Refactor LLM configuration to remove llamacpp support

- Removed references to llamacpp from configuration files, scripts, and documentation, streamlining the LLM setup process to focus solely on Ollama. - Updated environment variables and paths to eliminate llamacpp-related entries, ensuring clarity in the configuration. - Adjusted CLI commands and help messages to reflect the removal of llamacpp, enhancing user experience and reducing confusion. - Revised documentation to provide clear guidance on using Ollama exclusively, including updates to setup instructions and runtime options.
2026-08-21 08:51:36 +03:00
parent 9a4b87dc06
commit 2ab32a8ab5
45 changed files with 139 additions and 1521 deletions
@@ -4,7 +4,6 @@
 !.env.example
 gpu-rent.vars
 ollama-models.yaml
-llamacpp-models.yaml
 models.yaml
 extensions.yaml

@@ -55,8 +55,8 @@ Unix: `./gpu-rent.sh …` (один раз `chmod +x gpu-rent.sh`).
 | `doctor` | Preflight без create |
 | `flavors` / `dry-run` | Что выберет / план без денег |
 | `up` / `up --yes` | GPU + seed + туннель; без `--yes` — меню Selectel/LLM |
-| `up --ollama` / `--llamacpp` | + LLM (см. [docs/llm.md](docs/llm.md)) |
-| `up --llm-only --llamacpp` | только LLM, без SwarmUI |
+| `up --ollama` | + Ollama (см. [docs/llm.md](docs/llm.md)) |
+| `up --llm-only --ollama` | только Ollama, без SwarmUI |
 | `tunnel` / `open` | Снова UI / браузер |
 | `hold` / `status` | Пауза killer / состояние |
 | `stop` / `destroy --i-understand-data-loss` | Стоп GPU / + диски |
@@ -73,7 +73,7 @@ Unix: `./gpu-rent.sh …` (один раз `chmod +x gpu-rent.sh`).
 | `.env` | из `env.example` — `OS_*`, Civitai (не в git) |
 | `gpu-rent.vars` | из example — несекретные дефолты, лаунчер |
 | `models.yaml` / `extensions.yaml` | манифесты (из `*.example.yaml`) |
-| `ollama-models.yaml` / `llamacpp-models.yaml` | LLM (из example; gitignore) |
+| `ollama-models.yaml` | Ollama-модели (из example; gitignore) |
 | `Models/` … | локальный push на `up` |
 | `.gpu-rent/` | state, SSH-ключ (gitignore) |

@@ -87,7 +87,7 @@ Unix: `./gpu-rent.sh …` (один раз `chmod +x gpu-rent.sh`).
 | --- | --- |
 | [docs/setup.md](docs/setup.md) | **Пошаговая подготовка** до первого `up` |
 | [docs/cli.md](docs/cli.md) | Все команды и переменные |
-| [docs/llm.md](docs/llm.md) | Ollama / llama.cpp |
+| [docs/llm.md](docs/llm.md) | Ollama (opt-in LLM) |
 | [docs/models.md](docs/models.md) | Civitai + папка `Models/` |
 | [docs/spike-notes.md](docs/spike-notes.md) | Чеклист первого живого прогона |
 | [docs/README.md](docs/README.md) | Оглавление всего `docs/` |
@@ -20,7 +20,7 @@ CLI поднимает прерываемый GPU в Selectel, держит Swar
 | Git-расширения SwarmUI/Comfy | [extensions.md](extensions.md) |
 | Word-list промптов | [autocomplete.md](autocomplete.md) |
 | Push/pull папок | [local-folders.md](local-folders.md) |
-| Ollama / llama.cpp | [llm.md](llm.md) |
+| Ollama | [llm.md](llm.md) |
 | Capture ссылок с VM | [cli.md](cli.md) (`capture`) + [models.md](models.md) |
 | Как устроены диски и killer | [architecture.md](architecture.md) |
 | Контракт Selectel | [selectel.md](selectel.md) |
@@ -19,7 +19,7 @@
 │    SwarmUI : 127.0.0.1:7801                                   │
 │    idle-killer: очередь / hold / LLM busy → delete this server │
 │    app cred: DELETE/GET только этот server_id (fail closed)   │
-│    optional: Ollama :11434 / llama.cpp :8080 (туннель 17811/12)│
+│    optional: Ollama :11434 (туннель 17811)                    │
 │                                                               │
 │  boot volume (network)     ОС + NVIDIA + SwarmUI нативно + snapshot │
 │  data volume (network)     Models, Output, Data, workflows     │
@@ -49,7 +49,7 @@
 | `notify` | Toast/звук при backend Idle |
 | `tunnel` | sshtunnel + Nova EXPIRED watchdog |
 | `local_watchdog` | Опциональный локальный тик → stop при unclean exit |
-| `llm_runtime` / `setup_wizard` | Opt-in Ollama/llama.cpp + `ollama-models.yaml` / `llamacpp-models.yaml` |
+| `llm_runtime` / `setup_wizard` | Opt-in Ollama + `ollama-models.yaml` |
 | `idle_killer` / `hold` | systemd на VM + hold-файл |
 | `ready` / `snapshot` | Idle backend + boot snapshot |
 | `access_card` | URL / MCP panel после ready |
@@ -51,7 +51,7 @@ gpu-rent up --yes --ollama

 | Команда | Поведение |
 | --- | --- |
-| `gpu-rent setup` | Wizard: манифесты (в т.ч. ollama/llamacpp-models), `LLM_RUNTIME` (нумерованное меню), пресет, опционально local-watchdog |
+| `gpu-rent setup` | Wizard: манифесты (в т.ч. ollama-models), `LLM_RUNTIME` (нумерованное меню), пресет, опционально local-watchdog |
 | `gpu-rent doctor` | Preflight **без** create. Exit ≠ 0 → сессию начинать нельзя |
 | `gpu-rent flavors` | Скан `SCAN_POOLS` × `FLAVOR_PREFERENCE`, список в текущем регионе |
 | `gpu-rent dry-run` | План без mutating-вызовов |
@@ -59,8 +59,8 @@ gpu-rent up --yes --ollama
 | `gpu-rent up --yes` | Без вопросов; flavor из `FLAVOR_PREFERENCE` / `DEFAULT_FLAVOR_ID` |
 | `up --keep-on-fail` | Не гасить GPU при ошибке install (по умолчанию `UP_STOP_ON_FAIL=true` → `stop`) |
 | `gpu-rent up -v` / `--verbose` | Полная таблица doctor на `up` (по умолчанию кратко) |
-| `gpu-rent up --ollama` / `--llamacpp` / `--llm …` | LLM рядом со SwarmUI |
-| `gpu-rent up --no-swarm` / `--llm-only` | Только LLM (нужен `--ollama`/`--llamacpp`); без clone SwarmUI |
+| `gpu-rent up --ollama` / `--llm ollama` | Ollama рядом со SwarmUI |
+| `gpu-rent up --no-swarm` / `--llm-only` | Только Ollama (нужен `--ollama` / `LLM_RUNTIME=ollama`); без clone SwarmUI |
 | `gpu-rent up --no-update` | Без `git pull` SwarmUI/extensions (только недостающие clone) |
 | `gpu-rent up --no-tunnel` | Только облако |
 | `gpu-rent up --no-spot` | Не preemptible |
@@ -138,7 +138,7 @@ Exit 0 → можно `up`. Exit 1 → причина в таблице / кра
 | --- | --- |
 | `.env` | Секреты и OpenStack (`OS_*`, токены). Не в git |
 | `gpu-rent.vars` | Несекретные дефолты; читают лаунчеры и CLI. Пример: `gpu-rent.vars.example` |
-| `models.yaml` / `extensions.yaml` / `ollama-models.yaml` / `llamacpp-models.yaml` | Манифесты (gitignore; из `*.example.yaml`) |
+| `models.yaml` / `extensions.yaml` / `ollama-models.yaml` | Манифесты (gitignore; из `*.example.yaml`) |

 ### Лаунчер (`gpu-rent.vars`)

@@ -192,7 +192,6 @@ AUTOCOMPLETE_ENABLED=true
 SWARMUI_LOCAL_PORT=17801
 LLM_RUNTIME=none
 OLLAMA_LOCAL_PORT=17811
-LLAMACPP_LOCAL_PORT=17812
 UPDATE_GIT=true

 DEFAULT_FLAVOR_ID=
@@ -233,12 +232,12 @@ Application credential для idle-killer CLI создаёт на `up` (узко
 1. Nova `ACTIVE`
 2. TCP 22 / SSH
 3. Backend Idle (если SwarmUI) → toast (если `NOTIFY_READY`)
-4. На VM: HTTP сервисов стека (SwarmUI `:7801` / Ollama `:11434` / llama.cpp `:8080`) — `verify_stack_on_vm`
+4. На VM: HTTP сервисов стека (SwarmUI `:7801` / Ollama `:11434`) — `verify_stack_on_vm`
 5. На VM: **nvidia-smi / CUDA** (fail-fast) + **torch+cuda** в Comfy venv при SwarmUI (ждём) — `verify_gpu_env`
 6. В логе: строка **`тайминг up:`** (SSH / bootstrap / Idle / verify / …)
 7. Туннель + проверка **localhost** тех же сервисов → access-card (красная рамка, если killer failed)

-`gpu-rent logs --unit swarm|ollama|llamacpp|killer` — фильтр journalctl.
+`gpu-rent logs --unit swarm|ollama|killer` — фильтр journalctl.
 `gpu-rent status` — killer/hold, последний стек/GPU-env, тайминг up.

 Локальный порт UI: **17801** (на VM по-прежнему 7801 на loopback).
@@ -24,7 +24,7 @@ GPU в облаке дорогой. Веса для генерации карт
 - Первый clone git-реп расширений SwarmUI и ComfyUI nodes из `extensions.yaml`.
 - Autocomplete: word-list в `Data/Autocompletions` до старта UI, на каждом `up` проверка новой версии.
 - `doctor` до create; фоллбек flavor; интерактивный выбор flavor/диска; `hold`; toast Idle; `open` на 17801.
- Opt-in LLM (Ollama / llama.cpp) через `setup` / флаги / `LLM_RUNTIME`; `capture` ссылок с VM.
+- Opt-in LLM (Ollama) через `setup` / флаги / `LLM_RUNTIME`; `capture` ссылок с VM.
 - `status`: диск, окно preempt 24 ч, killer / LLM.
 - Snapshot boot после первого удачного bootstrap.

@@ -10,7 +10,7 @@
 | Civitai хост | Дефолт API `civitai.red` (полный каталог). `.com` — SFW-витрина, NSFW с неё часто 404. Ссылки `.com`/`.red`/`.green` в манифесте принимаем. 404 → один retry на второй хост. Один токен на оба домена |
 | Пул GPU | Перед `up`/`flavors` сканируем `SCAN_POOLS` (дефолт `ru-6,ru-7`). `ru-6` — мультизональный: ходим на `https://ru-6.cloud.api.selcloud.ru/compute/` тем же токеном (SDK-каталог часто знает только RC-пул). Собираем типы GPU из extra_specs и совпадения с `FLAVOR_PREFERENCE`. Автоматом `.env` не пишем — печатаем рекомендацию `OS_REGION_NAME` / `GPU_RENT_AZ` |
 | Манифест моделей | `<repo>/models.yaml`, типы: checkpoint / lora / vae / embedding / controlnet / upscaler. В git только `models.example.yaml` |
-| Расширения | `<repo>/extensions.yaml`: git-репы `swarmui` → `src/Extensions`, `comfy` → DLNodes. Поле `requires: none\|ollama\|llamacpp\|any-llm` фильтрует по `LLM_RUNTIME`. Пример: swarm-assistent с `requires: ollama` |
+| Расширения | `<repo>/extensions.yaml`: git-репы `swarmui` → `src/Extensions`, `comfy` → DLNodes. Поле `requires: none\|ollama\|any-llm` фильтрует по `LLM_RUNTIME`. Пример: swarm-assistent с `requires: ollama` |
 | Autocomplete | До первого старта: скачать word-list в `Data/Autocompletions`, прописать `DefaultUser.AutoComplete.Source`. На каждом `up` сверить GitHub blob sha и обновить файл, если изменился. Дефолт: `tags/danbooru.csv` из a1111-sd-webui-tagcomplete (как в доке SwarmUI) |
 | Доступ | Браузер на туннеле; MCP переключается на облако, пока оно живо; HTTP API SwarmUI через тот же туннель |
 | Preemptible | По умолчанию всегда. Обычный сервер — только `--no-spot` |
@@ -22,7 +22,7 @@
 | `up` / `tunnel` | `up` по умолчанию после ready открывает туннель `:17801`, печатает URL и ждёт. `--no-tunnel` — только облако. Ctrl+C на туннеле GPU не гасит (`stop` отдельно). Команда `tunnel` остаётся для повторного входа |
 | Interactive `up` | Без `--yes`: нумерованные меню (LLM, пресет, flavor #, data GB, preemptible) → confirm. Изменения можно сохранить в `gpu-rent.vars` |
 | Git update | На каждом `up` по умолчанию: `git pull` SwarmUI + репы из `extensions.yaml` + уже установленные на data (`Extensions`/`DLNodes`). `--no-update` или `UPDATE_GIT=false` — не тянуть |
-| LLM (opt-in) | `none` по умолчанию. Флаги / `LLM_RUNTIME` / меню. Манифесты: `ollama-models.yaml`, `llamacpp-models.yaml`. Порты 17811 / 17812. Prompt-help, не замена SwarmUI |
+| LLM (opt-in) | `none` по умолчанию. Флаги / `LLM_RUNTIME` / меню. Манифест: `ollama-models.yaml`. Порт 17811. Prompt-help, не замена SwarmUI |
 | llm-only | `ENABLE_SWARMUI=false` / `WORKLOAD=llm` / `--no-swarm`/`--llm-only`: GPU + LLM без SwarmUI (bootstrap только data disk) |
 | Capture | `gpu-rent capture` — инвентарь VM → merge **ссылок** в локальные yaml (веса не качать) |
 | Perf auto-tune | На `up`: probe GPU → tier. Swarm/Comfy: sageattention ExtraArgs на Ampere+ ≥16 GiB. Ollama: flash/KV/keep-alive + `GPU_OVERHEAD` чтобы оставить VRAM под Krea |
@@ -38,7 +38,7 @@ swarmui:
  - url: https://gitea.hsrv.site/mrleo1nid/swarm-assistent.git
    ref: main
    dir: swarm-assistent
-    requires: ollama   # none | ollama | llamacpp | any-llm; default none = always
+    requires: ollama   # none | ollama | any-llm; default none = always
  - url: https://github.com/example/SwarmUI-SomeExt.git
    ref: main
    dir: SomeExt
@@ -51,8 +51,7 @@ comfy:
 | --- | --- |
 | `none` (или поле отсутствует) | всегда |
 | `ollama` | только при `LLM_RUNTIME=ollama` |
-| `llamacpp` | только при `LLM_RUNTIME=llamacpp` |
-| `any-llm` | при `ollama` или `llamacpp` |
+| `any-llm` | при `LLM_RUNTIME=ollama` |

 Строки с несовпавшим `requires` пропускаются (лог), остальные ставятся как обычно. В `extensions.example.yaml` по умолчанию — **swarm-assistent** с `requires: ollama` (чат/vision под Krea 2).

@@ -1,6 +1,6 @@
 # LLM рядом со SwarmUI (opt-in)

-По умолчанию поднимается **только SwarmUI**. Ollama или llama.cpp — отдельно (помощь с промптами и т.п.).
+По умолчанию поднимается **только SwarmUI**. Ollama — отдельно (помощь с промптами и т.п.).

 Интерактивные вопросы — **нумерованные меню** (Enter = вариант со ←). Ключ словом (`recommended`, `keep`) тоже принимается.

@@ -9,10 +9,10 @@
 | # | стек | смысл |
 | --- | --- | --- |
 | 1 | SwarmUI | только генерация картинок |
-| 2 | SwarmUI + LLM | UI + Ollama/llama.cpp на той же карте |
-| 3 | только LLM | **без** SwarmUI — Ollama или llama.cpp |
+| 2 | SwarmUI + LLM | UI + Ollama на той же карте |
+| 3 | только LLM | **без** SwarmUI — только Ollama |

-Флаги: `up --yes --llm-only --llamacpp` / `--no-swarm --ollama`. Vars: `ENABLE_SWARMUI=false` или `WORKLOAD=llm`.
+Флаги: `up --yes --llm-only --ollama` / `--no-swarm --ollama`. Vars: `ENABLE_SWARMUI=false` или `WORKLOAD=llm`.

 ---

@@ -30,8 +30,8 @@ gpu-rent setup

 ```text
 gpu-rent up --yes --ollama
-gpu-rent up --yes --llm llamacpp
-gpu-rent up --yes --llm-only --llamacpp   # без SwarmUI
+gpu-rent up --yes --llm ollama
+gpu-rent up --yes --llm-only --ollama   # без SwarmUI
 ```

 С `--yes` пресеты не спрашивает (берёт существующий yaml / example).
@@ -44,7 +44,7 @@ LLM_RUNTIME=ollama

 Потом `gpu-rent up --yes`.

-Выключить: `LLM_RUNTIME=none` (на `up` старые unit’ы `gpu-rent-ollama` / `gpu-rent-llamacpp` останавливаются).
+Выключить: `LLM_RUNTIME=none` (на `up` старые LLM unit’ы на VM останавливаются).

 Голый `gpu-rent` без args — **help**. Двойной клик: `GPU_RENT_DEFAULT_ARGS=up --yes`. Полный doctor на up: `up -v`.

@@ -56,7 +56,6 @@ LLM_RUNTIME=ollama
 | --- | --- | --- |
 | SwarmUI | 7801 | **17801** |
 | Ollama | 11434 | **17811** |
-| llama.cpp | 8080 | **17812** |

 ```text
 gpu-rent tunnel
@@ -111,61 +110,12 @@ Unit `gpu-rent-ollama` читает `/mnt/swarm_data/.gpu-rent-gpu.json`:

 ---

-## llama.cpp
-
-| Файл | Роль |
-| --- | --- |
-| `llamacpp-models.example.yaml` | шаблон |
-| `llamacpp-models.yaml` | HTTPS URL на `.gguf` (gitignore) |
-
-На `up`: скачать GGUF → `/mnt/swarm_data/llamacpp/models` → `llama-server` + systemd. Уже скачанные крупные файлы не трогает.
-
-**Бинарник:** в GitHub Releases **нет** Linux CUDA — только Windows CUDA + Ubuntu CPU/Vulkan. Драйвер/CUDA runtime на VM ≠ готовый `llama-server` с CUDA: его нужно **собрать** (`nvcc`) или взять Vulkan prebuilt.
-
-Порядок по умолчанию (`LLAMACPP_BACKEND=auto`):
-
-1. Если на VM уже есть `nvcc` (часто после прошлого `up`) → **CUDA-сборка** (тихо + heartbeat 5–15 мин).
-2. Иначе → Ubuntu **Vulkan** prebuilt (~30 MB).
-3. Fallback на другой путь при ошибке.
-
-Был только Vulkan-stamp, а `nvcc` появился — следующий `up` сам пересоберёт CUDA.
-
-| Var | Зачем |
-| --- | --- |
-| `LLAMACPP_BACKEND=auto\|cuda\|vulkan` | выбор пути (default auto) |
-| `LLAMACPP_TAG` | pin release (`b10545`) |
-| `LLAMACPP_ASSET_URL` + `LLAMACPP_SHA256` | свой архив |
-| `LLAMACPP_BUILD_CUDA=1` | форс CUDA; `=0` — никогда не собирать |
-| `LLAMACPP_FORCE_REINSTALL=1` | снести бинарь и поставить заново |
-| `LLAMACPP_NGL` / `LLAMACPP_CTX` | override GPU layers / context |
-| `LLAMACPP_HOST` / `LLAMACPP_PORT` | bind (default `127.0.0.1:8080`) |
-| `LLAMACPP_EXTRA_ARGS` | доп. флаги `llama-server` |
-| `OLLAMA_VERSION` / `OLLAMA_SHA256` | pin Ollama |
-
-Готовые кейсы A–H — в `gpu-rent.vars.example`.
-
-### Пресеты (меню)
-
-| # | ключ | что |
-| --- | --- | --- |
-| 1 | **recommended** | Qwen2.5-VL 7B abliterate + mmproj (~4.7+0.8 GB, картинки+RU) |
-| 2 | light | Qwen2.5 3B Instruct text Q4_K_M |
-| 3 | text | Qwen2.5 7B abliterate text-only |
-| 4 | stock | официальный 7B Instruct Q4_K_M |
-| 5 | empty | только runtime |
-| — | keep | не трогать yaml |
-
-`-ngl` / `-c` — по GPU probe. Для recommended abliterate GGUF нужен **`HF_TOKEN`** в `.env` (иначе 401). Вручную: положи GGUF в models и `systemctl restart gpu-rent-llamacpp`.
-
---
-
 ## Idle-killer и LLM

 Busy (не гасить GPU):

 - `ollama pull` (маркер младше ~45 мин; старше сбрасывается);
- загруженная модель в Ollama;
- llama.cpp: слоты заняты.
+- загруженная модель в Ollama.

 Ошибка установки LLM на `up` — **fail** (не тихий лог).

@@ -178,9 +128,6 @@ Busy (не гасить GPU):
 ```bash
 OLLAMA_VERSION=0.6.5
 OLLAMA_SHA256=<sha256 of ollama-linux-amd64.tgz>
-LLAMACPP_TAG=b10545
-# или LLAMACPP_ASSET_URL=... + LLAMACPP_SHA256=...
-# CUDA из исходников (медленно): LLAMACPP_BUILD_CUDA=1
-# Переустановка бинаря: LLAMACPP_FORCE_REINSTALL=1
-# VRAM: LLAMACPP_NGL=40  LLAMACPP_CTX=4096  LLAMACPP_EXTRA_ARGS=--flash-attn on
 ```
+
+Готовые кейсы — в `gpu-rent.vars.example`.
@@ -196,7 +196,7 @@ LOCAL_MODELS_DIR=  # пусто = <корень приложения>/Models

 ## Hugging Face (резерв)

- **Скачивание:** `HF_TOKEN` в `.env` → GGUF (llama.cpp) и строки `models.yaml` с `huggingface.co/…/resolve/…`.
+- **Скачивание:** `HF_TOKEN` в `.env` → строки `models.yaml` с `huggingface.co/…/resolve/…`.
 - **Capture:** если Civitai by-hash не нашёл файл → поиск на Hub по имени + LFS sha256 → URL в манифест.
 - **SwarmUI:** тот же токен прокидывается как `huggingface_api` (Model Downloader).
 - Abliterated / gated репозитории без токена почти всегда дают **401**.
@@ -49,7 +49,7 @@ chmod +x gpu-rent.sh
 - `env.example` → `.env`
 - `models.example.yaml` → `models.yaml`
 - `extensions.example.yaml` → `extensions.yaml`
- `ollama-models.example.yaml` / `llamacpp-models.example.yaml` → соответствующие yaml (лаунчер / setup)
+- `ollama-models.example.yaml` → `ollama-models.yaml` (лаунчер / setup)
 - `gpu-rent.vars.example` → `gpu-rent.vars`

 Секреты только в `<repo>/.env` и runtime в `<repo>/.gpu-rent/` (оба в `.gitignore`).
@@ -243,7 +243,7 @@ copy models.example.yaml models.yaml
 | --- | --- |
 | `--no-tunnel` | только облако; UI потом: `tunnel --open` |
 | `--no-update` | не `git pull` SwarmUI/extensions |
-| `--ollama` / `--llamacpp` | LLM рядом ([llm.md](llm.md)) |
+| `--ollama` / `--llm ollama` | LLM рядом ([llm.md](llm.md)) |
 | `--no-swarm` / `--llm-only` | только LLM, без SwarmUI (нужен runtime) |
 | `--no-spot` | обычный (не preemptible) тариф |
 | `--flavor ID` | явный flavor (пропускает меню flavor) |
@@ -108,7 +108,7 @@

 | Шаг | OK? |
 | --- | --- |
-| Ollama :17811 или llama.cpp :17812 через туннель | |
+| Ollama :17811 через туннель | |
 | pull / GGUF из манифеста | |
 | `LLM_RUNTIME=none` гасит unit на следующем up | |

@@ -20,7 +20,7 @@
 6. Если есть Civitai-токен и манифест моделей: seed весов **до** первого старта SwarmUI; иначе — дефолтная модель установщика.
 7. Push непустых `Models/` / `Wildcards/` / `CustomWorkflows/`.
 8. systemd unit `swarmui`: `./launch-linux.sh --launch_mode none --host 127.0.0.1 --port 7801`.
-9. **GPU probe** → `/mnt/swarm_data/.gpu-rent-gpu.json` (VRAM / compute cap / tier) — до старта UI; Ollama/llama.cpp читают его при install.
+9. **GPU probe** → `/mnt/swarm_data/.gpu-rent-gpu.json` (VRAM / compute cap / tier) — до старта UI; Ollama читает его при install.
 10. Старт SwarmUI → seed LLM → idle-killer → **wait backend Idle**.
 11. **Perf tune после Idle** — `triton`+`sageattention` в Comfy venv и `--use-sage-attention` в `Data/Backends.fds` (маркер `.gpu-rent-perf-tuned`; повтор при смене GPU).
 12. Авторизация SwarmUI включена, токен в `Data` на диске.
@@ -38,29 +38,18 @@ AUTOCOMPLETE_GITHUB_REF=main
 AUTOCOMPLETE_FILENAME=danbooru.csv

 SWARMUI_LOCAL_PORT=17801
-# Optional LLM beside SwarmUI: none | ollama | llamacpp (setup / --ollama / --llamacpp)
+# Optional LLM beside SwarmUI: none | ollama (setup / --ollama / --llm ollama)
 # ENABLE_SWARMUI=true
-# WORKLOAD=llm          # llm-only (no SwarmUI); requires LLM_RUNTIME≠none
+# WORKLOAD=llm          # llm-only (no SwarmUI); requires LLM_RUNTIME=ollama
 LLM_RUNTIME=none
 OLLAMA_LOCAL_PORT=17811
-LLAMACPP_LOCAL_PORT=17812
 # OLLAMA_MODELS_MANIFEST=
-# LLAMACPP_MODELS_MANIFEST=
-# Pin / тонкая настройка LLM (несecреты; удобнее в gpu-rent.vars — см. кейсы A–H):
-# LLAMACPP_TAG=b10545
-# LLAMACPP_BACKEND=auto
-# LLAMACPP_BUILD_CUDA=1
-# LLAMACPP_FORCE_REINSTALL=1
-# LLAMACPP_ASSET_URL=
-# LLAMACPP_SHA256=
-# LLAMACPP_NGL=40
-# LLAMACPP_CTX=4096
-# LLAMACPP_EXTRA_ARGS=--flash-attn on
+# Pin Ollama (несecреты; удобнее в gpu-rent.vars):
 # OLLAMA_VERSION=0.6.5
 # OLLAMA_SHA256=
 # CIVITAI_API_TOKEN=
 # CIVITAI_API_HOST=civitai.red
-# Hugging Face (GGUF / gated HF URLs / capture fallback metadata):
+# Hugging Face (HF URLs in models.yaml / capture fallback metadata):
 # HF_TOKEN=   # or HUGGING_FACE_HUB_TOKEN — https://huggingface.co/settings/tokens
 # git pull SwarmUI + extensions on each up (default true). CLI: --no-update
 UPDATE_GIT=true
@@ -2,8 +2,8 @@
 # Empty/missing file → no extra extensions, stock SwarmUI.
 # swarmui = C# repos cloned to src/Extensions
 # comfy   = Python custom nodes cloned to ComfyUI DLNodes
-# requires: none (default) | ollama | llamacpp | any-llm
-#   — clone only when LLM_RUNTIME matches (ollama / llamacpp / either)
+# requires: none (default) | ollama | any-llm
+#   — clone only when LLM_RUNTIME matches (ollama / either LLM)

 swarmui:
  - url: https://gitea.hsrv.site/mrleo1nid/swarm-assistent.git
@@ -119,7 +119,6 @@ function Copy-IfMissing {
 Copy-IfMissing (Join-Path $Root "models.example.yaml") (Join-Path $Root "models.yaml") "models.yaml"
 Copy-IfMissing (Join-Path $Root "extensions.example.yaml") (Join-Path $Root "extensions.yaml") "extensions.yaml"
 Copy-IfMissing (Join-Path $Root "ollama-models.example.yaml") (Join-Path $Root "ollama-models.yaml") "ollama-models.yaml"
-Copy-IfMissing (Join-Path $Root "llamacpp-models.example.yaml") (Join-Path $Root "llamacpp-models.yaml") "llamacpp-models.yaml"
 Copy-IfMissing (Join-Path $Root "gpu-rent.vars.example") (Join-Path $Root "gpu-rent.vars") "gpu-rent.vars"

 Import-GpuRentVars (Join-Path $Root "gpu-rent.vars")
@@ -99,10 +99,6 @@ if [[ ! -f "$ROOT/ollama-models.yaml" && -f "$ROOT/ollama-models.example.yaml" ]
  cp "$ROOT/ollama-models.example.yaml" "$ROOT/ollama-models.yaml"
  echo "gpu-rent: created ollama-models.yaml"
 fi
-if [[ ! -f "$ROOT/llamacpp-models.yaml" && -f "$ROOT/llamacpp-models.example.yaml" ]]; then
-  cp "$ROOT/llamacpp-models.example.yaml" "$ROOT/llamacpp-models.yaml"
-  echo "gpu-rent: created llamacpp-models.yaml"
-fi
 if [[ ! -f "$ROOT/gpu-rent.vars" && -f "$ROOT/gpu-rent.vars.example" ]]; then
  cp "$ROOT/gpu-rent.vars.example" "$ROOT/gpu-rent.vars"
  echo "gpu-rent: created gpu-rent.vars"
@@ -31,59 +31,24 @@
 # Подробнее: docs/llm.md
 # ---------------------------------------------------------------------------

-# --- A) Default: SwarmUI + llama.cpp (auto: CUDA если nvcc на VM, иначе Vulkan) ---
-# LLM_RUNTIME=llamacpp
+# --- A) Default: SwarmUI + Ollama ---
+# LLM_RUNTIME=ollama
 # ENABLE_SWARMUI=true
-# LLAMACPP_TAG=b10545
-# LLAMACPP_BACKEND=auto
+# OLLAMA_LOCAL_PORT=17811

 # --- B) Только LLM, без SwarmUI ---
 # WORKLOAD=llm
-# LLM_RUNTIME=llamacpp
-# LLAMACPP_TAG=b10545
+# LLM_RUNTIME=ollama

-# --- C) Явно CUDA (и снести старый Vulkan-бинарь) ---
-# LLM_RUNTIME=llamacpp
-# LLAMACPP_TAG=b10545
-# LLAMACPP_BACKEND=cuda
-# LLAMACPP_FORCE_REINSTALL=1
-
-# --- C2) Явно быстрый Vulkan, без compile ---
-# LLM_RUNTIME=llamacpp
-# LLAMACPP_BACKEND=vulkan
-# LLAMACPP_BUILD_CUDA=0
-
-# --- D) Свой бинарь / pin URL (supply-chain) ---
-# LLM_RUNTIME=llamacpp
-# LLAMACPP_ASSET_URL=https://github.com/ggml-org/llama.cpp/releases/download/b10545/llama-b10545-bin-ubuntu-vulkan-x64.tar.gz
-# LLAMACPP_SHA256=<sha256 архива>
-# LLAMACPP_FORCE_REINSTALL=1
-
-# --- E) Делим 4090 со SwarmUI: меньше слоёв / короче контекст ---
-# LLM_RUNTIME=llamacpp
-# LLAMACPP_NGL=40
-# LLAMACPP_CTX=4096
-# LLAMACPP_EXTRA_ARGS=--flash-attn on
-
-# --- F) Длинный контекст / почти весь VRAM под LLM (llm-only) ---
-# WORKLOAD=llm
-# LLM_RUNTIME=llamacpp
-# LLAMACPP_NGL=99
-# LLAMACPP_CTX=32768
-# LLAMACPP_EXTRA_ARGS=--parallel 1
-
-# --- G) Ollama вместо llama.cpp + pin версии ---
+# --- C) Pin версии Ollama (supply-chain) ---
 # LLM_RUNTIME=ollama
 # OLLAMA_VERSION=0.6.5
 # OLLAMA_SHA256=<sha256 of ollama-linux-amd64.tgz>
-# OLLAMA_LOCAL_PORT=17811

-# --- H) Выключить LLM (гасит unit’ы на следующем up) ---
+# --- D) Выключить LLM (гасит unit’ы на следующем up) ---
 # LLM_RUNTIME=none

-# Порты туннеля (localhost):
+# Порт туннеля (localhost):
 # OLLAMA_LOCAL_PORT=17811
-# LLAMACPP_LOCAL_PORT=17812
-# Свой путь к манифесту GGUF/pull:
-# LLAMACPP_MODELS_MANIFEST=
+# Свой путь к манифесту pull:
 # OLLAMA_MODELS_MANIFEST=
@@ -1,16 +0,0 @@
-# Copy to llamacpp-models.yaml (gitignored). Used when LLM_RUNTIME=llamacpp.
-# url = direct HTTPS link to a .gguf; mmproj = vision projector (Qwen2.5-VL).
-# Empty models: [] → only llama-server, GGUF клади вручную.
-# Purpose: prompt-help / vision beside SwarmUI (RU/EN, low refusal).
-
-models:
-  # Recommended: Qwen2.5-VL 7B abliterate Q4_K_M + mmproj (~4.7 + 0.8 GB)
-  - url: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-abliterated-GGUF/resolve/main/Qwen2.5-VL-7B-Instruct-abliterated.Q4_K_M.gguf
-    mmproj: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-abliterated-GGUF/resolve/main/Qwen2.5-VL-7B-Instruct-abliterated.mmproj-Q8_0.gguf
-    default: true
-
-  # Text-only abliterate 7B (~4.7 GB):
-  # - url: https://huggingface.co/RichardErkhov/huihui-ai_-_Qwen2.5-7B-Instruct-abliterated-gguf/resolve/main/Qwen2.5-7B-Instruct-abliterated.Q4_K_M.gguf
-
-  # Lighter text (~2 GB):
-  # - url: https://huggingface.co/bartowski/Qwen2.5-3B-Instruct-GGUF/resolve/main/Qwen2.5-3B-Instruct-Q4_K_M.gguf
@@ -1,17 +1,15 @@
 # Copy to ollama-models.yaml (gitignored). Used when LLM_RUNTIME=ollama.
-# name = exact tag for `ollama pull`. Empty models: [] → runtime only, no pull.
-# Purpose: SwarmUI prompt help + vision (RU/EN, low refusal).
+# name = exact tag for `ollama pull` (Ollama library / community).
+# Empty models: [] → runtime only, no pull.
+# Purpose: SwarmUI prompt help — vision + RU/EN on the same GPU as diffusion.

 models:
-  # Recommended: vision + Russian/English, ~6GB, abliterated
+  # Recommended (~6GB): vision + Russian/English, low refusal
  - name: huihui_ai/qwen2.5-vl-abliterated:7b
    default: true

-  # Lighter vision:
-  # - name: huihui_ai/qwen2.5-vl-abliterated:3b
-
-  # Text-only abliterate (~5GB):
-  # - name: huihui_ai/qwen2.5-abliterate:7b
-
-  # Official stock (more refusals):
-  # - name: qwen2.5:7b
+  # Alternatives (uncomment / use presets on setup|up):
+  # light  — huihui_ai/qwen2.5-vl-abliterated:3b   (~3GB)
+  # stock  — qwen2.5vl:7b                          (official library)
+  # text   — huihui_ai/qwen2.5-abliterate:7b       (no vision, ~5GB)
+  # big    — qwen2.5vl:32b                         (~21GB; llm-only)
@@ -71,19 +71,6 @@ def collect_access_links(cfg: Config, *, tunneled: bool) -> list[AccessLink]:
                    ),
                ]
            )
-        elif runtime == "llamacpp":
-            p = cfg.llamacpp_local_port
-            links.extend(
-                [
-                    AccessLink("llama.cpp", f"http://127.0.0.1:{p}", "OpenAI-compatible"),
-                    AccessLink(
-                        "OpenAI /v1",
-                        f"http://127.0.0.1:{p}/v1/chat/completions",
-                        "chat completions",
-                    ),
-                    AccessLink("Models", f"http://127.0.0.1:{p}/v1/models", "list"),
-                ]
-            )
        if not links:
            links.append(
                AccessLink(
@@ -138,7 +125,7 @@ def render_access_panel(
    cmds.add_column(style="dim", no_wrap=True)
    cmds.add_column()
    cmds.add_row("открыть UI", "gpu-rent open")
-    if resolve_llm_runtime(cfg) in {"ollama", "llamacpp"}:
+    if resolve_llm_runtime(cfg) == "ollama":
        cmds.add_row("открыть LLM", "gpu-rent open --llm")
    cmds.add_row("hold killer", "gpu-rent hold")
    cmds.add_row("стоп GPU", "gpu-rent stop")
@@ -335,7 +335,7 @@ def status() -> None:
    rt = resolve_llm_runtime(cfg)
    noted = notes.get("llm_runtime")
    llm_err = notes.get("llm_error")
-    detail = f"{rt}; ollama :{cfg.ollama_local_port} / llamacpp :{cfg.llamacpp_local_port}"
+    detail = f"{rt}; ollama :{cfg.ollama_local_port}"
    if noted and noted != rt:
        detail += f" (notes: {noted})"
    if llm_err:
@@ -370,7 +370,7 @@ def status() -> None:
 def open(
    llm: bool = typer.Option(False, "--llm", help="Открыть LLM API URL вместо SwarmUI"),
 ) -> None:
-    """Открыть браузер на SwarmUI :17801 (или --llm / llm-only на Ollama/llama.cpp)."""
+    """Открыть браузер на SwarmUI :17801 (или --llm / llm-only на Ollama)."""
    cfg = load_config(require_auth=False)
    use_llm = llm or not bool(getattr(cfg, "enable_swarmui", True))
    if use_llm:
@@ -379,8 +379,6 @@ def open(
        runtime = resolve_llm_runtime(cfg)
        if runtime == "ollama":
            port = cfg.ollama_local_port
-        elif runtime == "llamacpp":
-            port = cfg.llamacpp_local_port
        else:
            console.print("[red]LLM не выбран[/red] (LLM_RUNTIME / gpu-rent setup)")
            raise typer.Exit(1)
@@ -398,9 +396,9 @@ def open(

@app.command()
 def setup(
-    llm: Optional[str] = typer.Option(None, "--llm", help="none|ollama|llamacpp"),
+    llm: Optional[str] = typer.Option(None, "--llm", help="none|ollama"),
    ollama_preset: Optional[str] = typer.Option(
-        None, "--ollama-preset", help="recommended|light|stock|alt|empty"
+        None, "--ollama-preset", help="recommended|light|stock|text|big|empty"
    ),
    watchdog: Optional[bool] = typer.Option(
        None, "--watchdog/--no-watchdog", help="Поставить local-watchdog"
@@ -466,10 +464,9 @@ def up(
        help="Полный doctor-таблица на up (по умолчанию кратко)",
    ),
    llm: Optional[str] = typer.Option(
-        None, "--llm", help="none|ollama|llamacpp (override LLM_RUNTIME)"
+        None, "--llm", help="none|ollama (override LLM_RUNTIME)"
    ),
    ollama: bool = typer.Option(False, "--ollama", help="То же что --llm ollama"),
-    llamacpp: bool = typer.Option(False, "--llamacpp", help="То же что --llm llamacpp"),
    no_swarm: bool = typer.Option(
        False,
        "--no-swarm",
@@ -489,7 +486,7 @@ def up(
            write_ollama_models_preset,
        )
        from gpu_rent.paths import vars_path
-        from gpu_rent.prompts import MenuItem, prompt_menu
+        from gpu_rent.prompts import prompt_menu
        from gpu_rent.timing import clock_elapsed, clock_reset, format_duration
        from gpu_rent.varsfile import upsert_vars

@@ -506,7 +503,6 @@ def up(
            runtime = decide_runtime(
                flag=llm,
                ollama_flag=ollama,
-                llamacpp_flag=llamacpp,
                from_config=cfg.llm_runtime,
            )
        except ValueError as exc:
@@ -514,11 +510,10 @@ def up(

        enable_swarm = False if no_swarm else cfg.enable_swarmui
        asked_model_preset = False
-        llm_flags = bool(llm or ollama or llamacpp or no_swarm)
+        llm_flags = bool(llm or ollama or no_swarm)

        if not yes and not llm_flags:
            from gpu_rent.llm_runtime import (
-                llamacpp_preset_menu,
                ollama_preset_menu,
                workload_menu,
            )
@@ -548,47 +543,11 @@ def up(
            elif stack == "both":
                enable_swarm = True
                if runtime == "none":
-                    try:
-                        choice = prompt_menu(
-                            "LLM runtime",
-                            [
-                                MenuItem("ollama", "Ollama (+ pull моделей)"),
-                                MenuItem("llamacpp", "llama.cpp server (+ GGUF)"),
-                            ],
-                            default="ollama",
-                            ask=_ask,
-                            show=log,
-                        )
-                        runtime = decide_runtime(
-                            flag=choice,
-                            ollama_flag=False,
-                            llamacpp_flag=False,
-                            from_config="none",
-                        )
-                    except ValueError as exc:
-                        raise GpuRentError(str(exc)) from exc
+                    runtime = "ollama"
            else:
                enable_swarm = False
                if runtime == "none":
-                    try:
-                        choice = prompt_menu(
-                            "LLM runtime",
-                            [
-                                MenuItem("ollama", "Ollama (+ pull моделей)"),
-                                MenuItem("llamacpp", "llama.cpp server (+ GGUF)"),
-                            ],
-                            default="llamacpp",
-                            ask=_ask,
-                            show=log,
-                        )
-                        runtime = decide_runtime(
-                            flag=choice,
-                            ollama_flag=False,
-                            llamacpp_flag=False,
-                            from_config="none",
-                        )
-                    except ValueError as exc:
-                        raise GpuRentError(str(exc)) from exc
+                    runtime = "ollama"

            if typer.confirm("Запомнить стек в gpu-rent.vars?", default=True):
                upsert_vars(
@@ -614,68 +573,31 @@ def up(
                if preset not in {"keep", "example"}:
                    write_ollama_models_preset(cfg.ollama_models_manifest, preset)
                asked_model_preset = True
-            elif runtime == "llamacpp":
-                from gpu_rent.llm_runtime import (
-                    ensure_llamacpp_manifest_from_example,
-                    write_llamacpp_models_preset,
-                )
-
-                ensure_llamacpp_manifest_from_example()
-                try:
-                    preset = prompt_menu(
-                        "llama.cpp GGUF",
-                        llamacpp_preset_menu(include_keep=False),
-                        default="recommended",
-                        ask=_ask,
-                        show=log,
-                    )
-                except ValueError as exc:
-                    raise GpuRentError(str(exc)) from exc
-                if preset not in {"keep", "example"}:
-                    write_llamacpp_models_preset(cfg.llamacpp_models_manifest, preset)
-                asked_model_preset = True

        # Runtime уже в vars — спросить пресет, default=keep.
-        if not yes and not asked_model_preset and runtime in {"llamacpp", "ollama"}:
-            from gpu_rent.llm_runtime import (
-                ensure_llamacpp_manifest_from_example,
-                llamacpp_preset_menu,
-                ollama_preset_menu,
-                write_llamacpp_models_preset,
-            )
+        if not yes and not asked_model_preset and runtime == "ollama":
+            from gpu_rent.llm_runtime import ollama_preset_menu

            def _ask2(msg: str, default: str = "") -> str:
                return typer.prompt(msg, default=default)

            try:
-                if runtime == "llamacpp":
-                    ensure_llamacpp_manifest_from_example()
-                    key = prompt_menu(
-                        "llama.cpp GGUF",
-                        llamacpp_preset_menu(include_keep=True),
-                        default="keep",
-                        ask=_ask2,
-                        show=log,
-                    )
-                    if key not in {"keep", "example", ""}:
-                        write_llamacpp_models_preset(cfg.llamacpp_models_manifest, key)
-                else:
-                    ensure_ollama_manifest_from_example()
-                    key = prompt_menu(
-                        "Ollama preset",
-                        ollama_preset_menu(include_keep=True),
-                        default="keep",
-                        ask=_ask2,
-                        show=log,
-                    )
-                    if key not in {"keep", "example", ""}:
-                        write_ollama_models_preset(cfg.ollama_models_manifest, key)
+                ensure_ollama_manifest_from_example()
+                key = prompt_menu(
+                    "Ollama preset",
+                    ollama_preset_menu(include_keep=True),
+                    default="keep",
+                    ask=_ask2,
+                    show=log,
+                )
+                if key not in {"keep", "example", ""}:
+                    write_ollama_models_preset(cfg.ollama_models_manifest, key)
            except ValueError as exc:
                raise GpuRentError(str(exc)) from exc

        if not enable_swarm and runtime == "none":
            raise GpuRentError(
-                "llm-only требует --ollama / --llamacpp / --llm … "
+                "llm-only требует --ollama / --llm ollama "
                "(или убери --no-swarm / ENABLE_SWARMUI=true)"
            )

@@ -796,7 +718,7 @@ def logs(
        None,
        "--unit",
        "-u",
-        help="swarm|ollama|llamacpp|killer|cloud-init (по умолчанию — всё)",
+        help="swarm|ollama|killer|cloud-init (по умолчанию — всё)",
    ),
    lines: int = typer.Option(80, "--lines", "-n", help="Строк journalctl"),
 ) -> None:
@@ -812,8 +734,6 @@ def logs(
            "swarm": "swarmui",
            "swarmui": "swarmui",
            "ollama": "ollama",
-            "llamacpp": "llamacpp",
-            "llama": "llamacpp",
            "killer": "gpu-rent-idle-killer",
            "idle-killer": "gpu-rent-idle-killer",
            "idle": "gpu-rent-idle-killer",
@@ -823,7 +743,7 @@ def logs(
        if key not in aliases:
            raise GpuRentError(
                f"неизвестный --unit={unit!r}; "
-                "ожидаю: swarm|ollama|llamacpp|killer|cloud-init|all"
+                "ожидаю: swarm|ollama|killer|cloud-init|all"
            )
        target = aliases[key]
        n = max(10, min(int(lines), 500))
@@ -835,7 +755,7 @@ def logs(
            )
        journal_units = []
        if target == "all":
-            journal_units = ["swarmui", "ollama", "llamacpp", "gpu-rent-idle-killer"]
+            journal_units = ["swarmui", "ollama", "gpu-rent-idle-killer"]
        elif target != "cloud-init":
            journal_units = [target]
        for ju in journal_units:
@@ -14,7 +14,6 @@ from gpu_rent.paths import (
    default_ssh_key_path,
    env_path,
    extensions_manifest_path,
-    llamacpp_models_manifest_path,
    migrate_legacy_if_needed,
    models_manifest_path,
    ollama_models_manifest_path,
@@ -107,9 +106,7 @@ class Config:
    llm_runtime: str
    enable_swarmui: bool
    ollama_models_manifest: Path
-    llamacpp_models_manifest: Path
    ollama_local_port: int
-    llamacpp_local_port: int

    default_flavor_id: str
    flavor_preference: tuple[str, ...]
@@ -185,10 +182,6 @@ def load_config(*, require_auth: bool = True) -> Config:
        (os.environ.get("OLLAMA_MODELS_MANIFEST") or "").strip()
        or str(ollama_models_manifest_path())
    ).expanduser()
-    llamacpp_manifest = Path(
-        (os.environ.get("LLAMACPP_MODELS_MANIFEST") or "").strip()
-        or str(llamacpp_models_manifest_path())
-    ).expanduser()

    try:
        llm_runtime = normalize_runtime(os.environ.get("LLM_RUNTIME"))
@@ -243,9 +236,7 @@ def load_config(*, require_auth: bool = True) -> Config:
        llm_runtime=llm_runtime,
        enable_swarmui=enable_swarmui,
        ollama_models_manifest=ollama_manifest,
-        llamacpp_models_manifest=llamacpp_manifest,
        ollama_local_port=_as_int(os.environ.get("OLLAMA_LOCAL_PORT"), 17811),
-        llamacpp_local_port=_as_int(os.environ.get("LLAMACPP_LOCAL_PORT"), 17812),
        default_flavor_id=(os.environ.get("DEFAULT_FLAVOR_ID") or "").strip(),
        flavor_preference=_csv(
            os.environ.get("FLAVOR_PREFERENCE"),
@@ -282,7 +282,6 @@ def _civitai(cfg: Config, checks: list[Check]) -> None:

 def _huggingface(cfg: Config, checks: list[Check]) -> None:
    from gpu_rent.huggingface import is_huggingface_url, probe_whoami
-    from gpu_rent.llm_runtime import normalize_runtime, parse_llamacpp_models
    from gpu_rent.manifests import parse_models

    needs_hf = False
@@ -293,14 +292,6 @@ def _huggingface(cfg: Config, checks: list[Check]) -> None:
                break
    except Exception:
        pass
-    try:
-        if normalize_runtime(cfg.llm_runtime) == "llamacpp":
-            for e in parse_llamacpp_models(cfg.llamacpp_models_manifest):
-                if e.url and is_huggingface_url(e.url):
-                    needs_hf = True
-                    break
-    except Exception:
-        pass

    if not cfg.hf_token:
        checks.append(
@@ -309,7 +300,7 @@ def _huggingface(cfg: Config, checks: list[Check]) -> None:
                True,
                False,
                (
-                    "HF_TOKEN нет — gated GGUF / HF в models.yaml дадут 401. "
+                    "HF_TOKEN нет — gated HF URL в models.yaml дадут 401. "
                    "https://huggingface.co/settings/tokens"
                    if needs_hf
                    else "токена нет (опционально для HF URL / capture fallback)"
@@ -1,39 +1,37 @@
-"""Optional LLM runtimes (Ollama / llama.cpp) beside SwarmUI."""
+"""Optional LLM runtime (Ollama) beside SwarmUI."""

 from __future__ import annotations

 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any
-from urllib.parse import unquote, urlparse

 import yaml

 from gpu_rent.paths import (
-    llamacpp_models_example_path,
-    llamacpp_models_manifest_path,
    ollama_models_example_path,
    ollama_models_manifest_path,
 )

-VALID_RUNTIMES = frozenset({"none", "ollama", "llamacpp"})
+VALID_RUNTIMES = frozenset({"none", "ollama"})

 OLLAMA_PRESETS: dict[str, list[str]] = {
-    # Vision + RU/EN + low refusal — best default for swarm-assistent / prompt help with images
-    "recommended": ["huihui_ai/qwen2.5-vl-abliterated:7b"],
-    "light": ["huihui_ai/qwen2.5-vl-abliterated:3b"],
-    "text": ["huihui_ai/qwen2.5-abliterate:7b"],
-    "stock": ["qwen2.5:7b"],
-    "alt": ["richardyoung/qwen2.5-7b-instruct-abliterated"],
+    # Ollama library / community tags only (no GGUF). Tuned for SwarmUI prompt help:
+    # vision + RU/EN, share VRAM with diffusion on typical 24 GiB.
+    "recommended": ["huihui_ai/qwen2.5-vl-abliterated:7b"],  # ~6 GB, low refusal
+    "light": ["huihui_ai/qwen2.5-vl-abliterated:3b"],  # ~3 GB, tight VRAM
+    "stock": ["qwen2.5vl:7b"],  # official library vision
+    "text": ["huihui_ai/qwen2.5-abliterate:7b"],  # ~5 GB, no vision
+    "big": ["qwen2.5vl:32b"],  # ~21 GB — llm-only or ≥40 GiB free
    "empty": [],
 }

 OLLAMA_PRESET_LABELS: dict[str, str] = {
    "recommended": "Qwen2.5-VL 7B abliterate (картинки+RU, ~6GB)",
-    "light": "Qwen2.5-VL 3B abliterate (vision, быстрее, ~3GB)",
+    "light": "Qwen2.5-VL 3B abliterate (мало VRAM, ~3GB)",
+    "stock": "официальный qwen2.5vl:7b (library, больше отказов)",
    "text": "Qwen2.5 7B abliterate text-only (~5GB)",
-    "stock": "официальный qwen2.5:7b (больше цензуры)",
-    "alt": "другой text abliterate-пак 7B",
+    "big": "qwen2.5vl:32b (~21GB; llm-only / большой GPU)",
    "empty": "только runtime, без pull",
    "keep": "не менять ollama-models.yaml",
 }
@@ -43,64 +41,9 @@ PRESET_HELP = "\n".join(
    f"{k} — {v}" for k, v in OLLAMA_PRESET_LABELS.items() if k != "keep"
 )

-# Each preset entry: {"url": "...gguf", "mmproj": optional vision projector url}
-LLAMACPP_PRESETS: dict[str, list[dict[str, str]]] = {
-    "recommended": [
-        {
-            "url": (
-                "https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-abliterated-GGUF/"
-                "resolve/main/Qwen2.5-VL-7B-Instruct-abliterated.Q4_K_M.gguf"
-            ),
-            "mmproj": (
-                "https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-abliterated-GGUF/"
-                "resolve/main/Qwen2.5-VL-7B-Instruct-abliterated.mmproj-Q8_0.gguf"
-            ),
-        },
-    ],
-    "light": [
-        {
-            "url": (
-                "https://huggingface.co/bartowski/Qwen2.5-3B-Instruct-GGUF/"
-                "resolve/main/Qwen2.5-3B-Instruct-Q4_K_M.gguf"
-            ),
-        },
-    ],
-    "text": [
-        {
-            "url": (
-                "https://huggingface.co/RichardErkhov/huihui-ai_-_Qwen2.5-7B-Instruct-abliterated-gguf/"
-                "resolve/main/Qwen2.5-7B-Instruct-abliterated.Q4_K_M.gguf"
-            ),
-        },
-    ],
-    "stock": [
-        {
-            "url": (
-                "https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF/"
-                "resolve/main/Qwen2.5-7B-Instruct-Q4_K_M.gguf"
-            ),
-        },
-    ],
-    "empty": [],
-}
-
-LLAMACPP_PRESET_LABELS: dict[str, str] = {
-    "recommended": "Qwen2.5-VL 7B abliterate + mmproj (~4.7+0.8GB, картинки+RU)",
-    "light": "Qwen2.5 3B Instruct Q4_K_M text (~2GB)",
-    "text": "Qwen2.5 7B abliterate text-only Q4_K_M (~4.7GB)",
-    "stock": "официальный Qwen2.5 7B Instruct Q4_K_M",
-    "empty": "только llama-server, GGUF вручную",
-    "keep": "не менять llamacpp-models.yaml",
-}
-
-LLAMACPP_PRESET_HELP = "\n".join(
-    f"{k} — {v}" for k, v in LLAMACPP_PRESET_LABELS.items() if k != "keep"
-)
-
 LLM_RUNTIME_LABELS: dict[str, str] = {
    "none": "только SwarmUI",
    "ollama": "Ollama (+ pull моделей)",
-    "llamacpp": "llama.cpp server (+ GGUF)",
 }

 WORKLOAD_LABELS: dict[str, str] = {
@@ -113,7 +56,7 @@ WORKLOAD_LABELS: dict[str, str] = {
 def llm_runtime_menu() -> list:
    from gpu_rent.prompts import MenuItem

-    return [MenuItem(k, f"{k} — {LLM_RUNTIME_LABELS[k]}") for k in ("none", "ollama", "llamacpp")]
+    return [MenuItem(k, f"{k} — {LLM_RUNTIME_LABELS[k]}") for k in ("none", "ollama")]


 def workload_menu() -> list:
@@ -131,53 +74,29 @@ def ollama_preset_menu(*, include_keep: bool = False) -> list:
    return [MenuItem(k, OLLAMA_PRESET_LABELS.get(k, k)) for k in keys]


-def llamacpp_preset_menu(*, include_keep: bool = False) -> list:
-    from gpu_rent.prompts import MenuItem
-
-    keys = list(LLAMACPP_PRESETS.keys())
-    if include_keep:
-        keys.append("keep")
-    return [MenuItem(k, LLAMACPP_PRESET_LABELS.get(k, k)) for k in keys]
-
-
@dataclass(frozen=True)
 class OllamaModelEntry:
    name: str
    default: bool = False


-@dataclass(frozen=True)
-class LlamaCppModelEntry:
-    url: str
-    filename: str | None = None
-    mmproj_url: str | None = None
-    default: bool = False
-
-
 def normalize_runtime(value: str | None) -> str:
    raw = (value or "none").strip().lower().replace("-", "").replace("_", "")
    if raw in {"", "none", "off", "no", "0"}:
        return "none"
    if raw in {"ollama"}:
        return "ollama"
-    if raw in {"llamacpp", "llama", "llamacppserver"}:
-        return "llamacpp"
-    raise ValueError(f"неизвестный LLM_RUNTIME={value!r}; жду none|ollama|llamacpp")
+    raise ValueError(f"неизвестный LLM_RUNTIME={value!r}; жду none|ollama")


 def decide_runtime(
    *,
    flag: str | None,
    ollama_flag: bool,
-    llamacpp_flag: bool,
    from_config: str,
 ) -> str:
-    if ollama_flag and llamacpp_flag:
-        raise ValueError("укажи только --ollama или --llamacpp, не оба")
    if ollama_flag:
        return "ollama"
-    if llamacpp_flag:
-        return "llamacpp"
    if flag is not None and str(flag).strip() != "":
        return normalize_runtime(flag)
    return normalize_runtime(from_config)
@@ -210,65 +129,6 @@ def parse_ollama_models(path: Path) -> list[OllamaModelEntry]:
    return out


-def gguf_filename_from_url(url: str) -> str:
-    path = unquote(urlparse(url).path)
-    name = path.rsplit("/", 1)[-1] if path else ""
-    if name.lower().endswith(".gguf"):
-        return name
-    return "model.gguf"
-
-
-# Dead / moved HF mirrors → current resolve URL (same Q4_K_M abliterate weights).
-_LLAMACPP_URL_ALIASES: dict[str, str] = {
-    "https://huggingface.co/bartowski/huihui-ai_Qwen2.5-7B-Instruct-abliterated-GGUF/resolve/main/huihui-ai_Qwen2.5-7B-Instruct-abliterated-Q4_K_M.gguf": (
-        "https://huggingface.co/RichardErkhov/huihui-ai_-_Qwen2.5-7B-Instruct-abliterated-gguf/resolve/main/Qwen2.5-7B-Instruct-abliterated.Q4_K_M.gguf"
-    ),
-}
-
-
-def remap_llamacpp_url(url: str) -> str:
-    """Rewrite known-dead GGUF mirrors so old llamacpp-models.yaml still works."""
-    key = (url or "").strip()
-    return _LLAMACPP_URL_ALIASES.get(key, key)
-
-
-def parse_llamacpp_models(path: Path) -> list[LlamaCppModelEntry]:
-    if not path.is_file():
-        return []
-    raw = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
-    if not isinstance(raw, dict):
-        return []
-    items = raw.get("models")
-    if items is None:
-        return []
-    if not isinstance(items, list):
-        raise ValueError(f"{path}: models должен быть списком")
-    out: list[LlamaCppModelEntry] = []
-    for item in items:
-        if isinstance(item, str):
-            url = remap_llamacpp_url(item.strip())
-            if url:
-                out.append(LlamaCppModelEntry(url=url))
-            continue
-        if not isinstance(item, dict):
-            continue
-        url = remap_llamacpp_url(str(item.get("url") or "").strip())
-        if not url:
-            continue
-        fname = item.get("filename")
-        filename = str(fname).strip() if fname else None
-        mmproj = remap_llamacpp_url(str(item.get("mmproj") or item.get("mmproj_url") or "").strip())
-        out.append(
-            LlamaCppModelEntry(
-                url=url,
-                filename=filename or None,
-                mmproj_url=mmproj or None,
-                default=bool(item.get("default")),
-            )
-        )
-    return out
-
-
 def write_ollama_models_preset(path: Path, preset: str) -> None:
    key = (preset or "recommended").strip().lower()
    if key not in OLLAMA_PRESETS:
@@ -289,32 +149,6 @@ def write_ollama_models_preset(path: Path, preset: str) -> None:
    path.write_text("\n".join(lines) + "\n", encoding="utf-8")


-def write_llamacpp_models_preset(path: Path, preset: str) -> None:
-    key = (preset or "recommended").strip().lower()
-    if key not in LLAMACPP_PRESETS:
-        raise ValueError(f"пресет {preset!r}; варианты: {', '.join(LLAMACPP_PRESETS)}")
-    specs = LLAMACPP_PRESETS[key]
-    lines = [
-        "# Локальный манифест llama.cpp GGUF (не коммить). Пример: llamacpp-models.example.yaml",
-        "# url = прямой HTTPS на .gguf; mmproj = projector для vision (Qwen2.5-VL и т.п.).",
-        "models:",
-    ]
-    if not specs:
-        lines.append("  []")
-    else:
-        for i, spec in enumerate(specs):
-            url = str(spec.get("url") or "").strip()
-            if not url:
-                continue
-            lines.append(f"  - url: {url}")
-            mmproj = str(spec.get("mmproj") or "").strip()
-            if mmproj:
-                lines.append(f"    mmproj: {mmproj}")
-            if i == 0:
-                lines.append("    default: true")
-    path.write_text("\n".join(lines) + "\n", encoding="utf-8")
-
-
 def ensure_ollama_manifest_from_example() -> Path:
    dest = ollama_models_manifest_path()
    if dest.is_file():
@@ -327,24 +161,10 @@ def ensure_ollama_manifest_from_example() -> Path:
    return dest


-def ensure_llamacpp_manifest_from_example() -> Path:
-    dest = llamacpp_models_manifest_path()
-    if dest.is_file():
-        return dest
-    example = llamacpp_models_example_path()
-    if example.is_file():
-        dest.write_text(example.read_text(encoding="utf-8"), encoding="utf-8")
-    else:
-        write_llamacpp_models_preset(dest, "recommended")
-    return dest
-
-
 def llm_local_port(cfg: Any) -> int | None:
    runtime = normalize_runtime(getattr(cfg, "llm_runtime", "none"))
    if runtime == "ollama":
        return int(getattr(cfg, "ollama_local_port", 17811))
-    if runtime == "llamacpp":
-        return int(getattr(cfg, "llamacpp_local_port", 17812))
    return None


@@ -352,47 +172,9 @@ def llm_remote_port(runtime: str) -> int | None:
    runtime = normalize_runtime(runtime)
    if runtime == "ollama":
        return 11434
-    if runtime == "llamacpp":
-        return 8080
    return None


-def pick_llamacpp_linux_asset_url(assets: list[dict[str, Any]]) -> str:
-    """Choose a Linux llama.cpp release asset URL.
-
-    Upstream ships Windows CUDA zips first; never pick win/macos/cudart-only.
-    Prefer ubuntu+cuda → linux+cuda → ubuntu vulkan x64 → ubuntu x64 CPU.
-    Mirrored in remote/install_llamacpp.sh (pick_linux_asset_url).
-    """
-    cands: list[tuple[int, str]] = []
-    for a in assets:
-        name = str(a.get("name") or "").lower()
-        url = str(a.get("browser_download_url") or "")
-        if not (url.endswith(".zip") or url.endswith(".tar.gz")):
-            continue
-        if any(x in name for x in ("win", "macos", "android", "darwin", "xcframework", "-ui.")):
-            continue
-        if "cudart" in name:
-            continue
-        score = 0
-        if "ubuntu" in name and "x64" in name and "cuda" in name:
-            score = 100
-        elif "linux" in name and "cuda" in name:
-            score = 90
-        elif "ubuntu" in name and "vulkan" in name and "x64" in name:
-            score = 50
-        elif "ubuntu" in name and "x64" in name and not any(
-            x in name for x in ("sycl", "openvino", "arm", "s390", "rocm")
-        ):
-            score = 30
-        elif "ubuntu" in name or "linux" in name:
-            score = 10
-        if score:
-            cands.append((score, url))
-    cands.sort(key=lambda t: t[0], reverse=True)
-    return cands[0][1] if cands else ""
-
-
 def append_vars_llm_runtime(vars_file: Path, runtime: str) -> None:
    from gpu_rent.varsfile import upsert_vars

@@ -45,7 +45,7 @@ class ModelEntry:
    url: str | None


-VALID_REQUIRES = frozenset({"none", "ollama", "llamacpp", "any-llm"})
+VALID_REQUIRES = frozenset({"none", "ollama", "any-llm"})


@dataclass
@@ -107,7 +107,7 @@ def normalize_requires(value: object | None) -> str:
    if raw in VALID_REQUIRES:
        return raw
    raise ConfigError(
-        f"requires={value!r}: жду none|ollama|llamacpp|any-llm"
+        f"requires={value!r}: жду none|ollama|any-llm"
    )


@@ -118,7 +118,7 @@ def repo_matches_runtime(repo: GitRepo, llm_runtime: str) -> bool:
    if req == "none":
        return True
    if req == "any-llm":
-        return runtime in {"ollama", "llamacpp"}
+        return runtime == "ollama"
    return runtime == req


@@ -56,14 +56,6 @@ def ollama_models_example_path() -> Path:
    return app_root() / "ollama-models.example.yaml"


-def llamacpp_models_manifest_path() -> Path:
-    return app_root() / "llamacpp-models.yaml"
-
-
-def llamacpp_models_example_path() -> Path:
-    return app_root() / "llamacpp-models.example.yaml"
-
-
 def extensions_manifest_path() -> Path:
    return app_root() / "extensions.yaml"

@@ -31,19 +31,6 @@ DATA = "/mnt/swarm_data"

 # Forwarded to remote install_*.sh (from .env / gpu-rent.vars → os.environ).
 _OLLAMA_INSTALL_ENV = ("OLLAMA_VERSION", "OLLAMA_SHA256")
-_LLAMACPP_INSTALL_ENV = (
-    "LLAMACPP_TAG",
-    "LLAMACPP_ASSET_URL",
-    "LLAMACPP_SHA256",
-    "LLAMACPP_BUILD_CUDA",
-    "LLAMACPP_BACKEND",
-    "LLAMACPP_FORCE_REINSTALL",
-    "LLAMACPP_NGL",
-    "LLAMACPP_CTX",
-    "LLAMACPP_HOST",
-    "LLAMACPP_PORT",
-    "LLAMACPP_EXTRA_ARGS",
-)


 def _remote_llm_env(cfg: Config, *keys: str) -> dict[str, str]:
@@ -450,7 +437,7 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
                check=False,
            )

-    # Always drop the other runtime so VRAM is not held by a leftover unit.
+    # Drop LLM units that should not hold VRAM for this runtime.
    if runtime == "none":
        log("LLM: none — останавливаю gpu-rent-ollama / gpu-rent-llamacpp если были")
        _stop_units("gpu-rent-ollama", "gpu-rent-llamacpp")
@@ -490,71 +477,8 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
                timeout=7200,
                log=log,
            )
-    elif runtime == "llamacpp":
-        _stop_units("gpu-rent-ollama")
-        from gpu_rent.llm_runtime import (
-            gguf_filename_from_url,
-            parse_llamacpp_models,
-        )
-
-        entries = parse_llamacpp_models(cfg.llamacpp_models_manifest)
-        defaults = [e for e in entries if e.default]
-        if defaults:
-            log(
-                "llama.cpp preferred: "
-                f"{defaults[0].filename or gguf_filename_from_url(defaults[0].url)}"
-            )
-        if entries:
-            jobs = []
-            for e in entries:
-                jobs.append(
-                    {
-                        "url": e.url,
-                        "filename": e.filename or gguf_filename_from_url(e.url),
-                    }
-                )
-                if e.mmproj_url:
-                    jobs.append(
-                        {
-                            "url": e.mmproj_url,
-                            "filename": gguf_filename_from_url(e.mmproj_url),
-                        }
-                    )
-            put_text(
-                cfg, host, "/tmp/gpu-rent-llamacpp-models.json", json.dumps(jobs, indent=2)
-            )
-            hf = (cfg.hf_token or "").strip()
-            if hf:
-                put_text(cfg, host, "/tmp/gpu-rent-hf.token", hf + "\n", mode=0o600)
-            else:
-                log(
-                    "⚠ HF_TOKEN не задан — gated GGUF (abliterated и др.) часто дают 401. "
-                    "Добавь HF_TOKEN=hf_… в .env → https://huggingface.co/settings/tokens"
-                )
-            log(f"llama.cpp: скачиваю {len(jobs)} GGUF из манифеста")
-            run_python(
-                cfg,
-                host,
-                _pkg_text("llamacpp_fetch.py"),
-                remote_path="/tmp/gpu-rent-llamacpp_fetch.py",
-                timeout=7200,
-                log=log,
-            )
-        else:
-            log("llamacpp-models.yaml пуст — GGUF skip (положи вручную)")
-        log("LLM: ставим/запускаем llama.cpp server")
-        import os
-
-        # Vulkan finishes in seconds; CUDA compile needs up to ~15–20 min.
-        run_script_sudo(
-            cfg,
-            host,
-            _pkg_text("install_llamacpp.sh"),
-            remote_path="/tmp/gpu-rent-install_llamacpp.sh",
-            timeout=3600,
-            env=_remote_llm_env(cfg, *_LLAMACPP_INSTALL_ENV),
-            log=log,
-        )
+    else:
+        raise CloudError(f"неизвестный LLM_RUNTIME={runtime!r}")
    st = load_state()
    st.notes = dict(st.notes or {})
    st.notes["llm_runtime"] = runtime
@@ -614,7 +538,7 @@ def provision_vm(
    rt = normalize_runtime(cfg.llm_runtime)
    if not swarm and rt == "none":
        raise CloudError(
-            "llm-only: нужен LLM_RUNTIME=ollama|llamacpp (или --ollama / --llamacpp)"
+            "llm-only: нужен LLM_RUNTIME=ollama (или --ollama / --llm ollama)"
        )

    # Arm ASAP so mid-provision failures still leave auto-stop on the VM.
@@ -692,6 +616,4 @@ def provision_vm(
        log("SwarmUI слушает 127.0.0.1:7801 — gpu-rent tunnel")
    if rt == "ollama":
        log(f"Ollama API → localhost:{cfg.ollama_local_port} (туннель)")
-    elif rt == "llamacpp":
-        log(f"llama.cpp → localhost:{cfg.llamacpp_local_port} (туннель)")
    log("Hold killer: gpu-rent hold   |  Стоп GPU: gpu-rent stop")
@@ -95,7 +95,6 @@ def unit_active(name):
 checks = []
 want_swarm = WANT_SWARM
 want_ollama = WANT_OLLAMA
-want_llama = WANT_LLAMA

 if want_swarm:
    ok, detail = http_ok("http://127.0.0.1:7801/")
@@ -131,18 +130,6 @@ if want_ollama:
        "unit": unit_active("gpu-rent-ollama"),
    })

-if want_llama:
-    ok, detail = http_ok("http://127.0.0.1:8080/health")
-    if not ok:
-        ok2, d2 = http_ok("http://127.0.0.1:8080/v1/models")
-        ok, detail = ok2, d2
-    checks.append({
-        "name": "llamacpp",
-        "ok": ok,
-        "detail": detail,
-        "unit": unit_active("gpu-rent-llamacpp"),
-    })
-
 print(json.dumps({"checks": checks}, ensure_ascii=False))
 '''

@@ -192,18 +179,17 @@ def wait_backend_idle(
    )


-def _expected_services(cfg: Config) -> tuple[bool, bool, bool]:
+def _expected_services(cfg: Config) -> tuple[bool, bool]:
    swarm = bool(getattr(cfg, "enable_swarmui", True))
    rt = normalize_runtime(getattr(cfg, "llm_runtime", "none"))
-    return swarm, rt == "ollama", rt == "llamacpp"
+    return swarm, rt == "ollama"


 def _probe_vm_once(cfg: Config, host: str) -> list[ServiceCheck]:
-    want_swarm, want_ollama, want_llama = _expected_services(cfg)
+    want_swarm, want_ollama = _expected_services(cfg)
    script = (
        _REMOTE_STACK_PROBE.replace("WANT_SWARM", "True" if want_swarm else "False")
        .replace("WANT_OLLAMA", "True" if want_ollama else "False")
-        .replace("WANT_LLAMA", "True" if want_llama else "False")
    )
    out = run_ssh(
        cfg,
@@ -251,8 +237,8 @@ def verify_stack_on_vm(
    raise_on_fail: bool = True,
 ) -> list[ServiceCheck]:
    """Poll until every enabled service answers on the VM loopback."""
-    want_swarm, want_ollama, want_llama = _expected_services(cfg)
-    if not (want_swarm or want_ollama or want_llama):
+    want_swarm, want_ollama = _expected_services(cfg)
+    if not (want_swarm or want_ollama):
        log("проверка стека: нечего ждать (swarm off, LLM none)")
        return []

@@ -261,8 +247,6 @@ def verify_stack_on_vm(
        names.append("SwarmUI :7801")
    if want_ollama:
        names.append("Ollama :11434")
-    if want_llama:
-        names.append("llama.cpp :8080")
    log(f"проверка на VM: {', '.join(names)}")

    deadline = time.time() + timeout
@@ -447,7 +431,7 @@ def verify_stack_local(
    raise_on_fail: bool = True,
 ) -> list[ServiceCheck]:
    """After tunnel: local ports + light HTTP for enabled services."""
-    want_swarm, want_ollama, want_llama = _expected_services(cfg)
+    want_swarm, want_ollama = _expected_services(cfg)
    targets: list[tuple[str, int, str | None]] = []
    if want_swarm:
        targets.append(("swarmui", int(cfg.swarmui_local_port), None))
@@ -459,9 +443,6 @@ def verify_stack_local(
                f"http://127.0.0.1:{cfg.ollama_local_port}/api/tags",
            )
        )
-    if want_llama:
-        p = int(cfg.llamacpp_local_port)
-        targets.append(("llamacpp", p, f"http://127.0.0.1:{p}/health"))

    if not targets:
        return []
@@ -481,8 +462,6 @@ def verify_stack_local(
                continue
            if url:
                ok, detail = _http_local(url)
-                if not ok and name == "llamacpp":
-                    ok, detail = _http_local(f"http://127.0.0.1:{port}/v1/models")
                last.append(ServiceCheck(name, ok, detail, "local"))
            else:
                ok, detail = _http_local(f"http://127.0.0.1:{port}/")
@@ -108,8 +108,7 @@ mkdir -p \
  "${DATA_ROOT}/Extensions" \
  "${DATA_ROOT}/DLNodes" \
  "${DATA_ROOT}/CustomWorkflows" \
-  "${DATA_ROOT}/ollama" \
-  "${DATA_ROOT}/llamacpp/models"
+  "${DATA_ROOT}/ollama"

 # LLM-only: data disk + tools, no SwarmUI clone / unit.
 if [[ "${GPU_RENT_SKIP_SWARMUI:-0}" == "1" ]]; then
@@ -107,7 +107,7 @@ def swarm_busy(swarm_url: str, timeout: float = 8.0) -> tuple[bool, str]:


 def llm_busy(timeout: float = 3.0) -> tuple[bool, str]:
-    """Ollama pull / loaded models or llama.cpp with a model count as busy."""
+    """Ollama pull / loaded models count as busy."""
    pull_marker = DATA / ".gpu-rent-ollama-pulling"
    if pull_marker.is_file():
        try:
@@ -138,24 +138,6 @@ def llm_busy(timeout: float = 3.0) -> tuple[bool, str]:
            return True, f"ollama running {names}"
    except (urllib.error.URLError, urllib.error.HTTPError, TimeoutError, json.JSONDecodeError, OSError):
        pass
-    # llama.cpp: slots in use
-    try:
-        req = urllib.request.Request("http://127.0.0.1:8080/health", method="GET")
-        with urllib.request.urlopen(req, timeout=timeout, context=ctx) as resp:
-            if getattr(resp, "status", 200) == 200:
-                try:
-                    req2 = urllib.request.Request("http://127.0.0.1:8080/props", method="GET")
-                    with urllib.request.urlopen(req2, timeout=timeout, context=ctx) as resp2:
-                        props = json.loads(resp2.read().decode("utf-8"))
-                    total = int(props.get("total_slots") or 0)
-                    avail = int(props.get("available_slots") or total)
-                    in_use = total - avail if total else 0
-                    if in_use > 0:
-                        return True, f"llamacpp slots_in_use={in_use}"
-                except Exception:
-                    pass
-    except (urllib.error.URLError, urllib.error.HTTPError, TimeoutError, OSError):
-        pass
    return False, "llm idle"


@@ -1,471 +0,0 @@
-#!/usr/bin/env bash
-# Install llama-server for OpenAI-compatible API on loopback :8080.
-#
-# Default: official Linux release asset (Ubuntu Vulkan — GPU without compile).
-# CUDA source build only as last resort (or LLAMACPP_BUILD_CUDA=1).
-# Pin: LLAMACPP_TAG=b10545  LLAMACPP_ASSET_URL=...  LLAMACPP_SHA256=...
-set -euo pipefail
-
-SWARM_USER="${SWARM_USER:-ubuntu}"
-DATA_ROOT="/mnt/swarm_data"
-LLAMA_ROOT="${DATA_ROOT}/llamacpp"
-MODELS_DIR="${LLAMA_ROOT}/models"
-BIN_DIR="${LLAMA_ROOT}/bin"
-SRC_DIR="${LLAMA_ROOT}/src"
-STAMP="${BIN_DIR}/.build-id"
-UNIT="gpu-rent-llamacpp"
-REPO="https://github.com/ggml-org/llama.cpp.git"
-API_BASE="https://api.github.com/repos/ggml-org/llama.cpp"
-
-log() { echo "[gpu-rent-llamacpp] $*" >&2; }
-
-if [[ "$(id -u)" -ne 0 ]]; then
-  echo "нужен root" >&2
-  exit 1
-fi
-
-mkdir -p "$MODELS_DIR" "$BIN_DIR"
-chown -R "${SWARM_USER}:${SWARM_USER}" "$LLAMA_ROOT"
-
-SERVER_BIN="${BIN_DIR}/llama-server"
-LLAMACPP_TAG="${LLAMACPP_TAG:-}"
-LLAMACPP_ASSET_URL="${LLAMACPP_ASSET_URL:-}"
-LLAMACPP_SHA256="${LLAMACPP_SHA256:-}"
-# 1 = force CUDA compile; 0 = never compile (Vulkan/CPU prebuilt only)
-LLAMACPP_BUILD_CUDA="${LLAMACPP_BUILD_CUDA:-}"
-# auto | cuda | vulkan — default auto: CUDA if nvcc already on VM, else Vulkan prebuilt
-LLAMACPP_BACKEND="${LLAMACPP_BACKEND:-auto}"
-LLAMACPP_FORCE_REINSTALL="${LLAMACPP_FORCE_REINSTALL:-}"
-LLAMACPP_NGL="${LLAMACPP_NGL:-}"
-LLAMACPP_CTX="${LLAMACPP_CTX:-}"
-LLAMACPP_HOST="${LLAMACPP_HOST:-127.0.0.1}"
-LLAMACPP_PORT="${LLAMACPP_PORT:-8080}"
-LLAMACPP_EXTRA_ARGS="${LLAMACPP_EXTRA_ARGS:-}"
-
-have_nvcc() {
-  if command -v nvcc >/dev/null 2>&1; then
-    return 0
-  fi
-  if [[ -x /usr/local/cuda/bin/nvcc ]]; then
-    export PATH="/usr/local/cuda/bin:${PATH}"
-    return 0
-  fi
-  return 1
-}
-
-# Prefer CUDA when toolkit already present (GPU images / prior up). Vulkan = fast no-compile.
-want_cuda_build() {
-  case "${LLAMACPP_BUILD_CUDA}" in
-    1|yes|true) return 0 ;;
-    0|no|false) return 1 ;;
-  esac
-  case "${LLAMACPP_BACKEND}" in
-    cuda) return 0 ;;
-    vulkan) return 1 ;;
-    *)
-      if have_nvcc; then
-        return 0
-      fi
-      return 1
-      ;;
-  esac
-}
-
-if [[ "${LLAMACPP_FORCE_REINSTALL}" == "1" ]]; then
-  log "LLAMACPP_FORCE_REINSTALL=1 — удаляю старый бинарь"
-  rm -f "$SERVER_BIN" "$STAMP"
-fi
-
-# Upgrade path: previous default was Vulkan prebuilt; if nvcc is here, prefer CUDA.
-if [[ -x "$SERVER_BIN" && -f "$STAMP" && "${LLAMACPP_BACKEND}" != "vulkan" && "${LLAMACPP_BUILD_CUDA}" != "0" ]]; then
-  if grep -q '^asset:' "$STAMP" 2>/dev/null && want_cuda_build; then
-    log "был Vulkan/CPU prebuilt, nvcc есть — пересобираю CUDA (лучше на NVIDIA)"
-    rm -f "$SERVER_BIN" "$STAMP"
-  fi
-fi
-
-# Prefer ubuntu CUDA (rare) → vulkan → cpu. Never Windows/macOS/cudart-only.
-pick_linux_asset_url() {
-  python3 -c '
-import json,sys
-data=json.load(sys.stdin)
-assets=data.get("assets") or []
-cands=[]
-for a in assets:
-    n=(a.get("name") or "").lower()
-    u=a.get("browser_download_url") or ""
-    if not (u.endswith(".zip") or u.endswith(".tar.gz")):
-        continue
-    if any(x in n for x in ("win","macos","android","darwin","xcframework","-ui.")):
-        continue
-    if "cudart" in n:
-        continue
-    score=0
-    if "ubuntu" in n and "x64" in n and "cuda" in n:
-        score=100
-    elif "linux" in n and "cuda" in n:
-        score=90
-    elif "ubuntu" in n and "vulkan" in n and "x64" in n:
-        score=50
-    elif "ubuntu" in n and "x64" in n and not any(
-        x in n for x in ("sycl","openvino","arm","s390","rocm")
-    ):
-        score=30
-    elif "ubuntu" in n or "linux" in n:
-        score=10
-    if score:
-        cands.append((score, u, n))
-cands.sort(reverse=True)
-print(cands[0][1] if cands else "")
-'
-}
-
-resolve_release_tag() {
-  # stdout = tag only (no log lines — callers capture via $())
-  if [[ -n "$LLAMACPP_TAG" ]]; then
-    echo "$LLAMACPP_TAG"
-    return
-  fi
-  curl -fsSL "${API_BASE}/releases/latest" | python3 -c \
-    'import json,sys; print(json.load(sys.stdin).get("tag_name") or "")'
-}
-
-cuda_architectures() {
-  python3 - <<'PY'
-import json
-from pathlib import Path
-p = Path("/mnt/swarm_data/.gpu-rent-gpu.json")
-cap = "8.9"
-if p.is_file():
-    try:
-        cap = str(json.loads(p.read_text()).get("compute_cap") or cap)
-    except Exception:
-        pass
-parts = cap.split(".")
-try:
-    maj, mnr = int(parts[0]), int(parts[1]) if len(parts) > 1 else 0
-    print(f"{maj}{mnr}")
-except ValueError:
-    print("89")
-PY
-}
-
-ensure_build_deps() {
-  export DEBIAN_FRONTEND=noninteractive
-  apt-get install -y -qq \
-    cmake build-essential git curl ca-certificates \
-    libcurl4-openssl-dev >/dev/null
-  if command -v nvcc >/dev/null 2>&1; then
-    return 0
-  fi
-  if [[ -x /usr/local/cuda/bin/nvcc ]]; then
-    export PATH="/usr/local/cuda/bin:${PATH}"
-    return 0
-  fi
-  log "ставлю nvidia-cuda-toolkit (нужен nvcc)…"
-  apt-get install -y -qq nvidia-cuda-toolkit >/dev/null
-  if command -v nvcc >/dev/null 2>&1; then
-    return 0
-  fi
-  if [[ -x /usr/local/cuda/bin/nvcc ]]; then
-    export PATH="/usr/local/cuda/bin:${PATH}"
-    return 0
-  fi
-  return 1
-}
-
-ensure_vulkan_runtime() {
-  if ldconfig -p 2>/dev/null | grep -q 'libvulkan\.so'; then
-    return 0
-  fi
-  export DEBIAN_FRONTEND=noninteractive
-  log "ставлю libvulkan1 (для Ubuntu Vulkan prebuilt)…"
-  apt-get install -y -qq libvulkan1 mesa-vulkan-drivers >/dev/null 2>&1 || \
-    apt-get install -y -qq libvulkan1 >/dev/null 2>&1 || true
-}
-
-install_from_archive_url() {
-  local url="$1"
-  local tmp kind
-  tmp="$(mktemp -d)"
-  (
-    cd "$tmp"
-    log "скачиваю prebuilt: $url"
-    curl -fL --progress-bar "$url" -o pkg.bin
-    if [[ -n "$LLAMACPP_SHA256" ]]; then
-      echo "${LLAMACPP_SHA256}  pkg.bin" | sha256sum -c -
-    else
-      log "WARN: LLAMACPP_SHA256 не задан — checksum skip"
-    fi
-    mkdir -p out
-    # Do NOT grep -i zip — that matches "gzip" and breaks .tar.gz.
-    kind="$(file -b pkg.bin 2>/dev/null || true)"
-    case "$url" in
-      *.zip)
-        apt-get install -y -qq unzip >/dev/null 2>&1 || true
-        unzip -qo pkg.bin -d out
-        ;;
-      *)
-        if [[ "$kind" == Zip\ archive* ]] || [[ "$kind" == *"Zip archive"* ]]; then
-          apt-get install -y -qq unzip >/dev/null 2>&1 || true
-          unzip -qo pkg.bin -d out
-        else
-          tar -xaf pkg.bin -C out 2>/dev/null \
-            || tar -xzf pkg.bin -C out 2>/dev/null \
-            || tar -xf pkg.bin -C out
-        fi
-        ;;
-    esac
-    local found
-    found="$(find out -type f -name 'llama-server' | head -n1 || true)"
-    if [[ -z "$found" ]]; then
-      found="$(find out -type f -name 'server' | head -n1 || true)"
-    fi
-    if [[ -z "$found" ]]; then
-      log "в архиве нет llama-server (file says: ${kind:-unknown})"
-      exit 1
-    fi
-    install -m 755 "$found" "$SERVER_BIN"
-    # Shared libs next to binary (release tarballs ship .so alongside).
-    find out -type f \( -name '*.so' -o -name '*.so.*' \) -print0 2>/dev/null \
-      | while IFS= read -r -d '' so; do
-          install -m 755 "$so" "${BIN_DIR}/$(basename "$so")"
-        done
-    chown -R "${SWARM_USER}:${SWARM_USER}" "$BIN_DIR"
-  )
-  local rc=$?
-  rm -rf "$tmp"
-  return "$rc"
-}
-
-install_linux_release() {
-  local tag="$1"
-  local api url
-  api="${API_BASE}/releases/tags/${tag}"
-  url="$(curl -fsSL "$api" | pick_linux_asset_url)"
-  if [[ -z "$url" ]]; then
-    log "в release ${tag} нет Linux-ассета"
-    return 1
-  fi
-  if [[ "$url" == *vulkan* ]]; then
-    log "беру Ubuntu Vulkan prebuilt (GPU без compile; CUDA-сборка — LLAMACPP_BUILD_CUDA=1)"
-    ensure_vulkan_runtime
-  elif [[ "$url" == *cuda* ]]; then
-    log "беру Linux CUDA prebuilt"
-  else
-    log "WARN: Linux prebuilt без GPU backend (CPU) — ${url##*/}"
-  fi
-  if ! install_from_archive_url "$url"; then
-    return 1
-  fi
-  echo "asset:${tag}" >"$STAMP"
-  chown "${SWARM_USER}:${SWARM_USER}" "$STAMP"
-}
-
-# Quiet CUDA build: no cmake spam; heartbeat every 30s with last %.
-build_cuda_from_source() {
-  local tag="$1"
-  local arch build_log pid pct line
-  arch="$(cuda_architectures)"
-  build_log="${LLAMA_ROOT}/build-cuda.log"
-  log "крайний случай: сборка CUDA из исходников (tag=${tag}, arch=${arch}, 5–15 мин)…"
-  log "полный лог: ${build_log}"
-  if ! ensure_build_deps; then
-    log "нет nvcc — CUDA-сборку пропускаем"
-    return 1
-  fi
-  log "nvcc $(nvcc --version 2>/dev/null | tail -n1 || echo '?')"
-
-  mkdir -p "$SRC_DIR"
-  export GIT_TERMINAL_PROMPT=0
-  if [[ -d "${SRC_DIR}/.git" ]]; then
-    git -C "$SRC_DIR" -c advice.detachedHead=false fetch --depth 1 origin tag "$tag" 2>>"$build_log" || true
-    if ! git -C "$SRC_DIR" -c advice.detachedHead=false checkout -f "$tag" >>"$build_log" 2>&1; then
-      rm -rf "$SRC_DIR"
-      git -c advice.detachedHead=false clone --depth 1 --branch "$tag" "$REPO" "$SRC_DIR" >>"$build_log" 2>&1
-    fi
-  else
-    rm -rf "$SRC_DIR"
-    git -c advice.detachedHead=false clone --depth 1 --branch "$tag" "$REPO" "$SRC_DIR" >>"$build_log" 2>&1
-  fi
-
-  cmake -S "$SRC_DIR" -B "${SRC_DIR}/build" \
-    -DCMAKE_BUILD_TYPE=Release \
-    -DGGML_CUDA=ON \
-    -DCMAKE_CUDA_ARCHITECTURES="${arch}" \
-    -DLLAMA_BUILD_SERVER=ON \
-    -DLLAMA_BUILD_UI=OFF \
-    -DLLAMA_USE_PREBUILT_UI=OFF \
-    -DGGML_CCACHE=OFF \
-    >>"$build_log" 2>&1
-
-  # Background build + heartbeat (keeps SSH stream alive without 200 cmake lines).
-  cmake --build "${SRC_DIR}/build" -j"$(nproc)" --target llama-server \
-    >>"$build_log" 2>&1 &
-  pid=$!
-  while kill -0 "$pid" 2>/dev/null; do
-    pct="$(grep -oE '\[[[:space:]]*[0-9]+%\]' "$build_log" 2>/dev/null | tail -n1 || true)"
-    line="$(grep -E 'Building CUDA|Built target|Linking' "$build_log" 2>/dev/null | tail -n1 || true)"
-    if [[ -n "$pct" ]]; then
-      log "сборка CUDA ещё идёт… ${pct}${line:+ · ${line}}"
-    else
-      log "сборка CUDA ещё идёт… (cmake/nvcc, см. build.log)"
-    fi
-    sleep 30
-  done
-  if ! wait "$pid"; then
-    log "сборка упала — хвост ${build_log}:"
-    tail -n 40 "$build_log" >&2 || true
-    return 1
-  fi
-
-  local built="${SRC_DIR}/build/bin/llama-server"
-  if [[ ! -x "$built" ]]; then
-    log "сборка не дала ${built}"
-    return 1
-  fi
-  install -m 755 "$built" "$SERVER_BIN"
-  # CUDA build may need libs from build/bin
-  find "${SRC_DIR}/build/bin" -maxdepth 1 -type f \( -name '*.so' -o -name '*.so.*' \) -print0 2>/dev/null \
-    | while IFS= read -r -d '' so; do
-        install -m 755 "$so" "${BIN_DIR}/$(basename "$so")"
-      done
-  chown -R "${SWARM_USER}:${SWARM_USER}" "$BIN_DIR"
-  echo "cuda:${tag}:${arch}" >"$STAMP"
-  chown "${SWARM_USER}:${SWARM_USER}" "$STAMP"
-  log "CUDA binary → ${SERVER_BIN}"
-  return 0
-}
-
-normalize_tag() {
-  printf '%s' "$1" | tr -d '\r' | head -n1 | awk 'NF{print; exit}'
-}
-
-if [[ -x "$SERVER_BIN" ]]; then
-  log "llama-server уже есть: ${SERVER_BIN}"
-else
-  if [[ -n "$LLAMACPP_ASSET_URL" ]]; then
-    log "скачиваю по LLAMACPP_ASSET_URL…"
-    install_from_archive_url "$LLAMACPP_ASSET_URL"
-    echo "asset-url" >"$STAMP"
-    chown "${SWARM_USER}:${SWARM_USER}" "$STAMP"
-  else
-    if [[ -z "$LLAMACPP_TAG" ]]; then
-      log "WARN: LLAMACPP_TAG не задан — latest (см. docs/llm.md)"
-    fi
-    tag="$(normalize_tag "$(resolve_release_tag)")"
-    if [[ -z "$tag" || "$tag" == *" "* || "$tag" == *"["* ]]; then
-      log "не удалось определить release tag (got: ${tag:-empty})"
-      exit 1
-    fi
-
-    installed=0
-    if want_cuda_build; then
-      log "backend: CUDA (nvcc есть или LLAMACPP_BACKEND/BUILD_CUDA) — сборка, Vulkan только если упадёт"
-      if build_cuda_from_source "$tag"; then
-        installed=1
-      else
-        log "CUDA-сборка не вышла — fallback на Linux prebuilt (Vulkan/CPU)"
-        if install_linux_release "$tag"; then
-          installed=1
-        fi
-      fi
-    else
-      log "backend: Linux prebuilt (нет nvcc / LLAMACPP_BACKEND=vulkan) — без compile"
-      if install_linux_release "$tag"; then
-        installed=1
-      else
-        log "prebuilt не вышел — крайний случай: CUDA из исходников"
-        if build_cuda_from_source "$tag"; then
-          installed=1
-        fi
-      fi
-    fi
-    if [[ "$installed" != "1" ]]; then
-      log "не удалось поставить llama-server"
-      exit 1
-    fi
-  fi
-fi
-
-if [[ ! -x "$SERVER_BIN" ]]; then
-  log "нет исполняемого ${SERVER_BIN}"
-  exit 1
-fi
-
-# Prefer a weights GGUF (skip mmproj), then attach --mmproj if present.
-MODEL_ARG=""
-MMPROJ_ARG=""
-FIRST_GGUF="$(
-  find "$MODELS_DIR" -type f \( -name '*.gguf' -o -name '*.GGUF' \) \
-    ! -iname '*mmproj*' 2>/dev/null | head -n1 || true
-)"
-MMPROJ_GGUF="$(
-  find "$MODELS_DIR" -type f \( -iname '*mmproj*.gguf' -o -iname '*mmproj*.GGUF' \) \
-    2>/dev/null | head -n1 || true
-)"
-if [[ -n "$FIRST_GGUF" ]]; then
-  MODEL_ARG="-m ${FIRST_GGUF}"
-  log "модель ${FIRST_GGUF}"
-else
-  log "нет GGUF в ${MODELS_DIR} — положи файл вручную и systemctl restart ${UNIT}"
-fi
-if [[ -n "$MMPROJ_GGUF" ]]; then
-  MMPROJ_ARG="--mmproj ${MMPROJ_GGUF}"
-  log "mmproj ${MMPROJ_GGUF}"
-fi
-
-# GPU layers: share card with Swarm — full offload on mid+, leave headroom on low.
-NGL=99
-CTX=8192
-if [[ -f "${DATA_ROOT}/.gpu-rent-gpu.json" ]]; then
-  eval "$(python3 - <<'PY'
-import json
-from pathlib import Path
-gpu=json.loads(Path("/mnt/swarm_data/.gpu-rent-gpu.json").read_text())
-vram=int(gpu.get("vram_mib") or 0)
-gib=vram/1024.0
-if gib < 16:
-    print("NGL=40"); print("CTX=4096")
-elif gib < 24:
-    print("NGL=99"); print("CTX=8192")
-elif gib < 48:
-    print("NGL=99"); print("CTX=16384")
-else:
-    print("NGL=99"); print("CTX=32768")
-PY
-)" || true
-fi
-# Explicit overrides from gpu-rent.vars / .env (forwarded by provision).
-if [[ -n "$LLAMACPP_NGL" ]]; then
-  NGL="$LLAMACPP_NGL"
-fi
-if [[ -n "$LLAMACPP_CTX" ]]; then
-  CTX="$LLAMACPP_CTX"
-fi
-log "llama.cpp -ngl ${NGL} -c ${CTX} host=${LLAMACPP_HOST} port=${LLAMACPP_PORT}${LLAMACPP_EXTRA_ARGS:+ extra=${LLAMACPP_EXTRA_ARGS}}"
-
-cat >/etc/systemd/system/${UNIT}.service <<EOF
-[Unit]
-Description=gpu-rent llama.cpp server (loopback, GPU-tuned)
-After=network-online.target local-fs.target
-Wants=network-online.target
-
-[Service]
-Type=simple
-User=${SWARM_USER}
-Group=${SWARM_USER}
-WorkingDirectory=${LLAMA_ROOT}
-Environment=LD_LIBRARY_PATH=${BIN_DIR}
-ExecStart=${SERVER_BIN} ${MODEL_ARG} ${MMPROJ_ARG} --host ${LLAMACPP_HOST} --port ${LLAMACPP_PORT} -ngl ${NGL} -c ${CTX} ${LLAMACPP_EXTRA_ARGS}
-Restart=on-failure
-RestartSec=8
-
-[Install]
-WantedBy=multi-user.target
-EOF
-
-systemctl daemon-reload
-systemctl enable "$UNIT"
-systemctl restart "$UNIT" || log "unit стартовал с ошибкой (часто нет GGUF) — проверь journalctl -u ${UNIT}"
-log "ok — http://127.0.0.1:8080 models=${MODELS_DIR}"
@@ -1,182 +0,0 @@
-#!/usr/bin/env python3
-"""Download GGUF files for llama.cpp from a JSON job list. Stdlib only."""
-from __future__ import annotations
-
-import json
-import os
-import sys
-import time
-import urllib.error
-import urllib.parse
-import urllib.request
-from pathlib import Path
-
-JOBS = Path("/tmp/gpu-rent-llamacpp-models.json")
-MODELS_DIR = Path("/mnt/swarm_data/llamacpp/models")
-TOKEN_FILE = Path("/tmp/gpu-rent-hf.token")
-
-
-def fmt_bytes(n: float) -> str:
-    n = float(n)
-    for unit, div in (("GB", 1024**3), ("MB", 1024**2), ("KB", 1024), ("B", 1)):
-        if n >= div or unit == "B":
-            if unit == "B":
-                return f"{int(n)}B"
-            return f"{n / div:.1f}{unit}"
-    return f"{n:.0f}B"
-
-
-def progress_line(
-    label: str,
-    done: int,
-    total: int | None,
-    speed: float,
-    *,
-    width: int = 22,
-) -> str:
-    if total and total > 0:
-        pct = min(100.0, 100.0 * done / total)
-        filled = int(width * done / total)
-        filled = min(width, max(0, filled))
-        bar = "#" * filled + "-" * (width - filled)
-        return (
-            f"{label} [{bar}] {pct:5.1f}%  "
-            f"{fmt_bytes(done)}/{fmt_bytes(total)}  {fmt_bytes(speed)}/s"
-        )
-    return f"{label} {fmt_bytes(done)}  {fmt_bytes(speed)}/s"
-
-
-class DownloadProgress:
-    def __init__(self, label: str, total: int | None) -> None:
-        self.label = label
-        self.total = total if total and total > 0 else None
-        self.done = 0
-        self.t0 = time.monotonic()
-        self.last_print = 0.0
-
-    def add(self, n: int) -> None:
-        self.done += n
-        now = time.monotonic()
-        if now - self.last_print < 1.0 and not (
-            self.total is not None and self.done >= self.total
-        ):
-            return
-        self.last_print = now
-        self._emit()
-
-    def finish(self) -> None:
-        self._emit(final=True)
-
-    def _emit(self, *, final: bool = False) -> None:
-        elapsed = max(time.monotonic() - self.t0, 0.001)
-        line = progress_line(self.label, self.done, self.total, self.done / elapsed)
-        if final:
-            print(line, flush=True)
-        else:
-            print(line, end="\r", flush=True)
-
-
-def download(url: str, dest: Path, headers: dict[str, str], *, label: str) -> None:
-    partial = dest.with_suffix(dest.suffix + ".partial")
-
-    class StripAuthRedirect(urllib.request.HTTPRedirectHandler):
-        def redirect_request(self, req, fp, code, msg, headers_resp, newurl):
-            new = urllib.request.HTTPRedirectHandler.redirect_request(
-                self, req, fp, code, msg, headers_resp, newurl
-            )
-            if new is None:
-                return None
-            host = (urllib.parse.urlparse(new.full_url).hostname or "").lower()
-            # Hub needs Bearer; CDN (cdn-lfs.*) is pre-signed — drop Authorization.
-            if host in {"huggingface.co", "hf.co"}:
-                return new
-            return urllib.request.Request(
-                new.full_url, headers={"User-Agent": headers.get("User-Agent", "gpu-rent/1")}
-            )
-
-    opener = urllib.request.build_opener(StripAuthRedirect)
-    req = urllib.request.Request(url, headers=headers)
-    with opener.open(req, timeout=600) as resp, partial.open("wb") as out:
-        cl = resp.headers.get("Content-Length")
-        try:
-            total_n = int(cl) if cl else None
-        except ValueError:
-            total_n = None
-        prog = DownloadProgress(label, total_n)
-        while True:
-            chunk = resp.read(1024 * 1024)
-            if not chunk:
-                break
-            out.write(chunk)
-            prog.add(len(chunk))
-        prog.finish()
-    partial.replace(dest)
-
-
-def main() -> int:
-    if TOKEN_FILE.is_file():
-        try:
-            os.environ["HF_TOKEN"] = TOKEN_FILE.read_text(encoding="utf-8").strip()
-        finally:
-            try:
-                TOKEN_FILE.unlink(missing_ok=True)
-            except OSError:
-                pass
-    if not JOBS.is_file():
-        print("no jobs file")
-        return 1
-    jobs = json.loads(JOBS.read_text(encoding="utf-8"))
-    if not isinstance(jobs, list) or not jobs:
-        print("llamacpp fetch: пустой список — skip")
-        return 0
-    MODELS_DIR.mkdir(parents=True, exist_ok=True)
-    token = (os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN") or "").strip()
-    failed = 0
-    for i, job in enumerate(jobs, 1):
-        if not isinstance(job, dict):
-            continue
-        url = str(job.get("url") or "").strip()
-        name = str(job.get("filename") or "").strip()
-        if not url:
-            continue
-        if not name:
-            name = url.rstrip("/").rsplit("/", 1)[-1] or "model.gguf"
-        dest = MODELS_DIR / name
-        prefix = f"[{i}/{len(jobs)}]"
-        if dest.is_file() and dest.stat().st_size > 1_000_000:
-            print(f"{prefix} уже есть {name} ({fmt_bytes(dest.stat().st_size)})")
-            continue
-        print(f"{prefix} качаю {name}", flush=True)
-        headers = {"User-Agent": "gpu-rent/1"}
-        if token:
-            headers["Authorization"] = f"Bearer {token}"
-        try:
-            download(url, dest, headers, label=f"{prefix} {name}")
-            print(f"{prefix} ok {name} ({fmt_bytes(dest.stat().st_size)})")
-        except (urllib.error.URLError, urllib.error.HTTPError, OSError, TimeoutError) as exc:
-            failed += 1
-            msg = str(exc)
-            if "401" in msg or "403" in msg:
-                if not token:
-                    msg += (
-                        " — нет HF_TOKEN: добавь в .env "
-                        "(https://huggingface.co/settings/tokens) и прими условия репо"
-                    )
-                else:
-                    msg += (
-                        " — токен есть, но отказано: проверь scopes / "
-                        "Accept license на странице модели"
-                    )
-            print(f"FAIL {name}: {msg}")
-            try:
-                dest.with_suffix(dest.suffix + ".partial").unlink(missing_ok=True)
-            except OSError:
-                pass
-    if failed:
-        return 1
-    print("llamacpp fetch ok")
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main())
@@ -8,21 +8,16 @@ from pathlib import Path

 from gpu_rent.llm_runtime import (
    append_vars_llm_runtime,
-    ensure_llamacpp_manifest_from_example,
    ensure_ollama_manifest_from_example,
-    llamacpp_preset_menu,
    llm_runtime_menu,
    normalize_runtime,
    ollama_preset_menu,
-    write_llamacpp_models_preset,
    write_ollama_models_preset,
 )
 from gpu_rent.paths import (
    app_root,
    env_path,
    extensions_manifest_path,
-    llamacpp_models_example_path,
-    llamacpp_models_manifest_path,
    models_manifest_path,
    ollama_models_example_path,
    ollama_models_manifest_path,
@@ -66,12 +61,6 @@ def run_setup(
    _copy_if_missing(
        ollama_models_example_path(), ollama_models_manifest_path(), "ollama-models.yaml", log
    )
-    _copy_if_missing(
-        llamacpp_models_example_path(),
-        llamacpp_models_manifest_path(),
-        "llamacpp-models.yaml",
-        log,
-    )

    runtime = llm
    if runtime is None:
@@ -107,24 +96,6 @@ def run_setup(
        else:
            write_ollama_models_preset(ollama_models_manifest_path(), preset)
            log(f"ollama-models.yaml пресет={preset}")
-    elif runtime == "llamacpp":
-        preset = ollama_preset
-        if preset is None and ask:
-            preset = prompt_menu(
-                "llama.cpp GGUF",
-                llamacpp_preset_menu(include_keep=False),
-                default="recommended",
-                ask=ask,
-                show=log,
-            )
-        if preset is None:
-            preset = "recommended"
-        if preset.strip().lower() in {"keep", "example", ""}:
-            ensure_llamacpp_manifest_from_example()
-            log("llamacpp-models.yaml из example")
-        else:
-            write_llamacpp_models_preset(llamacpp_models_manifest_path(), preset)
-            log(f"llamacpp-models.yaml пресет={preset}")

    do_wd = install_watchdog
    if do_wd is None and confirm:
@@ -79,8 +79,6 @@ def tunnel_forwards(cfg: Config) -> list[tuple[int, int]]:
    runtime = normalize_runtime(cfg.llm_runtime)
    if runtime == "ollama":
        pairs.append((cfg.ollama_local_port, 11434))
-    elif runtime == "llamacpp":
-        pairs.append((cfg.llamacpp_local_port, 8080))
    if not pairs:
        # Failsafe: at least SwarmUI port so tunnel isn't empty.
        pairs.append((cfg.swarmui_local_port, 7801))
@@ -189,8 +187,6 @@ def run_tunnel(
        open_url = f"http://127.0.0.1:{cfg.swarmui_local_port}"
    elif runtime == "ollama":
        open_url = f"http://127.0.0.1:{cfg.ollama_local_port}"
-    elif runtime == "llamacpp":
-        open_url = f"http://127.0.0.1:{cfg.llamacpp_local_port}"
    else:
        open_url = f"http://127.0.0.1:{cfg.swarmui_local_port}"

@@ -5,7 +5,6 @@ class _Cfg:
    swarmui_local_port = 17801
    llm_runtime = "ollama"
    ollama_local_port = 17811
-    llamacpp_local_port = 17812
    enable_swarmui = True


@@ -1,31 +0,0 @@
-from pathlib import Path
-
-from gpu_rent.llm_runtime import (
-    gguf_filename_from_url,
-    parse_llamacpp_models,
-    write_llamacpp_models_preset,
-)
-
-
-def test_gguf_filename_from_url():
-    url = (
-        "https://huggingface.co/org/repo/resolve/main/"
-        "Qwen2.5-3B-Instruct-Q4_K_M.gguf"
-    )
-    assert gguf_filename_from_url(url) == "Qwen2.5-3B-Instruct-Q4_K_M.gguf"
-
-
-def test_write_and_parse_llamacpp_preset(tmp_path: Path):
-    path = tmp_path / "llamacpp-models.yaml"
-    write_llamacpp_models_preset(path, "light")
-    entries = parse_llamacpp_models(path)
-    assert len(entries) == 1
-    assert entries[0].default is True
-    assert "Qwen2.5-3B" in entries[0].url
-    assert entries[0].url.startswith("https://")
-
-
-def test_parse_llamacpp_empty(tmp_path: Path):
-    path = tmp_path / "llamacpp-models.yaml"
-    write_llamacpp_models_preset(path, "empty")
-    assert parse_llamacpp_models(path) == []
@@ -15,13 +15,12 @@ def test_parse_enable_swarmui_workload(monkeypatch):
 def test_tunnel_forwards_llm_only(monkeypatch):
    class Cfg:
        swarmui_local_port = 17801
-        llm_runtime = "llamacpp"
+        llm_runtime = "ollama"
        enable_swarmui = False
        ollama_local_port = 17811
-        llamacpp_local_port = 17812

    monkeypatch.setattr("gpu_rent.tunnel.load_state", lambda: type("S", (), {"notes": {}})())
-    assert tunnel_forwards(Cfg()) == [(17812, 8080)]
+    assert tunnel_forwards(Cfg()) == [(17811, 11434)]


 def test_access_links_llm_only(monkeypatch):
@@ -29,10 +28,9 @@ def test_access_links_llm_only(monkeypatch):

    class Cfg:
        swarmui_local_port = 17801
-        llm_runtime = "llamacpp"
+        llm_runtime = "ollama"
        enable_swarmui = False
        ollama_local_port = 17811
-        llamacpp_local_port = 17812

    monkeypatch.setattr(
        "gpu_rent.access_card.load_state",
@@ -40,5 +38,5 @@ def test_access_links_llm_only(monkeypatch):
    )
    labels = [x.label for x in collect_access_links(Cfg(), tunneled=True)]
    assert "SwarmUI UI" not in labels
-    assert "llama.cpp" in labels
+    assert "Ollama API" in labels
    assert mcp_snippet_lines(Cfg())[0].startswith("#")
@@ -5,10 +5,7 @@ import pytest
 from gpu_rent.llm_runtime import (
    decide_runtime,
    normalize_runtime,
-    parse_llamacpp_models,
    parse_ollama_models,
-    pick_llamacpp_linux_asset_url,
-    remap_llamacpp_url,
    write_ollama_models_preset,
 )

@@ -16,22 +13,20 @@ from gpu_rent.llm_runtime import (
 def test_normalize_runtime():
    assert normalize_runtime(None) == "none"
    assert normalize_runtime("OLLAMA") == "ollama"
-    assert normalize_runtime("llama-cpp") == "llamacpp"
+    with pytest.raises(ValueError):
+        normalize_runtime("llamacpp")
    with pytest.raises(ValueError):
        normalize_runtime("foo")


 def test_decide_runtime_flags_win():
    assert (
-        decide_runtime(flag=None, ollama_flag=True, llamacpp_flag=False, from_config="none")
-        == "ollama"
+        decide_runtime(flag=None, ollama_flag=True, from_config="none") == "ollama"
    )
    assert (
-        decide_runtime(flag="llamacpp", ollama_flag=False, llamacpp_flag=False, from_config="ollama")
-        == "llamacpp"
+        decide_runtime(flag="ollama", ollama_flag=False, from_config="none") == "ollama"
    )
-    with pytest.raises(ValueError):
-        decide_runtime(flag=None, ollama_flag=True, llamacpp_flag=True, from_config="none")
+    assert decide_runtime(flag=None, ollama_flag=False, from_config="ollama") == "ollama"


 def test_parse_ollama_models(tmp_path: Path):
@@ -60,66 +55,3 @@ def test_write_preset(tmp_path: Path):
    write_ollama_models_preset(path, "recommended")
    entries = parse_ollama_models(path)
    assert entries[0].name == "huihui_ai/qwen2.5-vl-abliterated:7b"
-
-
-def test_write_llamacpp_preset_includes_mmproj(tmp_path: Path):
-    from gpu_rent.llm_runtime import write_llamacpp_models_preset
-
-    path = tmp_path / "lc.yaml"
-    write_llamacpp_models_preset(path, "recommended")
-    entries = parse_llamacpp_models(path)
-    assert len(entries) == 1
-    assert "VL" in entries[0].url or "vl" in entries[0].url.lower()
-    assert entries[0].mmproj_url
-    assert "mmproj" in entries[0].mmproj_url
-
-
-def test_remap_dead_bartowski_abliterate_url(tmp_path: Path):
-    dead = (
-        "https://huggingface.co/bartowski/huihui-ai_Qwen2.5-7B-Instruct-abliterated-GGUF/"
-        "resolve/main/huihui-ai_Qwen2.5-7B-Instruct-abliterated-Q4_K_M.gguf"
-    )
-    fixed = remap_llamacpp_url(dead)
-    assert "RichardErkhov" in fixed
-    assert "Q4_K_M.gguf" in fixed
-    path = tmp_path / "lc.yaml"
-    path.write_text(f"models:\n  - url: {dead}\n    default: true\n", encoding="utf-8")
-    entries = parse_llamacpp_models(path)
-    assert len(entries) == 1
-    assert entries[0].url == fixed
-
-
-def test_pick_llamacpp_linux_asset_skips_windows_cuda():
-    assets = [
-        {
-            "name": "cudart-llama-bin-win-cuda-12.4-x64.zip",
-            "browser_download_url": "https://example/cudart-win.zip",
-        },
-        {
-            "name": "llama-b10545-bin-win-cuda-12.4-x64.zip",
-            "browser_download_url": "https://example/win-cuda.zip",
-        },
-        {
-            "name": "llama-b10545-bin-ubuntu-x64.tar.gz",
-            "browser_download_url": "https://example/ubuntu-cpu.tar.gz",
-        },
-        {
-            "name": "llama-b10545-bin-ubuntu-vulkan-x64.tar.gz",
-            "browser_download_url": "https://example/ubuntu-vulkan.tar.gz",
-        },
-    ]
-    assert pick_llamacpp_linux_asset_url(assets) == "https://example/ubuntu-vulkan.tar.gz"
-
-
-def test_pick_llamacpp_linux_asset_prefers_ubuntu_cuda():
-    assets = [
-        {
-            "name": "llama-b1-bin-ubuntu-vulkan-x64.tar.gz",
-            "browser_download_url": "https://example/vulkan.tar.gz",
-        },
-        {
-            "name": "llama-b1-bin-ubuntu-cuda-12.4-x64.tar.gz",
-            "browser_download_url": "https://example/cuda.tar.gz",
-        },
-    ]
-    assert pick_llamacpp_linux_asset_url(assets) == "https://example/cuda.tar.gz"
@@ -53,7 +53,6 @@ def test_extensions_requires_ollama(tmp_path: Path):
    assert repos[1].requires == "none"
    assert repo_matches_runtime(repos[0], "ollama")
    assert not repo_matches_runtime(repos[0], "none")
-    assert not repo_matches_runtime(repos[0], "llamacpp")
    assert repo_matches_runtime(repos[1], "none")
    assert repo_matches_runtime(repos[1], "ollama")

@@ -67,7 +66,6 @@ def test_extensions_requires_any_llm(tmp_path: Path):
    repo = parse_extensions(path)[0]
    assert repo.requires == "any-llm"
    assert repo_matches_runtime(repo, "ollama")
-    assert repo_matches_runtime(repo, "llamacpp")
    assert not repo_matches_runtime(repo, "none")


@@ -8,7 +8,6 @@ class _Cfg:
    notify_ready = True
    llm_runtime = "none"
    ollama_local_port = 17811
-    llamacpp_local_port = 17812


 def test_ensure_boot_snapshot_skips_existing():
@@ -1,27 +1,21 @@
 from types import SimpleNamespace

-from gpu_rent.provision import _LLAMACPP_INSTALL_ENV, _remote_llm_env
+from gpu_rent.provision import _OLLAMA_INSTALL_ENV, _remote_llm_env


-def test_remote_llm_env_forwards_llamacpp_vars(monkeypatch):
-    monkeypatch.setenv("LLAMACPP_TAG", "b10545")
-    monkeypatch.setenv("LLAMACPP_BUILD_CUDA", "1")
-    monkeypatch.setenv("LLAMACPP_NGL", "40")
-    monkeypatch.setenv("LLAMACPP_ASSET_URL", "https://example/a.tar.gz")
-    monkeypatch.delenv("LLAMACPP_SHA256", raising=False)
+def test_remote_llm_env_forwards_ollama_vars(monkeypatch):
+    monkeypatch.setenv("OLLAMA_VERSION", "0.6.5")
+    monkeypatch.setenv("OLLAMA_SHA256", "abc123")
    cfg = SimpleNamespace(ssh_user="ubuntu")
-    env = _remote_llm_env(cfg, *_LLAMACPP_INSTALL_ENV)
+    env = _remote_llm_env(cfg, *_OLLAMA_INSTALL_ENV)
    assert env["SWARM_USER"] == "ubuntu"
-    assert env["LLAMACPP_TAG"] == "b10545"
-    assert env["LLAMACPP_BUILD_CUDA"] == "1"
-    assert env["LLAMACPP_NGL"] == "40"
-    assert env["LLAMACPP_ASSET_URL"] == "https://example/a.tar.gz"
-    assert "LLAMACPP_SHA256" not in env
+    assert env["OLLAMA_VERSION"] == "0.6.5"
+    assert env["OLLAMA_SHA256"] == "abc123"


 def test_remote_llm_env_skips_empty(monkeypatch):
-    for key in _LLAMACPP_INSTALL_ENV:
+    for key in _OLLAMA_INSTALL_ENV:
        monkeypatch.delenv(key, raising=False)
    cfg = SimpleNamespace(ssh_user="ubuntu")
-    env = _remote_llm_env(cfg, *_LLAMACPP_INSTALL_ENV)
+    env = _remote_llm_env(cfg, *_OLLAMA_INSTALL_ENV)
    assert env == {"SWARM_USER": "ubuntu"}
@@ -55,7 +55,7 @@ def test_prompt_server_plan_with_ranked(monkeypatch, tmp_path):
 def test_upsert_vars(tmp_path):
    path = tmp_path / "gpu-rent.vars"
    path.write_text("# c\nLLM_RUNTIME=none\n", encoding="utf-8")
-    upsert_vars(path, {"LLM_RUNTIME": "llamacpp", "DATA_VOLUME_SIZE_GB": "200"})
+    upsert_vars(path, {"LLM_RUNTIME": "ollama", "DATA_VOLUME_SIZE_GB": "200"})
    data = parse_vars_file(path)
-    assert data["LLM_RUNTIME"] == "llamacpp"
+    assert data["LLM_RUNTIME"] == "ollama"
    assert data["DATA_VOLUME_SIZE_GB"] == "200"
@@ -32,7 +32,6 @@ def test_tunnel_forwards_swarm_only(monkeypatch):
        swarmui_local_port = 17801
        llm_runtime = "none"
        ollama_local_port = 17811
-        llamacpp_local_port = 17812

    monkeypatch.setattr("gpu_rent.tunnel.load_state", lambda: type("S", (), {"notes": {}})())
    assert tunnel_forwards(Cfg()) == [(17801, 7801)]
@@ -43,7 +42,6 @@ def test_tunnel_forwards_prefers_cfg_over_stale_notes(monkeypatch):
        swarmui_local_port = 17801
        llm_runtime = "none"
        ollama_local_port = 17811
-        llamacpp_local_port = 17812

    monkeypatch.setattr(
        "gpu_rent.tunnel.load_state",
@@ -65,10 +63,10 @@ def test_resolve_llm_notes_only_when_cfg_none(monkeypatch):
    assert resolve_llm_runtime(Cfg()) == "ollama"

    class Cfg2:
-        llm_runtime = "llamacpp"
+        llm_runtime = "ollama"

    monkeypatch.setattr(
        "gpu_rent.access_card.load_state",
-        lambda: type("S", (), {"notes": {"llm_runtime": "ollama"}})(),
+        lambda: type("S", (), {"notes": {"llm_runtime": "none"}})(),
    )
-    assert resolve_llm_runtime(Cfg2()) == "llamacpp"
+    assert resolve_llm_runtime(Cfg2()) == "ollama"
@@ -7,19 +7,18 @@ class _Cfg:
    llm_runtime = "none"
    swarmui_local_port = 17801
    ollama_local_port = 17811
-    llamacpp_local_port = 17812


 def test_expected_services_swarm_only():
-    assert _expected_services(_Cfg()) == (True, False, False)
+    assert _expected_services(_Cfg()) == (True, False)


 def test_expected_services_llm_only():
    class C:
        enable_swarmui = False
-        llm_runtime = "llamacpp"
+        llm_runtime = "ollama"

-    assert _expected_services(C()) == (False, False, True)
+    assert _expected_services(C()) == (False, True)


 def test_verify_stack_local_empty_when_nothing():
@@ -28,7 +27,6 @@ def test_verify_stack_local_empty_when_nothing():
        llm_runtime = "none"
        swarmui_local_port = 17801
        ollama_local_port = 17811
-        llamacpp_local_port = 17812

    logs: list[str] = []
    assert verify_stack_local(C(), logs.append, timeout=0.1) == []
@@ -40,7 +38,6 @@ def test_verify_stack_local_fails_closed_port(monkeypatch):
        llm_runtime = "ollama"
        swarmui_local_port = 17801
        ollama_local_port = 17999
-        llamacpp_local_port = 17812

    monkeypatch.setattr(
        "gpu_rent.ready._tcp_ok", lambda port, host="127.0.0.1", timeout=0.8: False
@@ -157,7 +154,7 @@ def test_verify_gpu_env_llm_only_skips_torch_requirement(monkeypatch):

    class C:
        enable_swarmui = False
-        llm_runtime = "llamacpp"
+        llm_runtime = "ollama"

    payload = {
        "ok": True,