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 == []

tests/test_ollama_warmup.py

@@ -0,0 +1,101 @@
+"""Unit tests for remote ollama_warmup (stdlib helpers)."""
+
+from importlib.util import module_from_spec, spec_from_file_location
+from pathlib import Path
+
+_ROOT = Path(__file__).resolve().parents[1]
+_SPEC = spec_from_file_location(
+    "ollama_warmup_remote",
+    _ROOT / "src" / "gpu_rent" / "remote" / "ollama_warmup.py",
+)
+assert _SPEC and _SPEC.loader
+_mod = module_from_spec(_SPEC)
+_SPEC.loader.exec_module(_mod)
+
+
+class FakeResp:
+    def __init__(self, body: str):
+        self._body = body.encode("utf-8")
+
+    def read(self) -> bytes:
+        return self._body
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *a):
+        return False
+
+
+def test_parse_env_file(tmp_path):
+    path = tmp_path / ".gpu-rent-ollama.env"
+    path.write_text(
+        "# comment\nOLLAMA_KEEP_ALIVE=15m\nOLLAMA_CONTEXT_LENGTH=16384\n",
+        encoding="utf-8",
+    )
+    env = _mod.parse_env_file(path)
+    assert env["OLLAMA_KEEP_ALIVE"] == "15m"
+    assert env["OLLAMA_CONTEXT_LENGTH"] == "16384"
+    assert _mod.parse_env_file(tmp_path / "missing") == {}
+
+
+def test_warmup_skips_when_ps_has_model(monkeypatch):
+    def fake_urlopen(req, timeout=None):
+        url = getattr(req, "full_url", str(req))
+        assert "/api/ps" in url
+        return FakeResp('{"models":[{"name":"huihui_ai/qwen2.5-vl-abliterated:7b"}]}')
+
+    monkeypatch.setattr(_mod.urllib.request, "urlopen", fake_urlopen)
+    msg = _mod.warmup("huihui_ai/qwen2.5-vl-abliterated:7b", keep_alive="15m", num_ctx=16384, timeout=5)
+    assert "skip" in msg
+    assert "VRAM" in msg
+
+
+def test_warmup_posts_one_token_chat(monkeypatch):
+    seen: dict = {}
+
+    def fake_urlopen(req, timeout=None):
+        url = getattr(req, "full_url", str(req))
+        method = getattr(req, "method", "GET")
+        if "/api/ps" in url:
+            return FakeResp('{"models":[]}')
+        seen["url"] = url
+        seen["method"] = method
+        seen["body"] = req.data
+        return FakeResp('{"message":{"role":"assistant","content":"ok"}}')
+
+    monkeypatch.setattr(_mod.urllib.request, "urlopen", fake_urlopen)
+    msg = _mod.warmup("foo:7b", keep_alive="15m", num_ctx=16384, timeout=5)
+    assert msg.startswith("ok foo:7b")
+    assert "/api/chat" in seen["url"]
+    assert seen["method"] == "POST"
+    import json
+
+    body = json.loads(seen["body"].decode())
+    assert body["model"] == "foo:7b"
+    assert body["stream"] is False
+    assert body["keep_alive"] == "15m"
+    assert body["options"]["num_predict"] == 1
+    assert body["options"]["num_ctx"] == 16384
+    assert body["messages"][0]["content"] == "ok"
+
+
+def test_main_never_fails_on_http_error(monkeypatch, tmp_path):
+    jobs = tmp_path / "jobs.json"
+    jobs.write_text('{"model":"foo:7b"}', encoding="utf-8")
+    monkeypatch.setattr(_mod, "JOBS", jobs)
+    monkeypatch.setattr(_mod, "ENV_FILE", tmp_path / "missing.env")
+
+    def boom(*_a, **_k):
+        raise _mod.urllib.error.URLError("down")
+
+    monkeypatch.setattr(_mod.urllib.request, "urlopen", boom)
+    assert _mod.main() == 0
+
+
+def test_warmup_body_matches_assistent_ctx():
+    import json
+
+    body = json.loads(_mod.warmup_body("m:7b", keep_alive="15m", num_ctx=16384))
+    assert body["options"]["num_ctx"] == 16384
+    assert body["options"]["num_predict"] == 1

tests/test_perf_tiers.py

@@ -82,4 +82,6 @@ def test_ollama_mid_4090_context_16k():
     )
     tune = ollama_tune_for(info)
     assert tune.context_length == 16384
+    assert tune.keep_alive == "15m"
     assert "OLLAMA_CONTEXT_LENGTH=16384" in "\n".join(ollama_env_lines(tune))
+    assert "OLLAMA_KEEP_ALIVE=15m" in "\n".join(ollama_env_lines(tune))

tests/test_tunnel_watch.py

@@ -141,6 +141,7 @@ def _stub_tunnel(monkeypatch) -> None:
     monkeypatch.setattr("gpu_rent.tunnel._ssh_tunnel_forwarder", lambda: object)
     monkeypatch.setattr("gpu_rent.tunnel._start_forwarder", lambda *a, **k: _Fwd())
     monkeypatch.setattr("gpu_rent.ready.verify_stack_local", lambda *a, **k: [])
+    monkeypatch.setattr("gpu_rent.llm_runtime.maybe_warmup_ollama_local", lambda *a, **k: None)
     monkeypatch.setattr("gpu_rent.tunnel.load_state", lambda: st)
     monkeypatch.setattr("gpu_rent.tunnel.save_state", lambda s: None)
     monkeypatch.setattr("gpu_rent.access_card.print_access_card", lambda *a, **k: None)

tests/test_warm_up_skips.py

@@ -86,6 +86,7 @@ def test_install_ollama_skips_restart_when_unit_unchanged():
     text = files("gpu_rent.remote").joinpath("install_ollama.sh").read_text(encoding="utf-8")
     assert "cmp -s" in text
     assert "skip restart" in text
+    assert '"mid", 10 * 1024**3, "15m"' in text
 
 
 def test_provision_llm_skips_on_api_tags_not_cli_list():
@@ -97,6 +98,7 @@ def test_provision_llm_skips_on_api_tags_not_cli_list():
     assert "_ollama_api_tags" in text
     assert "awk 'NR>1" not in text
     assert "GPU не гасим" in text
+    assert "ollama_warmup.py" in text
     assert "без моделей из ollama-models.yaml" not in text