tomasz-syn-grzegorza
359afb3a59
Infrastructure configs for GMKtec K11 (Docker, vLLM, LocalAI, ComfyUI, control-plane, gpu-fan agent, Server UI with CLI/file explorer/GPU fan curve). Co-authored-by: Cursor <cursoragent@cursor.com>
148 lines
5.6 KiB
Markdown
148 lines
5.6 KiB
Markdown
# Stan systemu i repo
|
|
|
|
Snapshot na koniec sesji. Weryfikuj na serwerze przed działaniem (`docker ps`, `mountpoint /data`).
|
|
|
|
## Co działa
|
|
|
|
| Komponent | Status | Uwagi |
|
|
|-----------|--------|-------|
|
|
| Docker CE | OK | Root: `/data/docker` |
|
|
| NVIDIA driver | OK | ~595.x, `nvidia-smi` na hoście |
|
|
| GPU w kontenerze | OK | `--gpus all` testowany wcześniej |
|
|
| Mount `/data` | OK | ~1 TB, ext4, fstab UUID |
|
|
| vLLM image | OK | `vllm/vllm-openai:latest` pulled |
|
|
| vLLM container | STOP | `VLLM_MODEL` puste |
|
|
| LocalAI image | OK | `v4.4.3-gpu-nvidia-cuda-13` |
|
|
| LocalAI container | RUN | profil `localai`, port host **8070** |
|
|
| Server UI | stack w repo | `stacks/server-ui/` :8091 — native (systemd) lub Docker; zakładki: **CLI**, Pliki, Stacki, GPU Fan |
|
|
| gpu-fan agent | host systemd | `/opt/gpu-fan` :18090 localhost — **nie w `docker ps`** |
|
|
| ComfyUI | stack w repo | `stacks/comfyui/` — **nie uruchomiony na hoście** (rozdział 07) |
|
|
| Firewall / TLS | — | nie wdrożony (rozdział 07) |
|
|
|
|
## Ścieżki na `/data`
|
|
|
|
```
|
|
/data/
|
|
├── docker/ # obrazy i warstwy Docker
|
|
└── apps/
|
|
├── vllm/huggingface/ # cache HF (AWQ) — bez pobranego modelu
|
|
├── gguf/
|
|
│ ├── qwen3.6-27b/ # pod GGUF z katalogu vLLM
|
|
│ └── gemma-4-12b/
|
|
├── localai/
|
|
│ ├── models/ # modele LocalAI (pobieranie w toku)
|
|
│ ├── backends/
|
|
│ ├── configuration/ # api_keys.json, runtime_settings.json
|
|
│ ├── images/
|
|
│ └── data/ # auth DB, agents (przyszłe LOCALAI_AUTH)
|
|
└── comfyui/ # rozdział 07
|
|
├── storage/
|
|
├── models/
|
|
├── cache/hf-hub/
|
|
├── cache/torch-hub/
|
|
├── input/
|
|
├── output/
|
|
├── custom_nodes/
|
|
└── workflows/
|
|
```
|
|
|
|
## Pliki `.env` na serwerze (nie w git)
|
|
|
|
| Plik | Kluczowe ustawienia |
|
|
|------|---------------------|
|
|
| `/opt/control-plane/.env` | `API_KEY`, `SERVER_UI_*`, `GPU_FAN_*` — **jedyny** klucz panelu; instrukcja: [`04a-api-key.md`](../manual-tutorial/04a-api-key.md) |
|
|
| `stacks/vllm/.env` | `VLLM_MODEL=` (puste), `MAX_MODEL_LEN=131072`, port 8000 |
|
|
| `stacks/localai/.env` | `LOCALAI_PORT=8070`, `LOCALAI_API_KEY` ustawiony (sekret), obraz cuda-13 |
|
|
|
|
## Porty stacków (Server UI)
|
|
|
|
| Stack | Zmienna `.env` | Domyślny | Edycja w UI |
|
|
|-------|----------------|----------|-------------|
|
|
| LocalAI | `LOCALAI_PORT` | 8080 | tak |
|
|
| ComfyUI | `COMFYUI_PORT` | 8188 | tak |
|
|
| vLLM | `VLLM_PORT` | 8000 | tak |
|
|
| NPMPlus | — (host network) | 81 | nie (read-only) |
|
|
|
|
Szczegóły: [`SERVER-UI-PORT-CONFIG.md`](SERVER-UI-PORT-CONFIG.md)
|
|
|
|
## Control plane (instalacja)
|
|
|
|
| Komponent | Runtime | Instalator |
|
|
|-----------|---------|------------|
|
|
| gpu-fan | systemd native only | `install-control-plane.sh` lub `gpu-fan/scripts/install.sh` |
|
|
| Server UI | systemd **lub** Docker | `install-control-plane.sh`, `install.sh`, `install-docker.sh` |
|
|
|
|
Szczegóły: [`SERVER-UI-INSTALL-OPTIONS.md`](SERVER-UI-INSTALL-OPTIONS.md), [`CONTROL-PLANE-ENV.md`](CONTROL-PLANE-ENV.md), [`SERVER-UI-FILE-EXPLORER.md`](SERVER-UI-FILE-EXPLORER.md), tutorial [`08-server-ui-install.md`](../manual-tutorial/08-server-ui-install.md)
|
|
|
|
## Struktura repo (kluczowe pliki)
|
|
|
|
```
|
|
ubuntu-bare-metal/
|
|
├── README.md
|
|
├── scripts/setup-data-disk.sh
|
|
├── coding-agent/ # ten katalog
|
|
├── manual-tutorial/
|
|
│ ├── 00-prerequisites.md
|
|
│ ├── 01-system-update-and-docker.md
|
|
│ ├── 02-nvidia-driver.md
|
|
│ ├── 03-nvidia-container-toolkit.md
|
|
│ ├── 03b-system-tools.md
|
|
│ ├── 04-vllm-stack.md # część A dysk + część B vLLM
|
|
│ └── 05-localai-stack.md
|
|
│ ├── 06-gpu-fan-control.md
|
|
│ └── 07-comfyui-stack.md
|
|
└── stacks/
|
|
├── vllm/
|
|
│ ├── models.catalog.yaml
|
|
│ ├── docker-compose.yml
|
|
│ ├── profiles/
|
|
│ └── scripts/
|
|
├── localai/
|
|
│ ├── docker-compose.yml
|
|
│ └── scripts/
|
|
├── comfyui/
|
|
│ ├── docker-compose.yml
|
|
│ └── scripts/
|
|
└── llamacpp/
|
|
└── README.md
|
|
```
|
|
|
|
## Szybka weryfikacja (komendy)
|
|
|
|
```bash
|
|
# dysk i docker
|
|
mountpoint /data && df -h /data
|
|
docker info | grep "Docker Root Dir"
|
|
|
|
# vLLM
|
|
cd ~/cursor/ubuntu-bare-metal/stacks/vllm
|
|
docker compose --profile vllm ps
|
|
grep VLLM_MODEL .env
|
|
|
|
# LocalAI (port z .env użytkownika: 8070)
|
|
cd ~/cursor/ubuntu-bare-metal/stacks/localai
|
|
docker compose --profile localai ps
|
|
curl -s http://localhost:8070/readyz
|
|
|
|
# GPU
|
|
nvidia-smi
|
|
```
|
|
|
|
## Znane rozjazdy (docs vs runtime)
|
|
|
|
| Temat | Repo / docs | Faktycznie na serwerze |
|
|
|-------|-------------|------------------------|
|
|
| LocalAI port | 8080 w `.env.example`, tutorial 05 | **8070** w `.env` użytkownika |
|
|
| LOCALAI_API_KEY | brak w `docker-compose.yml` | ustawiony w `.env`, **może nie działać** w kontenerze |
|
|
| vLLM model | katalog + skrypty gotowe | model nie pobrany, stack nie startowany |
|
|
|
|
## Git / Gitea
|
|
|
|
- Repo zainicjowane lokalnie w `ubuntu-bare-metal/` (branch `main`)
|
|
- Root `.gitignore` — wyklucza `.env`, `.venv`, `upstream/`
|
|
- Push na Gitea: czeka na URL remote — [GITEA-PREP.md](GITEA-PREP.md)
|
|
|
|
## Server UI — wykres GPU Fan (ostatnie zmiany UI)
|
|
|
|
W `stacks/server-ui/static/index.html`: siatka osi, linijki kreskowe, odstępy etykiet, `aspect-ratio` SVG, wyśrodkowanie w karcie. Tutorial: [06-gpu-fan-control.md](../manual-tutorial/06-gpu-fan-control.md).
|