Warm Ollama into VRAM on up and tunnel so Assistent chat is not cold.

A 1-token /api/chat after tags (and again if /api/ps is empty) loads VL weights before the first message. Mid KEEP_ALIVE is 15m so a short image-gen burst does not unload the model.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/test_llm_runtime.py

@@ -7,6 +7,7 @@ from gpu_rent.llm_runtime import (
     decide_runtime,
     normalize_runtime,
     parse_ollama_models,
+    preferred_ollama_model,
     write_ollama_models_preset,
 )
 
@@ -68,3 +69,61 @@ def test_already_have_ollama_tag_exact_only():
     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")
+
+
+def test_preferred_ollama_model(tmp_path: Path):
+    path = tmp_path / "m.yaml"
+    path.write_text(
+        "models:\n  - name: a:3b\n  - name: b:7b\n    default: true\n",
+        encoding="utf-8",
+    )
+    assert preferred_ollama_model(path) == "b:7b"
+    path.write_text('models:\n  - "only:7b"\n', encoding="utf-8")
+    assert preferred_ollama_model(path) == "only:7b"
+    assert preferred_ollama_model(tmp_path / "missing.yaml") is None
+
+
+def test_warmup_ollama_http_skips_loaded(monkeypatch):
+    from gpu_rent.llm_runtime import warmup_ollama_http
+
+    class Resp:
+        def raise_for_status(self) -> None:
+            return None
+
+        def json(self):
+            return {"models": [{"name": "foo:7b"}]}
+
+    class Client:
+        def __init__(self, *a, **k):
+            pass
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+        def get(self, url, timeout=None):
+            assert url.endswith("/api/ps")
+            return Resp()
+
+        def post(self, *_a, **_k):
+            raise AssertionError("must not chat when already loaded")
+
+    monkeypatch.setattr("gpu_rent.llm_runtime.httpx.Client", Client)
+    msg = warmup_ollama_http("http://127.0.0.1:17811", "foo:7b")
+    assert "skip" in msg
+
+
+def test_maybe_warmup_skips_when_runtime_none():
+    from gpu_rent.llm_runtime import maybe_warmup_ollama_local
+
+    logs: list[str] = []
+
+    class Cfg:
+        llm_runtime = "none"
+        ollama_models_manifest = Path("missing.yaml")
+        ollama_local_port = 17811
+
+    maybe_warmup_ollama_local(Cfg(), logs.append)
+    assert logs == []