Update LLM support for llama.cpp and enhance configuration management

- Added support for `llamacpp-models.yaml` in `.gitignore` and implemented logic to copy it in `gpu-rent.ps1` and `gpu-rent.sh`.
- Enhanced CLI to prompt for llama.cpp model presets during setup and execution, improving user experience.
- Updated configuration handling to include `llamacpp_models_manifest` and related functions for managing llama.cpp models.
- Improved documentation in `cli.md` and `llm.md` to reflect changes in llama.cpp integration and model management.
- Refactored provisioning logic to handle llama.cpp model downloads and configurations effectively.

.gitignore

@@ -4,6 +4,7 @@
 !.env.example
 gpu-rent.vars
 ollama-models.yaml
+llamacpp-models.yaml
 models.yaml
 extensions.yaml
 

docs/cli.md

@@ -55,7 +55,7 @@ gpu-rent up --yes --ollama
 | `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 |
-| `gpu-rent up` / `up --yes` | Create/unshelve → bootstrap → optional LLM → **туннель** `:17801`. Ctrl+C = туннель off |
+| `gpu-rent up` / `up --yes` | Create/unshelve → bootstrap → optional LLM → **туннель** `:17801`. Без `--yes`: выбор flavor / data GB / preemptible, затем confirm. Ctrl+C = туннель off |
 | `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-update` | Без `git pull` SwarmUI/extensions (только недостающие clone) |

docs/llm.md

@@ -12,7 +12,7 @@
 gpu-rent setup
 ```
 
-Выбери `ollama` или `llamacpp`, при Ollama — пресет моделей. Значение пишется в `gpu-rent.vars` (`LLM_RUNTIME=…`).
+Выбери `ollama` или `llamacpp`. Для **обоих** спросит пресет моделей. Значение пишется в `gpu-rent.vars` (`LLM_RUNTIME=…`).
 
 ### Вариант B — флаг на `up`
 
@@ -103,14 +103,25 @@ Community abliterate-модели без гарантий безопасност
 
 ## llama.cpp
 
-1. `LLM_RUNTIME=llamacpp` или `up --llamacpp`
+Манифест GGUF:
-2. CLI ставит `llama-server` + systemd `gpu-rent-llamacpp`
-3. Положи GGUF в `/mnt/swarm_data/llamacpp/models` (SFTP / `gpu-rent ssh`)
-4. `systemctl restart gpu-rent-llamacpp` на VM
 
-Без GGUF unit может стартовать, но API бесполезен — смотри `gpu-rent logs` / `journalctl -u gpu-rent-llamacpp`.
+| Файл | Роль |
+| --- | --- |
+| `llamacpp-models.example.yaml` | шаблон в git |
+| `llamacpp-models.yaml` | URL на `.gguf` (gitignore) |
 
-Параметры `-ngl` / `-c` ставятся по тому же GPU probe (full offload на mid+, меньше слоёв и ctx на low).
+На interactive `up` / `setup` после выбора `llamacpp` спрашивается пресет (как у Ollama). На `up` CLI скачивает GGUF в `/mnt/swarm_data/llamacpp/models`, затем ставит `llama-server` + systemd.
+
+### Пресеты
+
+| preset | что |
+| --- | --- |
+| **recommended** | Qwen2.5 7B abliterate Q4_K_M (~4.7 GB) |
+| light | Qwen2.5 3B Instruct Q4_K_M |
+| stock | официальный Qwen2.5 7B Instruct Q4_K_M |
+| empty | только runtime |
+
+Параметры `-ngl` / `-c` — по GPU probe. Опционально `HF_TOKEN` для gated HF. Вручную: положи GGUF в каталог models и `systemctl restart gpu-rent-llamacpp`.
 
 ---
 

gpu-rent.ps1

@@ -119,6 +119,7 @@ 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")

gpu-rent.sh

@@ -99,6 +99,10 @@ 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"

llamacpp-models.example.yaml

@@ -0,0 +1,15 @@
+# Copy to llamacpp-models.yaml (gitignored). Used when LLM_RUNTIME=llamacpp.
+# url = direct HTTPS link to a .gguf (Hugging Face resolve/main/…).
+# Empty models: [] → only llama-server, GGUF клади вручную на VM.
+# Purpose: prompt-help beside SwarmUI (RU/EN).
+
+models:
+  # Recommended: Qwen2.5 7B abliterate Q4_K_M (~4.7 GB)
+  - url: 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
+    default: true
+
+  # Lighter (~2 GB):
+  # - url: https://huggingface.co/bartowski/Qwen2.5-3B-Instruct-GGUF/resolve/main/Qwen2.5-3B-Instruct-Q4_K_M.gguf
+
+  # Official stock 7B (more refusals):
+  # - url: https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF/resolve/main/Qwen2.5-7B-Instruct-Q4_K_M.gguf

src/gpu_rent/cli.py

@@ -436,6 +436,7 @@ def up(
         except ValueError as exc:
             raise GpuRentError(str(exc)) from exc
 
+        asked_model_preset = False
         if not yes and runtime == "none" and not llm and not ollama and not llamacpp:
             choice = typer.prompt(
                 "Поднять LLM рядом со SwarmUI? [none/ollama/llamacpp]",
@@ -457,6 +458,53 @@ def up(
                     write_ollama_models_preset(
                         cfg.ollama_models_manifest, preset.strip().lower()
                     )
+                asked_model_preset = True
+            elif runtime == "llamacpp":
+                from gpu_rent.llm_runtime import (
+                    LLAMACPP_PRESET_HELP,
+                    ensure_llamacpp_manifest_from_example,
+                    write_llamacpp_models_preset,
+                )
+
+                ensure_llamacpp_manifest_from_example()
+                console.print(LLAMACPP_PRESET_HELP)
+                preset = typer.prompt(
+                    "llama.cpp GGUF preset [recommended/light/stock/empty]",
+                    default="recommended",
+                )
+                if preset.strip().lower() not in {"keep", "example"}:
+                    write_llamacpp_models_preset(
+                        cfg.llamacpp_models_manifest, preset.strip().lower()
+                    )
+                asked_model_preset = True
+
+        # Runtime уже в vars (напр. llamacpp) — всё равно спросить модель, default=keep.
+        if not yes and not asked_model_preset and runtime == "llamacpp":
+            from gpu_rent.llm_runtime import (
+                LLAMACPP_PRESET_HELP,
+                ensure_llamacpp_manifest_from_example,
+                write_llamacpp_models_preset,
+            )
+
+            ensure_llamacpp_manifest_from_example()
+            console.print(LLAMACPP_PRESET_HELP)
+            preset = typer.prompt(
+                "llama.cpp GGUF preset [recommended/light/stock/empty/keep]",
+                default="keep",
+            )
+            key = preset.strip().lower()
+            if key not in {"keep", "example", ""}:
+                write_llamacpp_models_preset(cfg.llamacpp_models_manifest, key)
+        elif not yes and not asked_model_preset and runtime == "ollama":
+            ensure_ollama_manifest_from_example()
+            console.print(PRESET_HELP)
+            preset = typer.prompt(
+                "Ollama preset [recommended/light/stock/alt/empty/keep]",
+                default="keep",
+            )
+            key = preset.strip().lower()
+            if key not in {"keep", "example", ""}:
+                write_ollama_models_preset(cfg.ollama_models_manifest, key)
 
         cfg = replace(cfg, llm_runtime=runtime)
         if runtime != "none":
@@ -465,6 +513,9 @@ def up(
         def confirm(msg: str) -> bool:
             return typer.confirm(msg)
 
+        def ask(msg: str, default: str = "") -> str:
+            return typer.prompt(msg, default=default)
+
         state = cmd_up(
             cfg,
             no_spot=no_spot,
@@ -473,6 +524,7 @@ def up(
             adopt=adopt,
             update=False if no_update else None,
             confirm=confirm,
+            ask=None if yes else ask,
             log=lambda m: console.print(m),
         )
         if no_tunnel:

src/gpu_rent/config.py

@@ -14,6 +14,7 @@ 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,
@@ -84,6 +85,7 @@ class Config:
 
     llm_runtime: str
     ollama_models_manifest: Path
+    llamacpp_models_manifest: Path
     ollama_local_port: int
     llamacpp_local_port: int
 
@@ -157,6 +159,10 @@ 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"))
@@ -205,6 +211,7 @@ def load_config(*, require_auth: bool = True) -> Config:
         update_git=_as_bool(os.environ.get("UPDATE_GIT"), True),
         llm_runtime=llm_runtime,
         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(),

src/gpu_rent/llm_runtime.py

@@ -5,14 +5,19 @@ 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 ollama_models_example_path, ollama_models_manifest_path
+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"})
 
-# Presets for setup / interactive up (prompt help: RU + low refusal).
 OLLAMA_PRESETS: dict[str, list[str]] = {
     "recommended": ["huihui_ai/qwen2.5-abliterate:7b"],
     "light": ["qwen2.5:3b"],
@@ -29,6 +34,26 @@ PRESET_HELP = (
     "empty — только runtime, без pull"
 )
 
+LLAMACPP_PRESETS: dict[str, list[str]] = {
+    "recommended": [
+        "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",
+    ],
+    "light": [
+        "https://huggingface.co/bartowski/Qwen2.5-3B-Instruct-GGUF/resolve/main/Qwen2.5-3B-Instruct-Q4_K_M.gguf",
+    ],
+    "stock": [
+        "https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF/resolve/main/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
+    ],
+    "empty": [],
+}
+
+LLAMACPP_PRESET_HELP = (
+    "recommended — Qwen2.5 7B abliterate GGUF Q4_K_M (~4.7GB, мало отказов)\n"
+    "light — Qwen2.5 3B Instruct Q4_K_M (~2GB)\n"
+    "stock — официальный Qwen2.5 7B Instruct Q4_K_M\n"
+    "empty — только llama-server, GGUF положи вручную / правь llamacpp-models.yaml"
+)
+
 
 @dataclass(frozen=True)
 class OllamaModelEntry:
@@ -36,6 +61,13 @@ class OllamaModelEntry:
     default: bool = False
 
 
+@dataclass(frozen=True)
+class LlamaCppModelEntry:
+    url: str
+    filename: 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"}:
@@ -54,7 +86,6 @@ def decide_runtime(
     llamacpp_flag: bool,
     from_config: str,
 ) -> str:
-    """CLI flags win over config/vars."""
     if ollama_flag and llamacpp_flag:
         raise ValueError("укажи только --ollama или --llamacpp, не оба")
     if ollama_flag:
@@ -93,6 +124,49 @@ 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"
+
+
+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 = item.strip()
+            if url:
+                out.append(LlamaCppModelEntry(url=url))
+            continue
+        if not isinstance(item, dict):
+            continue
+        url = str(item.get("url") or "").strip()
+        if not url:
+            continue
+        fname = item.get("filename")
+        filename = str(fname).strip() if fname else None
+        out.append(
+            LlamaCppModelEntry(
+                url=url,
+                filename=filename 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:
@@ -113,6 +187,26 @@ 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)}")
+    urls = LLAMACPP_PRESETS[key]
+    lines = [
+        "# Локальный манифест llama.cpp GGUF (не коммить). Пример: llamacpp-models.example.yaml",
+        "# url = прямой HTTPS на .gguf. Пустой models: [] — без скачивания.",
+        "models:",
+    ]
+    if not urls:
+        lines.append("  []")
+    else:
+        for i, url in enumerate(urls):
+            lines.append(f"  - url: {url}")
+            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():
@@ -125,6 +219,18 @@ 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":
@@ -144,23 +250,6 @@ def llm_remote_port(runtime: str) -> int | None:
 
 
 def append_vars_llm_runtime(vars_file: Path, runtime: str) -> None:
-    runtime = normalize_runtime(runtime)
+    from gpu_rent.varsfile import upsert_vars
-    line = f"LLM_RUNTIME={runtime}\n"
+
-    if not vars_file.is_file():
+    upsert_vars(vars_file, {"LLM_RUNTIME": normalize_runtime(runtime)})
-        vars_file.write_text("# gpu-rent.vars — несекреты\n" + line, encoding="utf-8")
-        return
-    text = vars_file.read_text(encoding="utf-8")
-    lines = text.splitlines(keepends=True)
-    out: list[str] = []
-    replaced = False
-    for row in lines:
-        if row.lstrip().startswith("LLM_RUNTIME="):
-            out.append(line if row.endswith("\n") else line.rstrip("\n"))
-            replaced = True
-        else:
-            out.append(row)
-    if not replaced:
-        if out and not out[-1].endswith("\n"):
-            out[-1] = out[-1] + "\n"
-        out.append(line)
-    vars_file.write_text("".join(out), encoding="utf-8")

src/gpu_rent/paths.py

@@ -56,6 +56,14 @@ 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"
 

src/gpu_rent/provision.py

@@ -372,6 +372,48 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
             )
     elif runtime == "llamacpp":
         _stop_units("gpu-rent-ollama")
+        from gpu_rent.llm_runtime import (
+            gguf_filename_from_url,
+            parse_llamacpp_models,
+        )
+        import os
+
+        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 = [
+                {
+                    "url": e.url,
+                    "filename": e.filename or gguf_filename_from_url(e.url),
+                }
+                for e in entries
+            ]
+            put_text(
+                cfg, host, "/tmp/gpu-rent-llamacpp-models.json", json.dumps(jobs, indent=2)
+            )
+            hf = (
+                os.environ.get("HF_TOKEN")
+                or os.environ.get("HUGGING_FACE_HUB_TOKEN")
+                or ""
+            ).strip()
+            if hf:
+                put_text(cfg, host, "/tmp/gpu-rent-hf.token", hf + "\n", mode=0o600)
+            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")
         run_script_sudo(
             cfg,

src/gpu_rent/remote/llamacpp_fetch.py

@@ -0,0 +1,78 @@
+#!/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 urllib.error
+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 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", file=sys.stderr)
+        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
+        if dest.is_file() and dest.stat().st_size > 1_000_000:
+            print(f"[{i}/{len(jobs)}] уже есть {name} ({dest.stat().st_size} bytes)")
+            continue
+        print(f"[{i}/{len(jobs)}] download {name}")
+        partial = dest.with_suffix(dest.suffix + ".partial")
+        headers = {"User-Agent": "gpu-rent/1"}
+        if token:
+            headers["Authorization"] = f"Bearer {token}"
+        req = urllib.request.Request(url, headers=headers)
+        try:
+            with urllib.request.urlopen(req, timeout=600) as resp, partial.open("wb") as out:
+                while True:
+                    chunk = resp.read(1024 * 1024)
+                    if not chunk:
+                        break
+                    out.write(chunk)
+            partial.replace(dest)
+            print(f"ok {name} ({dest.stat().st_size} bytes)")
+        except (urllib.error.URLError, urllib.error.HTTPError, OSError, TimeoutError) as exc:
+            failed += 1
+            print(f"FAIL {name}: {exc}", file=sys.stderr)
+            try:
+                partial.unlink(missing_ok=True)
+            except OSError:
+                pass
+    if failed:
+        return 1
+    print("llamacpp fetch ok")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/gpu_rent/session.py

@@ -33,7 +33,7 @@ from gpu_rent.inventory import (
     pick_volume_type,
     resolve_flavor,
 )
-from gpu_rent.ux import print_up_preview
+from gpu_rent.ux import print_up_preview, prompt_server_plan
 from gpu_rent.pools import best_offer, format_pool_scan, scan_pools
 from gpu_rent.lock import SessionLock
 from gpu_rent.os_client import (
@@ -173,6 +173,7 @@ def cmd_up(
     adopt: bool = False,
     update: bool | None = None,
     confirm: Callable[[str], bool] | None = None,
+    ask: Callable[[str, str], str] | None = None,
     log: Log = _log_default,
 ) -> SessionState:
     do_update = cfg.update_git if update is None else update
@@ -291,6 +292,29 @@ def cmd_up(
 
         print_up_preview(cfg, flavors, picked=picked, spot=spot, log=log)
 
+        if not yes and ask is not None and flavor is None:
+            try:
+                plan = prompt_server_plan(
+                    cfg,
+                    flavors,
+                    picked=picked,
+                    spot=spot,
+                    ask=ask,
+                    confirm=confirm,
+                )
+            except ValueError as exc:
+                raise GpuRentError(str(exc)) from exc
+            picked = plan.flavor
+            spot = plan.spot
+            if plan.data_gb != cfg.data_volume_size_gb:
+                from dataclasses import replace
+
+                cfg = replace(cfg, data_volume_size_gb=plan.data_gb)
+            log(
+                f"выбрано: {'preemptible ' if spot else ''}{picked.name}, "
+                f"data {cfg.data_volume_size_gb} GB"
+            )
+
         prompt = (
             f"Создать {'preemptible ' if spot else ''}GPU {picked.name} "
             f"в {cfg.gpu_rent_az}, образ {getattr(image, 'name', image.id)}, "

src/gpu_rent/setup_wizard.py

@@ -7,16 +7,21 @@ from collections.abc import Callable
 from pathlib import Path
 
 from gpu_rent.llm_runtime import (
+    LLAMACPP_PRESET_HELP,
     PRESET_HELP,
     append_vars_llm_runtime,
+    ensure_llamacpp_manifest_from_example,
     ensure_ollama_manifest_from_example,
     normalize_runtime,
+    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,
@@ -59,6 +64,12 @@ 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:
@@ -86,6 +97,22 @@ 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  # reuse --ollama-preset flag as generic LLM preset in setup
+        if preset is None and ask:
+            log(LLAMACPP_PRESET_HELP)
+            preset = ask(
+                "llama.cpp GGUF preset [recommended/light/stock/empty]",
+                "recommended",
+            )
+        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:

src/gpu_rent/ux.py

@@ -3,12 +3,15 @@
 from __future__ import annotations
 
 from collections.abc import Callable
+from dataclasses import dataclass
 from typing import Any
 
 from gpu_rent.config import Config
 from gpu_rent.inventory import FlavorInfo, looks_like_gpu, rank_flavors, resolve_flavor
 
 Log = Callable[[str], None]
+Ask = Callable[[str, str], str]
+Confirm = Callable[[str], bool]
 
 
 def list_ranked_flavors(flavors: list[Any], cfg: Config) -> list[FlavorInfo]:
@@ -61,6 +64,93 @@ def print_up_preview(
         log(f"! {line}")
 
 
+@dataclass(frozen=True)
+class ServerPlan:
+    flavor: FlavorInfo
+    data_gb: int
+    spot: bool
+
+
+def prompt_server_plan(
+    cfg: Config,
+    flavors: list[Any],
+    *,
+    picked: FlavorInfo,
+    spot: bool,
+    ask: Ask,
+    confirm: Confirm | None = None,
+) -> ServerPlan:
+    """Interactive Selectel VM knobs before create (flavor / disk / spot)."""
+    ranked = list_ranked_flavors(flavors, cfg)
+    if not ranked:
+        return ServerPlan(flavor=picked, data_gb=cfg.data_volume_size_gb, spot=spot)
+
+    default_idx = 1
+    for i, info in enumerate(ranked, 1):
+        if info.id == picked.id:
+            default_idx = i
+            break
+
+    raw_idx = ask(
+        f"Flavor Selectel [1-{len(ranked)}] (Enter = рекомендация)",
+        str(default_idx),
+    ).strip()
+    try:
+        idx = int(raw_idx)
+    except ValueError as exc:
+        raise ValueError(f"номер flavor: жду 1…{len(ranked)}, получили {raw_idx!r}") from exc
+    if idx < 1 or idx > len(ranked):
+        raise ValueError(f"номер flavor: жду 1…{len(ranked)}, получили {idx}")
+    chosen = ranked[idx - 1]
+
+    raw_gb = ask(
+        "Data disk GB (тариф 24/7; рост только вверх)",
+        str(cfg.data_volume_size_gb),
+    ).strip()
+    try:
+        data_gb = int(raw_gb)
+    except ValueError as exc:
+        raise ValueError(f"Data disk GB: жду число, получили {raw_gb!r}") from exc
+    if data_gb < 20:
+        raise ValueError("Data disk GB: минимум 20")
+
+    spot_default = "Y" if spot else "n"
+    raw_spot = (
+        ask(
+            "Preemptible GPU (дешевле, могут усыпить ~24ч)? [Y/n]",
+            spot_default,
+        )
+        .strip()
+        .lower()
+    )
+    if raw_spot in {"", "y", "yes", "1", "true", "on"}:
+        use_spot = True
+    elif raw_spot in {"n", "no", "0", "false", "off"}:
+        use_spot = False
+    else:
+        raise ValueError(f"preemptible: жду Y/n, получили {raw_spot!r}")
+
+    changed = (
+        chosen.id != picked.id
+        or data_gb != cfg.data_volume_size_gb
+        or use_spot != spot
+    )
+    if confirm and changed and confirm("Запомнить flavor/disk/spot в gpu-rent.vars?"):
+        from gpu_rent.paths import vars_path
+        from gpu_rent.varsfile import upsert_vars
+
+        upsert_vars(
+            vars_path(),
+            {
+                "DEFAULT_FLAVOR_ID": chosen.id,
+                "DATA_VOLUME_SIZE_GB": str(data_gb),
+                "DEFAULT_SPOT": "true" if use_spot else "false",
+            },
+        )
+
+    return ServerPlan(flavor=chosen, data_gb=data_gb, spot=use_spot)
+
+
 def resolve_and_preview(
     cfg: Config,
     flavors: list[Any],

src/gpu_rent/varsfile.py

@@ -41,6 +41,40 @@ def apply_vars_file(path: Path, *, override: bool = False) -> dict[str, str]:
     return loaded
 
 
+def upsert_vars(path: Path, updates: dict[str, str]) -> None:
+    """Create or replace KEY=VALUE lines in a vars file (preserve comments/order)."""
+    if not updates:
+        return
+    if not path.is_file():
+        lines = ["# gpu-rent.vars — несекреты\n"]
+        for key, value in updates.items():
+            lines.append(f"{key}={value}\n")
+        path.write_text("".join(lines), encoding="utf-8")
+        return
+    text = path.read_text(encoding="utf-8")
+    rows = text.splitlines(keepends=True)
+    pending = dict(updates)
+    out: list[str] = []
+    for row in rows:
+        stripped = row.lstrip()
+        key = None
+        if "=" in stripped and not stripped.startswith("#"):
+            maybe = stripped.split("=", 1)[0].strip()
+            if maybe in pending:
+                key = maybe
+        if key is not None:
+            line = f"{key}={pending.pop(key)}\n"
+            out.append(line if row.endswith("\n") else line.rstrip("\n"))
+        else:
+            out.append(row)
+    if pending:
+        if out and not out[-1].endswith("\n"):
+            out[-1] = out[-1] + "\n"
+        for key, value in pending.items():
+            out.append(f"{key}={value}\n")
+    path.write_text("".join(out), encoding="utf-8")
+
+
 def split_args(value: str | None) -> list[str]:
     if not value or not value.strip():
         return []

tests/test_llamacpp_models.py

@@ -0,0 +1,31 @@
+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) == []

tests/test_server_plan.py

@@ -0,0 +1,61 @@
+from gpu_rent.inventory import FlavorInfo
+from gpu_rent.ux import prompt_server_plan
+from gpu_rent.varsfile import parse_vars_file, upsert_vars
+
+
+def _cfg_stub(monkeypatch, data_gb: int = 100):
+    monkeypatch.setenv("OS_AUTH_URL", "https://example.invalid/identity/v3")
+    monkeypatch.setenv("OS_USER_DOMAIN_NAME", "999")
+    monkeypatch.setenv("OS_USERNAME", "svc")
+    monkeypatch.setenv("OS_PASSWORD", "secret")
+    monkeypatch.setenv("OS_PROJECT_ID", "proj")
+    monkeypatch.setenv("OS_REGION_NAME", "ru-7")
+    monkeypatch.setenv("GPU_RENT_AZ", "ru-7a")
+    monkeypatch.setenv("DATA_VOLUME_SIZE_GB", str(data_gb))
+    from gpu_rent.config import load_config
+
+    return load_config(require_auth=True)
+
+
+def test_prompt_server_plan_with_ranked(monkeypatch, tmp_path):
+    cfg = _cfg_stub(monkeypatch)
+    monkeypatch.setattr("gpu_rent.paths.app_root", lambda: tmp_path)
+    f1 = FlavorInfo(
+        id="a", name="small", vcpus=4, ram_mb=16384, disabled=False, extra={}, label="4090-24"
+    )
+    f2 = FlavorInfo(
+        id="b", name="big", vcpus=12, ram_mb=65536, disabled=False, extra={}, label="4090-48"
+    )
+
+    monkeypatch.setattr(
+        "gpu_rent.ux.list_ranked_flavors",
+        lambda flavors, cfg: [f1, f2],
+    )
+    answers = iter(["2", "200", "n"])
+
+    def ask(msg: str, default: str = "") -> str:
+        return next(answers)
+
+    remembered: list[bool] = []
+
+    plan = prompt_server_plan(
+        cfg,
+        ["x"],
+        picked=f1,
+        spot=True,
+        ask=ask,
+        confirm=lambda m: remembered.append(True) or False,
+    )
+    assert plan.flavor.id == "b"
+    assert plan.data_gb == 200
+    assert plan.spot is False
+    assert remembered
+
+
+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"})
+    data = parse_vars_file(path)
+    assert data["LLM_RUNTIME"] == "llamacpp"
+    assert data["DATA_VOLUME_SIZE_GB"] == "200"