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>

src/gpu_rent/provision.py

@@ -1045,6 +1045,33 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
                     + f" (есть: {sorted(have) or 'пусто'}). "
                     "SwarmUI ок — GPU не гасим; Assistent будет пустой."
                 )
+            warm = next(
+                (
+                    n
+                    for n in list(defaults) + names
+                    if already_have_ollama_tag(have, n)
+                ),
+                "",
+            )
+            if warm:
+                log(f"Ollama warmup {warm} (гружу в VRAM)")
+                put_text(
+                    cfg,
+                    host,
+                    "/tmp/gpu-rent-ollama-warmup.json",
+                    json.dumps({"model": warm}, indent=2),
+                )
+                try:
+                    run_python(
+                        cfg,
+                        host,
+                        _pkg_text("ollama_warmup.py"),
+                        remote_path="/tmp/gpu-rent-ollama_warmup.py",
+                        timeout=240,
+                        log=log,
+                    )
+                except Exception as exc:
+                    log(f"⚠ Ollama warmup: {exc}")
     else:
         raise CloudError(f"неизвестный LLM_RUNTIME={runtime!r}")
     st = load_state()