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>
2026-08-21 20:47:14 +03:00
parent 79cb0a7e25
commit ffe5a031b4
12 changed files with 465 additions and 4 deletions
@@ -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))