Source files / . · raw · rendered

README.md

# Tersler

A ~€620 local-LLM box (GPUs included) built around two used **NVIDIA Tesla P100 16 GB** cards
(32 GB HBM2). It serves **Qwen3.8 Flash-Next (IQ3_S)** on [Strata](docs/research.md#strata-on-tersler-prepared-2026-11-10)
(qr0n's P100 build): 56–64 tok/s decode with MTP, ~350 tok/s prefill, 262k context, image input, lean_bench 0.983.
The original target was the reference build's ~50–67 tok/s on a 27B Q6_K model with MTP speculative decoding
([reference recipe](docs/reference-video-recipe.md)); the 27B on a P100-tuned llama.cpp fork reached 55–59 tok/s
and stays installed as the alternative.

The name comes from the homepage meme ("ERMAHGERD, TERSLER P-HERNDRERDS!"). The working title was
*junk-rig*; hostname, paths, commands and services were renamed on 2026-11-10 (`install/rename-from-junk-rig.sh`).
`/opt/junk-rig` stays as a symlink to `/opt/tersler` for anything that still has the old path built in.

**Status:** running since 2026-11-06. Since 2026-11-10 it serves Qwen3.8 Flash-Next (IQ3_S) on Strata: lean_bench 0.983,
~45-65 tok/s, images, 262k context. The 27B Q6_K on llama.cpp (55-59 tok/s) stays installed as the alternative
([benchmarks](bench/README.md), [changelog](docs/changelog.md)). To build one yourself, follow the
[rebuild guide](docs/recreate.md). Open items: [TODO.md](TODO.md).

## Hardware

| | |
|---|---|
| Board | ASUS Z10PA-U8 (C612, LGA2011-3); GPUs in PCIE6 + PCIE4 = x16/x16 Gen3 |
| CPU | Xeon E5-2690 v3 |
| RAM | 4 × 16 GB DDR4-2133 RDIMM in A1/B1/C1/C2 at 1866 MT/s; channel D dead (see hardware notes) |
| Storage | 256 GB SATA SSD |
| PSU | Seasonic SSP-650RT 650 W; GPUs capped at 150 W each |
| Cooling | one 120 × 38 mm fan + 3D-printed duct over both cards ([shroud/](shroud/)) |

Details: [docs/bom.md](docs/bom.md) (parts and links),
[docs/hardware-notes.md](docs/hardware-notes.md) (slots, BIOS, power cable,
cooling), [docs/research.md](docs/research.md) (how the parts were chosen).

## Software

OS: a prebuilt headless Ubuntu 24.04 image that boots into SSH, written to the SSD
with `dd` ([image/README.md](image/README.md)). Then, on the rig in `~/tersler`:

```bash
cp install/tersler.local.conf.example install/tersler.local.conf   # set MODEL_URL
sudo ./install/install.sh          # all stages; reboot when told, run it again
sudo ./install/install.sh --list   # stages and their state
sudo systemctl start tersler-llama
```

| Command (after install) | Purpose |
|---|---|
| `tersler-server` | run llama-server with the configured model and flags |
| `tersler-strata-server` | run Strata (Flash-Next) instead; `SERVE_ENGINE=strata` makes it the boot default ([research](docs/research.md#strata-on-tersler-prepared-2026-11-10)) |
| `tersler-bench [build]` | benchmark a build, append to `/opt/tersler/bench/results.csv` |
| `tersler-check-updates` | what changed upstream (fork, drivers, CUDA, model, new projects) |
| `tersler-fan-control --discover` | find the board's PWM fan outputs for GPU-temperature fan control |
| `tersler-gpu-setup` | persistence mode + power limit (runs at boot) |

Everything is pinned in [install/tersler.conf](install/tersler.conf). How to
follow upstream safely: [docs/keeping-up.md](docs/keeping-up.md).

## Layout

```
image/     build-image.sh: dd-able Ubuntu 24.04 server image (cloud-init seed)
install/   install.sh, config, systemd units, watched upstreams
scripts/   runtime tools (installed to /opt/tersler/bin)
bench/     benchmark prompts
shroud/    OpenSCAD duct design + STL + previews
docs/      rebuild guide, research, BOM, hardware notes, reference recipe, keeping up, changelog
bench/data/  benchmark results (tersler-bench, lean_bench, load and multi-session tests)
site/      static website generator (python3 site/build.py)
docs/reference/  ASUS Z10PA-U8 user guide, Dell T5810 spec sheet
```