Enhance Ollama model management and performance tuning

- Updated the `provision_llm` function to utilize the `/api/tags` endpoint for verifying available models, improving accuracy in model management.
- Introduced a new `already_have_ollama_tag` function to ensure exact tag matching, preventing mismatches during model checks.
- Enhanced the `pull_stream` function to require a successful status from the API before proceeding, ensuring reliable model downloads.
- Added logic to handle unwritten blob files, improving the robustness of the model pulling process.
- Updated documentation and tests to reflect these changes, ensuring clarity and reliability in Ollama model operations.

docs/llm.md

@@ -77,7 +77,7 @@ $env:OLLAMA_HOST = "http://127.0.0.1:17811"
 | `ollama-models.example.yaml` | шаблон в git |
 | `ollama-models.yaml` | список тегов (gitignore); лаунчер копирует example при отсутствии |
 
-На `up` — `ollama pull` по списку. Точные теги: уже есть `foo:7b` ≠ skip для `foo:3b`. Лишнее на диске не удаляет.
+На `up` — `ollama pull` по списку, сверка с **`/api/tags`**. Слой 100% без `status=success` не считается успехом (кэш ≠ модель в Assistent). Если тега всё ещё нет — warning, **GPU не гасим**. Лишнее на диске не удаляет.
 
 ### Пресеты (меню)
 

docs/swarmui.md

@@ -22,7 +22,7 @@
 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 читает его при install.
 10. Старт SwarmUI → **headless InstallConfirmWS** (ComfyUI в `dlbackend`, models=none — веса уже с Civitai) → seed LLM → idle-killer → **wait backend ready** (`running`; `empty`/`loading` ≠ ready; SwarmUI `idle` = бэкенды спят, не «готово»).
-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).
+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). Нужны `python3.12-dev`+gcc (ставим даже на light bootstrap); JIT прогревается на `up`, иначе флаг снимается — иначе первая генерация падает на `cuda_utils.c`.
 12. Авторизация SwarmUI включена, токен в `Data` на диске.
 13. Один snapshot boot volume `gpu-rent-boot-ok`, если ещё нет.
 

src/gpu_rent/llm_runtime.py

@@ -127,6 +127,17 @@ def parse_ollama_models(path: Path) -> list[OllamaModelEntry]:
     return out
 
 
+def already_have_ollama_tag(have: set[str], wanted: str) -> bool:
+    """Exact tag match only — qwen2.5:3b must not satisfy qwen2.5:7b."""
+    if wanted in have:
+        return True
+    if ":" not in wanted and f"{wanted}:latest" in have:
+        return True
+    if wanted.endswith(":latest") and wanted.rsplit(":", 1)[0] in have:
+        return True
+    return False
+
+
 def write_ollama_models_preset(path: Path, preset: str) -> None:
     key = (preset or "recommended").strip().lower()
     if key not in OLLAMA_PRESETS:

src/gpu_rent/provision.py

@@ -923,8 +923,49 @@ def ensure_swarmui_running(cfg: Config, host: str, log: Log, restart: bool) -> N
     run_ssh(cfg, host, "sudo -n systemctl enable swarmui", check=False)
 
 
+_OLLAMA_TAGS_PY = r"""
+import json, urllib.request
+try:
+    with urllib.request.urlopen("http://127.0.0.1:11434/api/tags", timeout=8) as r:
+        data = json.loads(r.read().decode())
+except Exception as exc:
+    print("ERR " + str(exc)[:200])
+    raise SystemExit(0)
+for m in data.get("models") or []:
+    if isinstance(m, dict):
+        for key in ("name", "model"):
+            n = m.get(key)
+            if n:
+                print(n)
+    elif isinstance(m, str) and m.strip():
+        print(m.strip())
+"""
+
+
+def _ollama_api_tags(cfg: Config, host: str) -> set[str]:
+    """Names from Ollama /api/tags (same source as Assistent / verify)."""
+    out = run_ssh(
+        cfg,
+        host,
+        "python3 - <<'PY'\n" + _OLLAMA_TAGS_PY + "\nPY",
+        check=False,
+        timeout=20,
+    )
+    names: set[str] = set()
+    for ln in out.splitlines():
+        s = ln.strip()
+        if not s or s.startswith("ERR "):
+            continue
+        names.add(s)
+    return names
+
+
 def provision_llm(cfg: Config, host: str, log: Log) -> None:
-    from gpu_rent.llm_runtime import normalize_runtime, parse_ollama_models
+    from gpu_rent.llm_runtime import (
+        already_have_ollama_tag,
+        normalize_runtime,
+        parse_ollama_models,
+    )
     from gpu_rent.ssh_ops import run_script_sudo
     from gpu_rent.state import load_state, save_state
 
@@ -950,6 +991,7 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
         st.notes.pop("llm_error", None)
         save_state(st)
         return
+    still: list[str] = []
     if runtime == "ollama":
         _stop_units("gpu-rent-llamacpp")
         log("LLM: ставим/запускаем Ollama")
@@ -967,48 +1009,53 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
         if defaults:
             log(f"Ollama preferred: {defaults[0]}")
         names = [e.name for e in entries]
+        still: list[str] = []
         if not names:
             log("ollama-models.yaml пуст — pull skip")
         else:
-            # Fast path: all tags already present — skip upload/pull script.
-            listed = run_ssh(
-                cfg,
-                host,
-                "ollama list 2>/dev/null | awk 'NR>1 {print $1}' || true",
-                check=False,
-                timeout=30,
-            )
-            have = {ln.strip() for ln in listed.splitlines() if ln.strip()}
-            missing = []
-            for name in names:
-                if name in have or (
-                    ":" not in name and f"{name}:latest" in have
-                ) or (
-                    name.endswith(":latest") and name.rsplit(":", 1)[0] in have
-                ):
-                    continue
-                missing.append(name)
+            have = _ollama_api_tags(cfg, host)
+            missing = [n for n in names if not already_have_ollama_tag(have, n)]
             if not missing:
-                log(f"ollama pull: skip — все {len(names)} уже есть")
+                log(f"ollama pull: skip — /api/tags уже {sorted(have)}")
             else:
                 put_text(
                     cfg, host, "/tmp/gpu-rent-ollama-models.json", json.dumps(missing, indent=2)
                 )
-                log(f"Ollama: pull {len(missing)} из манифеста (нет: {len(missing)})")
-                run_python(
-                    cfg,
-                    host,
-                    _pkg_text("ollama_pull.py"),
-                    remote_path="/tmp/gpu-rent-ollama_pull.py",
-                    timeout=7200,
-                    log=log,
+                log(
+                    f"Ollama: pull {len(missing)} из манифеста "
+                    f"(/api/tags={len(have)})"
+                )
+                try:
+                    run_python(
+                        cfg,
+                        host,
+                        _pkg_text("ollama_pull.py"),
+                        remote_path="/tmp/gpu-rent-ollama_pull.py",
+                        timeout=7200,
+                        log=log,
+                    )
+                except Exception as exc:
+                    log(f"⚠ Ollama pull: {exc}")
+            have = _ollama_api_tags(cfg, host)
+            still = [n for n in names if not already_have_ollama_tag(have, n)]
+            if still:
+                log(
+                    "⚠ Ollama /api/tags без "
+                    + ", ".join(still[:5])
+                    + f" (есть: {sorted(have) or 'пусто'}). "
+                    "SwarmUI ок — GPU не гасим; Assistent будет пустой."
                 )
     else:
         raise CloudError(f"неизвестный LLM_RUNTIME={runtime!r}")
     st = load_state()
     st.notes = dict(st.notes or {})
     st.notes["llm_runtime"] = runtime
-    st.notes.pop("llm_error", None)
+    if still:
+        st.notes["llm_error"] = (
+            "нет в /api/tags: " + ", ".join(still[:5])
+        )[:500]
+    else:
+        st.notes.pop("llm_error", None)
     save_state(st)
 
 

src/gpu_rent/ready.py

@@ -84,7 +84,8 @@ print("WAIT timeout-slice")
 
 # One-shot probe of configured stack endpoints on the VM (JSON line).
 _REMOTE_STACK_PROBE = r'''
-import json, urllib.error, urllib.request, subprocess
+import json, time, urllib.error, urllib.request, subprocess
+from pathlib import Path
 
 def http_ok(url, timeout=4.0):
     try:
@@ -107,9 +108,19 @@ def unit_active(name):
     except Exception:
         return "unknown"
 
+def pulling_age():
+    p = Path("/mnt/swarm_data/.gpu-rent-ollama-pulling")
+    try:
+        if p.is_file():
+            return max(0.0, time.time() - p.stat().st_mtime)
+    except OSError:
+        return None
+    return None
+
 checks = []
 want_swarm = WANT_SWARM
 want_ollama = WANT_OLLAMA
+want_ollama_models = WANT_OLLAMA_MODELS
 
 if want_swarm:
     ok, detail = http_ok("http://127.0.0.1:7801/")
@@ -134,9 +145,11 @@ if want_swarm:
         "ok": ok,
         "detail": detail,
         "unit": unit_active("swarmui"),
+        "retry": not ok,
     })
 
 if want_ollama:
+    retry = True
     try:
         req = urllib.request.Request("http://127.0.0.1:11434/api/tags", method="GET")
         with urllib.request.urlopen(req, timeout=5) as resp:
@@ -154,15 +167,27 @@ if want_ollama:
             preview = ", ".join(names[:3])
             extra = "" if len(names) <= 3 else f" +{len(names) - 3}"
             ok, detail = True, f"{len(names)} models ({preview}{extra})"
+            retry = False
+        elif want_ollama_models:
+            age = pulling_age()
+            if age is not None and age < 2700:
+                ok, detail = False, f"Ollama up, 0 models — pull идёт ({int(age)}s)"
+                retry = True
+            else:
+                ok, detail = True, "WARN 0 models — Assistent empty (GPU не гасим)"
+                retry = False
         else:
-            ok, detail = False, "Ollama up, 0 models — Assistent dropdown empty; ollama pull"
+            ok, detail = True, "0 models (манифест пуст)"
+            retry = False
     except Exception as exc:
         ok, detail = False, str(exc)[:160]
+        retry = True
     checks.append({
         "name": "ollama",
         "ok": ok,
         "detail": detail,
         "unit": unit_active("gpu-rent-ollama"),
+        "retry": retry,
     })
 
 print(json.dumps({"checks": checks}, ensure_ascii=False))
@@ -175,6 +200,7 @@ class ServiceCheck:
     ok: bool
     detail: str
     where: str = "vm"  # vm | local
+    retry: bool = True
 
 
 Log = Callable[[str], None]
@@ -343,10 +369,29 @@ def _expected_services(cfg: Config) -> tuple[bool, bool]:
     return swarm, rt == "ollama"
 
 
+def _want_ollama_models(cfg: Config) -> bool:
+    """True when ollama-models.yaml lists tags that must appear in /api/tags."""
+    if normalize_runtime(getattr(cfg, "llm_runtime", "none")) != "ollama":
+        return False
+    from gpu_rent.llm_runtime import parse_ollama_models
+
+    path = getattr(cfg, "ollama_models_manifest", None)
+    if path is None:
+        return False
+    try:
+        return bool(parse_ollama_models(path))
+    except (OSError, ValueError):
+        return False
+
+
 def _probe_vm_once(cfg: Config, host: str) -> list[ServiceCheck]:
     want_swarm, want_ollama = _expected_services(cfg)
     script = (
         _REMOTE_STACK_PROBE.replace("WANT_SWARM", "True" if want_swarm else "False")
+        .replace(
+            "WANT_OLLAMA_MODELS",
+            "True" if (want_ollama and _want_ollama_models(cfg)) else "False",
+        )
         .replace("WANT_OLLAMA", "True" if want_ollama else "False")
     )
     out = run_ssh(
@@ -380,7 +425,13 @@ def _probe_vm_once(cfg: Config, host: str) -> list[ServiceCheck]:
         if unit and unit != "unknown":
             detail = f"{detail}; unit={unit}"
         checks.append(
-            ServiceCheck(name=name, ok=bool(item.get("ok")), detail=detail, where="vm")
+            ServiceCheck(
+                name=name,
+                ok=bool(item.get("ok")),
+                detail=detail,
+                where="vm",
+                retry=bool(item.get("retry", True)),
+            )
         )
     return checks
 
@@ -417,9 +468,22 @@ def verify_stack_on_vm(
             last = [ServiceCheck("ssh", False, str(exc)[:200], "vm")]
         if last and all(c.ok for c in last):
             for c in last:
-                log(f"  [ok] {c.name}: {c.detail}")
+                mark = "warn" if c.detail.startswith("WARN") else "ok"
+                log(f"  [{mark}] {c.name}: {c.detail}")
             log("проверка VM: всё отвечает")
             return last
+        stuck = [c for c in last if not c.ok and not c.retry]
+        if stuck:
+            for c in last:
+                mark = "ok" if c.ok else "FAIL"
+                log(f"  [{mark}] {c.name}: {c.detail}")
+            if raise_on_fail:
+                failed = [c.name for c in stuck]
+                raise CloudError(
+                    f"{', '.join(failed)} не готов и ждать бесполезно: "
+                    f"{stuck[0].detail}. GPU жив — gpu-rent logs / повторный up (pull)"
+                )
+            return last
         bad = ", ".join(f"{c.name}={c.detail}" for c in last if not c.ok) or "пусто"
         wait.tick(f"  … ещё нет: {bad}")
         time.sleep(poll_every)

src/gpu_rent/remote/bootstrap.sh

@@ -109,6 +109,38 @@ else
   apt-get install -y -qq git python3 python3-venv python3-pip python3-full curl wget ca-certificates
 fi
 
+# Triton/sage JIT: python3.12-dev even on light (warm VMs skipped full apt).
+ensure_triton_build_deps() {
+  if [[ -f /usr/include/python3.12/Python.h ]] && command -v gcc >/dev/null 2>&1; then
+    return 0
+  fi
+  log "ставим gcc + python3.12-dev (Triton cuda_utils JIT)"
+  apt-get update -qq
+  apt-get install -y -qq build-essential python3-dev python3.12-dev
+}
+
+ensure_libcuda_for_triton() {
+  # Triton links -l:libcuda.so.1 -L/lib/x86_64-linux-gnu; Ubuntu driver is under /usr/lib.
+  local dst=/lib/x86_64-linux-gnu/libcuda.so.1
+  if [[ -e "$dst" ]]; then
+    return 0
+  fi
+  local src
+  for src in /usr/lib/x86_64-linux-gnu/libcuda.so.1 /usr/lib/x86_64-linux-gnu/nvidia/current/libcuda.so.1; do
+    if [[ -e "$src" ]]; then
+      mkdir -p /lib/x86_64-linux-gnu
+      ln -sf "$src" "$dst"
+      log "symlink $src → $dst (Triton gcc)"
+      return 0
+    fi
+  done
+}
+
+ensure_triton_build_deps
+ensure_libcuda_for_triton
+mkdir -p /var/tmp/gpu-rent-triton
+chown "${SWARM_USER}:${SWARM_USER}" /var/tmp/gpu-rent-triton 2>/dev/null || true
+
 ensure_data_mount
 
 mkdir -p \
@@ -257,6 +289,13 @@ Environment=HOME=/home/${SWARM_USER}
 Environment=DOTNET_ROOT=/home/${SWARM_USER}/.dotnet
 Environment=DOTNET_CLI_HOME=/home/${SWARM_USER}
 Environment=PATH=/home/${SWARM_USER}/.dotnet:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+Environment=TMPDIR=/tmp
+Environment=TEMP=/tmp
+Environment=TMP=/tmp
+Environment=TRITON_CACHE_DIR=/var/tmp/gpu-rent-triton
+Environment=TRITON_HOME=/var/tmp/gpu-rent-triton
+Environment=LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/lib/x86_64-linux-gnu
+Environment=LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu
 ExecStart=${SWARM_ROOT}/launch-linux.sh --launch_mode none --host 127.0.0.1 --port 7801
 Restart=on-failure
 RestartSec=8

src/gpu_rent/remote/ollama_pull.py

@@ -2,10 +2,12 @@
 """Pull Ollama models from a JSON list. Stdlib only. Runs on the VM.
 
 Uses POST /api/pull with stream JSON for completed/total + speed lines.
+Stream end without status=success is NOT ok (cached layers ≠ registered model).
 """
 from __future__ import annotations
 
 import json
+import os
 import subprocess
 import sys
 import time
@@ -16,6 +18,7 @@ from pathlib import Path
 JOBS = Path("/tmp/gpu-rent-ollama-models.json")
 MARKER = Path("/mnt/swarm_data/.gpu-rent-ollama-pulling")
 OLLAMA = "http://127.0.0.1:11434"
+MODELS_DIR = Path(os.environ.get("OLLAMA_MODELS") or "/mnt/swarm_data/ollama")
 
 
 def fmt_bytes(n: float) -> str:
@@ -42,8 +45,12 @@ def progress_line(label: str, done: int, total: int | None, speed: float, *, wid
 
 def listed() -> set[str]:
     """Exact tags from `ollama list` (NAME column), e.g. qwen2.5:7b."""
+    env = os.environ.copy()
+    env.setdefault("OLLAMA_HOST", "127.0.0.1:11434")
     try:
-        out = subprocess.check_output(["ollama", "list"], text=True, stderr=subprocess.DEVNULL)
+        out = subprocess.check_output(
+            ["ollama", "list"], text=True, stderr=subprocess.DEVNULL, env=env
+        )
     except (subprocess.CalledProcessError, FileNotFoundError):
         return set()
     names: set[str] = set()
@@ -56,6 +63,61 @@ def listed() -> set[str]:
     return names
 
 
+def api_tags() -> set[str]:
+    """Names from GET /api/tags (`name` and `model`)."""
+    try:
+        with urllib.request.urlopen(f"{OLLAMA}/api/tags", timeout=8) as resp:
+            data = json.loads(resp.read().decode())
+    except (urllib.error.URLError, urllib.error.HTTPError, OSError, TimeoutError, json.JSONDecodeError):
+        return set()
+    names: set[str] = set()
+    for m in data.get("models") or []:
+        if isinstance(m, dict):
+            for key in ("name", "model"):
+                val = m.get(key)
+                if val:
+                    names.add(str(val))
+        elif isinstance(m, str) and m.strip():
+            names.add(m.strip())
+    return names
+
+
+def is_unwritten_blob(path: Path, *, sample: int = 64) -> bool:
+    """True if the file is sparse/truncated zeros (Ollama skip-download bug)."""
+    try:
+        size = path.stat().st_size
+    except OSError:
+        return False
+    if size <= 0:
+        return False
+    n = min(sample, size)
+    try:
+        with path.open("rb") as fh:
+            head = fh.read(n)
+    except OSError:
+        return False
+    return bool(head) and head == b"\x00" * len(head)
+
+
+def purge_nul_blobs(models_dir: Path) -> int:
+    """Delete unwritten blob files so the next pull re-downloads layers."""
+    blobs = models_dir / "blobs"
+    if not blobs.is_dir():
+        return 0
+    n = 0
+    for path in blobs.iterdir():
+        if not path.is_file() or not is_unwritten_blob(path):
+            continue
+        size = path.stat().st_size
+        try:
+            path.unlink()
+        except OSError:
+            continue
+        n += 1
+        print(f"удалён пустой blob {path.name} ({size}B)", flush=True)
+    return n
+
+
 def already_have(have: set[str], wanted: str) -> bool:
     """Exact tag match only — qwen2.5:3b must not satisfy qwen2.5:7b."""
     if wanted in have:
@@ -67,8 +129,38 @@ def already_have(have: set[str], wanted: str) -> bool:
     return False
 
 
-def pull_stream(name: str, label: str) -> None:
-    body = json.dumps({"name": name, "stream": True}).encode("utf-8")
+def wait_in_tags(name: str, *, timeout: float) -> set[str]:
+    deadline = time.monotonic() + max(timeout, 0.0)
+    have: set[str] = set()
+    while True:
+        have = api_tags() | listed()
+        if already_have(have, name):
+            return have
+        if time.monotonic() >= deadline:
+            return have
+        time.sleep(1.0)
+
+
+def pull_cli(name: str, label: str) -> None:
+    env = os.environ.copy()
+    env["OLLAMA_HOST"] = "127.0.0.1:11434"
+    proc = subprocess.Popen(
+        ["ollama", "pull", name],
+        env=env,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+    )
+    assert proc.stdout is not None
+    for line in proc.stdout:
+        print(f"{label} {line.rstrip()}", flush=True)
+    code = proc.wait()
+    if code != 0:
+        raise RuntimeError(f"ollama pull CLI exit {code}")
+
+
+def pull_stream(name: str, label: str, *, tags_wait: float = 45.0) -> None:
+    body = json.dumps({"model": name, "name": name, "stream": True}).encode("utf-8")
     req = urllib.request.Request(
         f"{OLLAMA}/api/pull",
         data=body,
@@ -78,6 +170,8 @@ def pull_stream(name: str, label: str) -> None:
     t0 = time.monotonic()
     last_print = 0.0
     last_touch = 0.0
+    last_status = ""
+    got_success = False
     with urllib.request.urlopen(req, timeout=7200) as resp:
         while True:
             raw = resp.readline()
@@ -90,10 +184,10 @@ def pull_stream(name: str, label: str) -> None:
             if ev.get("error"):
                 raise RuntimeError(str(ev["error"]))
             status = str(ev.get("status") or "")
+            last_status = status or last_status
             total = int(ev.get("total") or 0) or None
             done = int(ev.get("completed") or 0)
             now = time.monotonic()
-            # Keep idle-killer marker fresh for long pulls (max-age 3h).
             if now - last_touch >= 60.0:
                 try:
                     MARKER.write_text(f"{int(time.time())}\n", encoding="utf-8")
@@ -112,8 +206,34 @@ def pull_stream(name: str, label: str) -> None:
                 print(f"{label} {status}", flush=True)
                 last_print = now
             if status == "success":
+                got_success = True
                 break
-    print(f"{label} ok", flush=True)
+    if not got_success:
+        print(
+            f"{label} HTTP stream без success (last={last_status!r}) — ollama pull CLI",
+            flush=True,
+        )
+        pull_cli(name, label)
+    have = wait_in_tags(name, timeout=tags_wait)
+    if already_have(have, name):
+        print(f"{label} ok (в /api/tags)", flush=True)
+        return
+    purged = purge_nul_blobs(MODELS_DIR)
+    if purged:
+        print(
+            f"{label} {purged} blob(ов) из нулей (pull skip по размеру) — ollama pull CLI",
+            flush=True,
+        )
+    else:
+        print(f"{label} нет в /api/tags — ollama pull CLI", flush=True)
+    pull_cli(name, label)
+    have = wait_in_tags(name, timeout=tags_wait)
+    if already_have(have, name):
+        print(f"{label} ok (в /api/tags)", flush=True)
+        return
+    raise RuntimeError(
+        f"после pull нет в /api/tags (есть: {sorted(have) or 'пусто'})"
+    )
 
 
 def main() -> int:
@@ -124,7 +244,7 @@ def main() -> int:
     if not isinstance(models, list) or not models:
         print("ollama pull: пустой список — skip")
         return 0
-    have = listed()
+    have = api_tags() | listed()
     MARKER.parent.mkdir(parents=True, exist_ok=True)
     MARKER.write_text(f"{int(time.time())}\n", encoding="utf-8")
     failed = 0
@@ -141,7 +261,13 @@ def main() -> int:
             try:
                 pull_stream(name, label=f"{prefix} {name}")
                 have.add(name)
-            except (urllib.error.URLError, urllib.error.HTTPError, OSError, TimeoutError, RuntimeError) as exc:
+            except (
+                urllib.error.URLError,
+                urllib.error.HTTPError,
+                OSError,
+                TimeoutError,
+                RuntimeError,
+            ) as exc:
                 failed += 1
                 print(f"FAIL pull {name}: {exc}")
     finally:

src/gpu_rent/remote/tune_swarm_perf.py

@@ -243,7 +243,70 @@ def patch_backends_extra_args(extra: str) -> bool:
     return True
 
 
-def sage_already_importable(py: Path) -> bool:
+def strip_sage_extra_args() -> bool:
+    """Remove --use-sage-attention so Comfy does not JIT-crash without python3-dev."""
+    if not BACKENDS.is_file():
+        return False
+    text = BACKENDS.read_text(encoding="utf-8")
+    out: list[str] = []
+    n = 0
+    for line in text.splitlines():
+        m = re.match(r"^(\s*ExtraArgs:\s*)(.*)$", line)
+        if not m or "--use-sage-attention" not in m.group(2):
+            out.append(line)
+            continue
+        rest = re.sub(r"(^|\s)--use-sage-attention\b", "", m.group(2)).strip()
+        out.append(f"{m.group(1)}{rest}")
+        n += 1
+    if not n:
+        return False
+    BACKENDS.write_text("\n".join(out) + "\n", encoding="utf-8")
+    print("stripped --use-sage-attention (triton JIT failed)")
+    return True
+
+
+_TRITON_JIT_PROBE = r"""
+import os
+os.environ.setdefault("TRITON_CACHE_DIR", "/var/tmp/gpu-rent-triton")
+os.environ.setdefault("TMPDIR", "/tmp")
+import torch
+import triton
+import triton.language as tl
+
+@triton.jit
+def _add(x_ptr, y_ptr, n):
+    i = tl.program_id(0)
+    if i < n:
+        tl.store(y_ptr + i, tl.load(x_ptr + i) + 1)
+
+x = torch.ones(32, device="cuda", dtype=torch.float16)
+y = torch.empty_like(x)
+_add[(32,)](x, y, 32)
+torch.cuda.synchronize()
+print("triton JIT ok")
+"""
+
+
+def triton_jit_ok(py: Path) -> bool:
+    """import sageattention is not enough — first gen compiles cuda_utils.c."""
+    env = _pip_env()
+    env["TRITON_CACHE_DIR"] = "/var/tmp/gpu-rent-triton"
+    env["TMPDIR"] = "/tmp"
+    try:
+        out = subprocess.check_output(
+            [str(py), "-c", _TRITON_JIT_PROBE],
+            text=True,
+            stderr=subprocess.STDOUT,
+            timeout=180,
+            env=env,
+        )
+        print((out or "").strip()[-400:] or "triton JIT ok")
+        return True
+    except (subprocess.CalledProcessError, OSError, subprocess.TimeoutExpired) as exc:
+        body = getattr(exc, "output", None) or str(exc)
+        print(f"WARN triton JIT failed: {body[-1500:]}")
+        print("нужны python3.12-dev + gcc; ExtraArgs без --use-sage-attention")
+        return False
     code, _ = _run(
         [str(py), "-c", "import triton, sageattention"],
         timeout=60,
@@ -269,6 +332,14 @@ def _pip_env() -> dict[str, str]:
     return env
 
 
+def sage_already_importable(py: Path) -> bool:
+    code, _ = _run(
+        [str(py), "-c", "import triton, sageattention"],
+        timeout=60,
+    )
+    return code == 0
+
+
 def pip_install_sage(py: Path) -> bool:
     """Install via ``python -m pip`` (never exec venv/bin/pip directly)."""
     if sage_already_importable(py):
@@ -299,16 +370,20 @@ def main() -> int:
         except json.JSONDecodeError:
             prev = {}
     same_gpu = bool(prev.get("uuid") and prev.get("uuid") == plan["uuid"])
-    # Always repair FDS corruption / relative StartScript even when tune is a no-op.
     fds_fixed = sanitize_backends_fds()
     if ensure_absolute_start_script():
         fds_fixed = True
     if same_gpu and prev.get("extra_args") == plan["extra_args"]:
-        if prev.get("pip_ok") or not plan["use_sage"]:
+        jit_ok_prev = bool(prev.get("jit_ok"))
+        if (prev.get("pip_ok") or not plan["use_sage"]) and (
+            not plan["use_sage"] or jit_ok_prev
+        ):
             print(f"perf tune already applied for {plan['name']} ({plan['tier']})")
             if fds_fixed:
                 print("RESTART_SWARMUI=1")
             return 0
+        if plan["use_sage"] and prev.get("pip_ok") and not jit_ok_prev:
+            print("perf tune: retry triton JIT (cuda_utils)")
 
     if same_gpu and plan["use_sage"] and prev.get("pip_ok") is False:
         print("perf tune: retry (previous pip_ok=false — sage/triton ещё не встали)")
@@ -316,16 +391,22 @@ def main() -> int:
     print(f"perf tune: {plan['name']} tier={plan['tier']} sage={plan['use_sage']}")
     restarted_needed = bool(fds_fixed)
     pip_ok = not plan["use_sage"]
+    jit_ok = not plan["use_sage"]
     if plan["use_sage"]:
         py = find_comfy_python()
         if py:
             pip_ok = pip_install_sage(py)
+            jit_ok = bool(pip_ok and triton_jit_ok(py))
         else:
             print("Comfy venv python not found yet — will retry next up")
             pip_ok = False
-        # Only patch ExtraArgs when wheels installed — otherwise Comfy may break.
-        if pip_ok and patch_backends_extra_args(plan["extra_args"]):
-            restarted_needed = True
+            jit_ok = False
+        if pip_ok and jit_ok:
+            if patch_backends_extra_args(plan["extra_args"]):
+                restarted_needed = True
+        else:
+            if strip_sage_extra_args():
+                restarted_needed = True
         if ensure_absolute_start_script():
             restarted_needed = True
 
@@ -333,8 +414,9 @@ def main() -> int:
         "uuid": plan["uuid"],
         "name": plan["name"],
         "tier": plan["tier"],
-        "extra_args": plan["extra_args"] if pip_ok else "",
+        "extra_args": plan["extra_args"] if (pip_ok and jit_ok) else "",
         "pip_ok": pip_ok,
+        "jit_ok": jit_ok,
         "restart_needed": restarted_needed,
     }
     MARKER.write_text(json.dumps(marker, indent=2) + "\n", encoding="utf-8")

tests/test_bootstrap.py

@@ -17,6 +17,9 @@ def test_bootstrap_script_is_native_swarmui():
     assert "GPU_RENT_BOOTSTRAP_LIGHT" in script
     assert "GPU_RENT_SKIP_SWARMUI" in script
     assert "light bootstrap — пропускаем apt-get" in script
+    assert "python3.12-dev" in script
+    assert "TRITON_CACHE_DIR" in script
+    assert "ensure_triton_build_deps" in script
     # llm-only re-up can skip apt when data marker exists (not only Swarm boot marker)
     assert 'MARKER_DATA' in script or ".gpu-rent-ready" in script
     # .NET: chown before Swarm install script; curl fallback if wget/perms fail

tests/test_llm_runtime.py

@@ -3,6 +3,7 @@ from pathlib import Path
 import pytest
 
 from gpu_rent.llm_runtime import (
+    already_have_ollama_tag,
     decide_runtime,
     normalize_runtime,
     parse_ollama_models,
@@ -59,3 +60,11 @@ def test_write_preset(tmp_path: Path):
     assert parse_ollama_models(path)[0].name == "huihui_ai/qwen2.5-vl-abliterated:32b"
     write_ollama_models_preset(path, "text")
     assert parse_ollama_models(path)[0].name == "huihui_ai/qwen2.5-abliterate:7b"
+
+
+def test_already_have_ollama_tag_exact_only():
+    have = {"qwen2.5:7b", "foo:latest"}
+    assert already_have_ollama_tag(have, "qwen2.5:7b")
+    assert not already_have_ollama_tag(have, "qwen2.5:3b")
+    assert already_have_ollama_tag(have, "foo")
+    assert already_have_ollama_tag(have, "foo:latest")

tests/test_ollama_pull.py

@@ -21,7 +21,95 @@ def test_exact_tag_only():
     assert already_have(have, "qwen2.5:3b")
 
 
-def test_latest_alias():
-    assert already_have({"foo:latest"}, "foo")
-    assert already_have({"foo"}, "foo:latest")
-    assert not already_have({"foo:3b"}, "foo")
+def test_pull_stream_requires_success_then_tags(monkeypatch):
+    class FakeResp:
+        def __init__(self, lines: list[str]):
+            self._lines = [ln.encode() for ln in lines]
+            self._i = 0
+
+        def readline(self) -> bytes:
+            if self._i >= len(self._lines):
+                return b""
+            row = self._lines[self._i]
+            self._i += 1
+            return row
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+    lines = [
+        '{"status":"pulling manifest"}\n',
+        '{"status":"downloading","total":100,"completed":100}\n',
+    ]
+    monkeypatch.setattr(
+        _mod.urllib.request, "urlopen", lambda *a, **k: FakeResp(lines)
+    )
+    cli = {"n": 0}
+
+    def fake_cli(*_a, **_k):
+        cli["n"] += 1
+
+    monkeypatch.setattr(_mod, "pull_cli", fake_cli)
+    monkeypatch.setattr(_mod, "api_tags", lambda: set())
+    monkeypatch.setattr(_mod, "listed", lambda: set())
+    monkeypatch.setattr(_mod.time, "sleep", lambda *_a, **_k: None)
+    try:
+        _mod.pull_stream("foo:7b", "x", tags_wait=0)
+        assert False, "expected RuntimeError"
+    except RuntimeError as exc:
+        assert "/api/tags" in str(exc)
+    assert cli["n"] == 2
+
+
+def test_pull_stream_ok_when_success_and_tags(monkeypatch):
+    class FakeResp:
+        def __init__(self, lines: list[str]):
+            self._lines = [ln.encode() for ln in lines]
+            self._i = 0
+
+        def readline(self) -> bytes:
+            if self._i >= len(self._lines):
+                return b""
+            row = self._lines[self._i]
+            self._i += 1
+            return row
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+    lines = ['{"status":"success"}\n']
+    monkeypatch.setattr(
+        _mod.urllib.request, "urlopen", lambda *a, **k: FakeResp(lines)
+    )
+    monkeypatch.setattr(_mod, "api_tags", lambda: {"foo:7b"})
+    monkeypatch.setattr(_mod, "listed", lambda: set())
+    _mod.pull_stream("foo:7b", "x", tags_wait=0)
+
+
+def test_unwritten_blob_is_all_nuls(tmp_path):
+    z = tmp_path / "zeros"
+    z.write_bytes(b"\x00" * 23)
+    real = tmp_path / "gguf"
+    real.write_bytes(b"GGUF" + b"\x00" * 12)
+    empty = tmp_path / "empty"
+    empty.write_bytes(b"")
+    assert _mod.is_unwritten_blob(z)
+    assert not _mod.is_unwritten_blob(real)
+    assert not _mod.is_unwritten_blob(empty)
+
+
+def test_purge_nul_blobs_keeps_gguf(tmp_path):
+    blobs = tmp_path / "blobs"
+    blobs.mkdir()
+    (blobs / "sha256-dead").write_bytes(b"\x00" * 28)
+    (blobs / "sha256-gguf").write_bytes(b"GGUF\x03\x00\x00\x00")
+    n = _mod.purge_nul_blobs(tmp_path)
+    assert n == 1
+    assert not (blobs / "sha256-dead").exists()
+    assert (blobs / "sha256-gguf").exists()

tests/test_tune_swarm_perf.py

@@ -82,6 +82,7 @@ def test_pip_ok_patches_extra_args(tmp_path, monkeypatch):
     monkeypatch.setattr(mod, "find_pip", lambda: pip)
     monkeypatch.setattr(mod, "find_comfy_python", lambda: pip)
     monkeypatch.setattr(mod, "pip_install_sage", lambda _p: True)
+    monkeypatch.setattr(mod, "triton_jit_ok", lambda _p: True)
 
     assert mod.main() == 0
     marker = json.loads((data / ".gpu-rent-perf-tuned").read_text(encoding="utf-8"))
@@ -131,6 +132,7 @@ def test_pip_fail_retries_next_run(tmp_path, monkeypatch):
     monkeypatch.setattr(mod, "find_pip", lambda: pip)
     monkeypatch.setattr(mod, "find_comfy_python", lambda: pip)
     monkeypatch.setattr(mod, "pip_install_sage", lambda _p: True)
+    monkeypatch.setattr(mod, "triton_jit_ok", lambda _p: True)
 
     assert mod.main() == 0
     new_m = json.loads(marker.read_text(encoding="utf-8"))
@@ -223,3 +225,37 @@ def test_ensure_absolute_start_script(tmp_path, monkeypatch):
     text = backends.read_text(encoding="utf-8")
     assert str(main_py.resolve()) in text
     assert mod.ensure_absolute_start_script() is False
+
+
+def test_jit_fail_strips_sage_extra_args(tmp_path, monkeypatch):
+    mod = _load()
+    data = tmp_path
+    backends = data / "Data" / "Backends.fds"
+    backends.parent.mkdir(parents=True)
+    backends.write_text("ExtraArgs: --use-sage-attention\n", encoding="utf-8")
+    gpu_json = data / ".gpu-rent-gpu.json"
+    gpu_json.write_text(
+        json.dumps(
+            {
+                "vram_mib": 24576,
+                "compute_cap": "8.9",
+                "uuid": "gpu-1",
+                "name": "RTX",
+            }
+        ),
+        encoding="utf-8",
+    )
+    pip = data / "fake-pip"
+    pip.write_text("#!/bin/sh\n", encoding="utf-8")
+    monkeypatch.setattr(mod, "DATA", data)
+    monkeypatch.setattr(mod, "GPU_JSON", gpu_json)
+    monkeypatch.setattr(mod, "MARKER", data / ".gpu-rent-perf-tuned")
+    monkeypatch.setattr(mod, "BACKENDS", backends)
+    monkeypatch.setattr(mod, "find_comfy_python", lambda: pip)
+    monkeypatch.setattr(mod, "pip_install_sage", lambda _p: True)
+    monkeypatch.setattr(mod, "triton_jit_ok", lambda _p: False)
+    assert mod.main() == 0
+    marker = json.loads((data / ".gpu-rent-perf-tuned").read_text(encoding="utf-8"))
+    assert marker["pip_ok"] is True
+    assert marker["jit_ok"] is False
+    assert "--use-sage-attention" not in backends.read_text(encoding="utf-8")

tests/test_verify_stack.py

@@ -188,4 +188,59 @@ def test_stack_probe_requires_ollama_models():
 
     assert "/api/tags" in _REMOTE_STACK_PROBE
     assert "0 models" in _REMOTE_STACK_PROBE
-    assert 'payload.get("models")' in _REMOTE_STACK_PROBE
+    assert "WANT_OLLAMA_MODELS" in _REMOTE_STACK_PROBE
+    assert "retry" in _REMOTE_STACK_PROBE
+    assert ".gpu-rent-ollama-pulling" in _REMOTE_STACK_PROBE
+    assert "WARN 0 models" in _REMOTE_STACK_PROBE
+    assert "GPU не гасим" in _REMOTE_STACK_PROBE
+
+
+def test_verify_stack_zero_models_does_not_fail_up(monkeypatch):
+    from gpu_rent.ready import ServiceCheck, verify_stack_on_vm
+
+    class C:
+        enable_swarmui = True
+        llm_runtime = "ollama"
+
+    def fake_probe(_cfg, _host):
+        return [
+            ServiceCheck("swarmui", True, "ok", "vm", retry=False),
+            ServiceCheck(
+                "ollama",
+                True,
+                "WARN 0 models — Assistent empty (GPU не гасим)",
+                "vm",
+                retry=False,
+            ),
+        ]
+
+    monkeypatch.setattr("gpu_rent.ready._probe_vm_once", fake_probe)
+    out = verify_stack_on_vm(C(), "1.2.3.4", [].append, timeout=5.0, poll_every=0.01)
+    assert all(c.ok for c in out)
+
+
+def test_verify_stack_retries_while_pulling(monkeypatch):
+    from gpu_rent.ready import ServiceCheck, verify_stack_on_vm
+
+    class C:
+        enable_swarmui = False
+        llm_runtime = "ollama"
+
+    n = {"i": 0}
+
+    def fake_probe(_cfg, _host):
+        n["i"] += 1
+        if n["i"] == 1:
+            return [
+                ServiceCheck(
+                    "ollama", False, "0 models — pull идёт (12s)", "vm", retry=True
+                )
+            ]
+        return [ServiceCheck("ollama", True, "1 models (qwen)", "vm", retry=False)]
+
+    monkeypatch.setattr("gpu_rent.ready._probe_vm_once", fake_probe)
+    out = verify_stack_on_vm(
+        C(), "1.2.3.4", [].append, timeout=5.0, poll_every=0.01
+    )
+    assert out[0].ok
+    assert n["i"] == 2

tests/test_warm_up_skips.py

@@ -88,6 +88,18 @@ def test_install_ollama_skips_restart_when_unit_unchanged():
     assert "skip restart" in text
 
 
+def test_provision_llm_skips_on_api_tags_not_cli_list():
+    from pathlib import Path
+
+    from gpu_rent import provision
+
+    text = Path(provision.__file__).read_text(encoding="utf-8")
+    assert "_ollama_api_tags" in text
+    assert "awk 'NR>1" not in text
+    assert "GPU не гасим" in text
+    assert "без моделей из ollama-models.yaml" not in text
+
+
 def test_cli_has_update_flag():
     from gpu_rent import cli