Add configurable LM Studio flushing and LLM retries.

test/test_lm_studio.py

@@ -0,0 +1,102 @@
+import json
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from photools3.lm_studio import LMStudioError, LMStudioModelManager
+
+
+def _http_response(payload: dict):
+    response = MagicMock()
+    response.__enter__.return_value = response
+    response.read.return_value = json.dumps(payload).encode("utf-8")
+    return response
+
+
+def test_reload_model_unloads_and_reloads_preserving_context_config():
+    model_response = {
+        "models": [
+            {
+                "key": "qwen/qwen3-vl-4b",
+                "loaded_instances": [
+                    {
+                        "id": "qwen/qwen3-vl-4b-instance",
+                        "config": {
+                            "context_length": 16384,
+                            "eval_batch_size": 256,
+                            "flash_attention": True,
+                            "offload_kv_cache_to_gpu": False,
+                            "parallel": 2,
+                        },
+                    }
+                ],
+            }
+        ]
+    }
+    with patch(
+        "photools3.lm_studio.urlopen",
+        side_effect=[
+            _http_response(model_response),
+            _http_response({"instance_id": "qwen/qwen3-vl-4b-instance"}),
+            _http_response({"instance_id": "qwen/qwen3-vl-4b-instance"}),
+        ],
+    ) as mocked_urlopen:
+        manager = LMStudioModelManager(
+            "http://localhost:1234/v1", "test-token", timeout=12
+        )
+        manager.reload_model("qwen/qwen3-vl-4b")
+
+    requests = [call.args[0] for call in mocked_urlopen.call_args_list]
+    assert requests[0].full_url == "http://localhost:1234/api/v1/models"
+    assert requests[0].get_method() == "GET"
+    assert requests[0].get_header("Authorization") == "Bearer test-token"
+
+    assert requests[1].full_url.endswith("/models/unload")
+    assert json.loads(requests[1].data) == {"instance_id": "qwen/qwen3-vl-4b-instance"}
+
+    assert requests[2].full_url.endswith("/models/load")
+    assert json.loads(requests[2].data) == {
+        "model": "qwen/qwen3-vl-4b",
+        "context_length": 16384,
+        "eval_batch_size": 256,
+        "flash_attention": True,
+        "offload_kv_cache_to_gpu": False,
+    }
+
+
+def test_reload_model_loads_by_key_if_no_instance_is_loaded():
+    model_response = {"models": [{"key": "qwen/model", "loaded_instances": []}]}
+    with patch(
+        "photools3.lm_studio.urlopen",
+        side_effect=[
+            _http_response(model_response),
+            _http_response({"status": "loaded"}),
+        ],
+    ) as mocked_urlopen:
+        manager = LMStudioModelManager("http://localhost:1234/v1", "token")
+        manager.reload_model("qwen/model")
+
+    assert mocked_urlopen.call_count == 2
+    load_request = mocked_urlopen.call_args.args[0]
+    assert load_request.full_url.endswith("/models/load")
+    assert json.loads(load_request.data) == {"model": "qwen/model"}
+
+
+def test_reload_model_rejects_ambiguous_model_selection():
+    model_response = {
+        "models": [
+            {
+                "key": "qwen/model",
+                "loaded_instances": [
+                    {"id": "instance-1", "config": {}},
+                    {"id": "instance-2", "config": {}},
+                ],
+            }
+        ]
+    }
+    with patch(
+        "photools3.lm_studio.urlopen", return_value=_http_response(model_response)
+    ):
+        manager = LMStudioModelManager("http://localhost:1234/v1", "token")
+        with pytest.raises(LMStudioError, match="Multiple loaded instances"):
+            manager.reload_model("qwen/model")