Support Ollama use: chat|memory and parallel embed beside VL.

Pull nomic-embed-text for Assistent memory, write ollama-roles.json, CPU Modelfile, and raise MAX_LOADED_MODELS/NUM_PARALLEL to 2.

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

tests/test_ollama_model_use.py

@@ -0,0 +1,55 @@
+"""Ollama manifest use: chat|memory parsing."""
+
+from pathlib import Path
+
+from gpu_rent.llm_runtime import (
+    MEMORY_EMBED_MODEL,
+    ensure_memory_model_entries,
+    ollama_roles_payload,
+    parse_ollama_models,
+    preferred_ollama_model,
+    write_ollama_models_preset,
+)
+
+
+def test_parse_use_chat_and_memory(tmp_path: Path):
+    path = tmp_path / "ollama-models.yaml"
+    path.write_text(
+        """
+models:
+  - name: chat-model:7b
+    use: chat
+    default: true
+  - name: nomic-embed-text
+    use: memory
+""",
+        encoding="utf-8",
+    )
+    entries = parse_ollama_models(path)
+    assert len(entries) == 2
+    assert entries[0].use == "chat" and entries[0].default
+    assert entries[1].use == "memory" and not entries[1].default
+    assert preferred_ollama_model(path) == "chat-model:7b"
+    roles = ollama_roles_payload(entries)
+    assert roles["chat"] == ["chat-model:7b"]
+    assert roles["memory"] == ["nomic-embed-text"]
+
+
+def test_ensure_memory_appended():
+    from gpu_rent.llm_runtime import OllamaModelEntry
+
+    entries = [OllamaModelEntry(name="vl:7b", default=True, use="chat")]
+    out = ensure_memory_model_entries(entries)
+    assert any(e.use == "memory" and e.name == MEMORY_EMBED_MODEL for e in out)
+
+
+def test_write_preset_includes_memory(tmp_path: Path):
+    path = tmp_path / "m.yaml"
+    write_ollama_models_preset(path, "recommended")
+    text = path.read_text(encoding="utf-8")
+    assert "use: chat" in text
+    assert "use: memory" in text
+    assert MEMORY_EMBED_MODEL in text
+    entries = parse_ollama_models(path)
+    assert preferred_ollama_model(path) is not None
+    assert preferred_ollama_model(path) != MEMORY_EMBED_MODEL

tests/test_perf_tiers.py

@@ -34,8 +34,8 @@ def test_ollama_high_reserves_vram_for_swarm():
     )
     tune = ollama_tune_for(info)
     assert tune.flash_attention
-    assert tune.num_parallel == 1
-    assert tune.max_loaded_models == 1
+    assert tune.num_parallel == 2
+    assert tune.max_loaded_models == 2
     assert tune.kv_cache_type == "q8_0"
     assert tune.gpu_overhead_bytes == 14 * 1024**3
     env = "\n".join(ollama_env_lines(tune))