diff --git a/docs/cli.md b/docs/cli.md
index 27662c3..349b110 100644
--- a/docs/cli.md
+++ b/docs/cli.md
@@ -231,9 +231,9 @@ Application credential для idle-killer CLI создаёт на `up` (узко
 
 1. Nova `ACTIVE`
 2. TCP 22 / SSH
-3. На VM: HTTP сервисов стека (SwarmUI `:7801` / Ollama `:11434` / llama.cpp `:8080`) — `verify_stack_on_vm`
-4. На VM: **nvidia-smi / CUDA** (+ **torch+cuda** в Comfy venv, если SwarmUI) — `verify_gpu_env`
-5. Backend Idle (если SwarmUI) → toast (если `NOTIFY_READY`)
+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`
+5. На VM: **nvidia-smi / CUDA** (+ **torch+cuda** в Comfy venv, если SwarmUI) — `verify_gpu_env`
 6. Туннель + проверка **localhost** тех же сервисов → access-card
 
 Локальный порт UI: **17801** (на VM по-прежнему 7801 на loopback).
diff --git a/tests/test_verify_stack.py b/tests/test_verify_stack.py
index 1c5b458..66345cc 100644
--- a/tests/test_verify_stack.py
+++ b/tests/test_verify_stack.py
@@ -59,6 +59,10 @@ def test_service_check_dataclass():
 
 
 def test_verify_gpu_env_ok(monkeypatch):
+    import json
+
+    import gpu_rent.ssh_ops as ssh_ops
+
     payload = {
         "ok": True,
         "checks": [
@@ -83,34 +87,12 @@ def test_verify_gpu_env_ok(monkeypatch):
         assert "GPU_RENT_CHECK_SWARM" in script
         return json.dumps(payload) + "\n"
 
-    import json
-
-    monkeypatch.setattr("gpu_rent.ssh_ops.run_python", fake_run_python)
-    # patch where used
-    monkeypatch.setattr(
-        "gpu_rent.ssh_ops.run_python",
-        fake_run_python,
-        raising=False,
-    )
-
-    import gpu_rent.ready as ready_mod
-
-    monkeypatch.setattr(
-        ready_mod,
-        "run_python",
-        fake_run_python,
-        raising=False,
-    )
-
-    # verify_gpu_env imports run_python inside the function
-    import gpu_rent.ssh_ops as ssh_ops
-
     monkeypatch.setattr(ssh_ops, "run_python", fake_run_python)
-
     logs: list[str] = []
     out = verify_gpu_env(_Cfg(), "1.2.3.4", logs.append, timeout=5.0, poll_every=0.1)
     assert all(c.ok for c in out)
     assert any(c.name == "torch" for c in out)
+    assert any("GPU-стека" in line for line in logs)
 
 
 def test_verify_gpu_env_fails_without_cuda(monkeypatch):
@@ -133,9 +115,41 @@ def test_verify_gpu_env_fails_without_cuda(monkeypatch):
     )
     logs: list[str] = []
     try:
-        verify_gpu_env(
-            _Cfg(), "1.2.3.4", logs.append, timeout=0.4, poll_every=0.1
-        )
+        verify_gpu_env(_Cfg(), "1.2.3.4", logs.append, timeout=0.4, poll_every=0.1)
         assert False, "expected CloudError"
     except CloudError as exc:
-        assert "cuda" in str(exc).lower() or "GPU-стек" in str(exc)
+        assert "GPU-стек" in str(exc) or "cuda" in str(exc).lower()
+
+
+def test_verify_gpu_env_llm_only_skips_torch_requirement(monkeypatch):
+    import json
+
+    import gpu_rent.ssh_ops as ssh_ops
+
+    class C:
+        enable_swarmui = False
+        llm_runtime = "llamacpp"
+
+    payload = {
+        "ok": True,
+        "checks": [
+            {"name": "nvidia-smi", "required": True, "ok": True, "detail": "ok"},
+            {"name": "cuda", "required": True, "ok": True, "detail": "ok"},
+            {
+                "name": "torch",
+                "required": False,
+                "ok": True,
+                "detail": "skip (llm-only, без Comfy venv)",
+            },
+        ],
+    }
+
+    def fake_run_python(cfg, host, script, **kw):
+        assert "GPU_RENT_CHECK_SWARM" in script
+        assert "os.environ['GPU_RENT_CHECK_SWARM']='0'" in script
+        return json.dumps(payload)
+
+    monkeypatch.setattr(ssh_ops, "run_python", fake_run_python)
+    logs: list[str] = []
+    out = verify_gpu_env(C(), "1.2.3.4", logs.append, timeout=5.0)
+    assert all(c.ok for c in out)