Add configurable LM Studio flushing and LLM retries.

test/test_llm_session_mocked.py

@@ -1,5 +1,7 @@
 from unittest.mock import MagicMock, patch
 
+import pytest
+
 from photools3.llm_session import LlmSession
 
 
@@ -84,3 +86,70 @@ def test_llm_session_captions_and_keywords_mocked(test_photo):
 
         # Verify that the mock was called
         assert mock_client.chat.completions.create.called
+
+
+def test_llm_session_retries_from_original_request_without_accumulating_messages():
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        def response(content):
+            message = MagicMock()
+            message.content = content
+            choice = MagicMock()
+            choice.message = message
+            result = MagicMock()
+            result.choices = [choice]
+            return result
+
+        mock_client.chat.completions.create.side_effect = [
+            response('{"wrong": "first"}'),
+            response('{"wrong": "second"}'),
+            response('{"text": "correct"}'),
+        ]
+
+        session = LlmSession(system_prompt="You are a helper.")
+        result = session.call(photos=[], prompt="Return JSON")
+
+    assert result.text == "correct"
+    requests = mock_client.chat.completions.create.call_args_list
+    messages_per_attempt = [call.kwargs["messages"] for call in requests]
+    assert [len(messages) for messages in messages_per_attempt] == [2, 3, 3]
+    final_correction = messages_per_attempt[2][-1]["content"]
+    assert '"wrong": "second"' in final_correction
+    assert '"wrong": "first"' not in final_correction
+
+
+def test_llm_session_uses_configured_retry_count(monkeypatch):
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_retries", 1)
+
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        def response(content):
+            message = MagicMock()
+            message.content = content
+            choice = MagicMock()
+            choice.message = message
+            result = MagicMock()
+            result.choices = [choice]
+            return result
+
+        mock_client.chat.completions.create.side_effect = [
+            response('{"wrong": "first"}'),
+            response('{"wrong": "second"}'),
+            response('{"text": "would be a third call"}'),
+        ]
+        session = LlmSession(system_prompt="You are a helper.")
+
+        with pytest.raises(ValueError, match="after 1 retries"):
+            session.call(photos=[])
+
+    assert mock_client.chat.completions.create.call_count == 2

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")

test/test_processing.py

@@ -127,3 +127,39 @@ def test_process_batch_is_sequential_reuses_session_and_continues_after_error():
     assert all(len(call.args[0]) == 1 for call in session.call.call_args_list)
     assert writer_class.return_value.write.call_count == 2
     session.close.assert_called_once()
+
+
+def test_process_batch_flushes_after_configured_attempt_count(monkeypatch):
+    monkeypatch.setattr("photools3.processing.settings.llm_flush_every_photos", 2)
+
+    class FakePhoto:
+        def __init__(self, path):
+            self.path = path
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, exc_type, exc_value, traceback):
+            return None
+
+    response = LlmResponseCaptionsAndKeywords(caption="Ein Haus", tags=["architektur"])
+    session = MagicMock()
+    session.call.side_effect = [
+        response,
+        RuntimeError("temporary error"),
+        response,
+        response,
+        response,
+    ]
+    paths = [Path(f"{i}.jpg") for i in range(5)]
+
+    with (
+        patch("photools3.processing._create_session", return_value=session),
+        patch("photools3.processing.Photo", FakePhoto),
+        patch("photools3.processing.MetadataWriter"),
+    ):
+        outcomes = process_batch(paths, ProcessingMode.BOTH)
+
+    assert len(outcomes) == len(paths)
+    assert session.flush_model.call_count == 2
+    session.close.assert_called_once()