runs |
|---|---|
| Main — ~80 GB · RX 6900XT | Coordinator gpt-oss:20b (~14 GB RAM) + SDXL (~15 GB VRAM) — ~60 GB RAM free
|
| Worker (.114) |
gemma4:26b workers |
⬇ Added over the same period: a cloud strategist (Sage, on Gemini 3 Flash) for the one role needing deep reasoning; the laptop as an SD 1.5 image node; and deterministic orchestration (a script owns the delegation loop, the model only plans and evaluates).
Stage 5 — Final
| Where | What it runs |
|---|---|
| Main — ~80 GB · RX 6900XT | Coordinator gpt-oss:20b + SDXL image generation |
| Worker (.114) — 32 GB · GTX 1060 |
gemma4:26b specialist workers |
| Laptop (.247) — 32 GB DDR5 · 780M | SD 1.5 image node (512 px) |
| Cloud | Sage strategist — Gemini 3 Flash (free tier), consulted when blocked |
The laptop's detour
The 8840U laptop ran its own small arc inside the larger one.
| Attempt | Result |
|---|---|
| Laptop 8840U / 780M as an LLM worker | CPU prefill ~25 tok/s → 13-minute stalls |
⬇ Why: no discrete GPU, so prefill is hopeless for large agent prompts (Part 2).
Repurposed for the one job the hardware could do:
| Repurposed | Result |
|---|---|
| Laptop as an image node | SDXL hangs on the 780M; SD 1.5 @ 512 px works (~37–42 s) |
Initial versus final
| Initial setup | Final setup | |
|---|---|---|
| Machines | 1 (main) | 3 (main, worker, laptop) + 1 cloud model |
| Coordinator |
gemma4:26b, local |
gpt-oss:20b, local |
| Workers |
gemma4:26b, same machine as the coordinator |
gemma4:26b, dedicated worker machine (.114) |
| Strategist | — | Gemini 3 Flash, cloud (consulted only when blocked) |
| Image generation | — | SDXL on the main 6900XT; SD 1.5 on the laptop node (.247) |
| Orchestration | model-driven loop | deterministic script owns the loop; model only plans and evaluates |
The pattern
Read top to bottom, the chain has a shape. Nearly every change was one of two moves: placement — which machine a thing runs on — or right-sizing — how large each model needs to be for its job. The workers moved for placement; the coordinator shrank for right-sizing; the strategist went to the cloud because that was the right placement for the one role that needed deep reasoning. Not once did the answer turn out to be "buy a bigger model and hope." The capability was almost always already there in the hardware on hand; the work was in arranging it so each piece had room to do the thing it was good at.
Top comments (0)