fix(ml): preloaded model loaded twice (#32610)

fix(ml): reuse cached model across names that clean to the same model

(cherry picked from commit 166e93559f123313cf9cb4e7324c4ee30e21c618)

machine-learning/immich_ml/models/cache.py

@@ -3,6 +3,7 @@ from collections.abc import Hashable
 from typing import Any
 
 from immich_ml import allocator
+from immich_ml.config import clean_name
 from immich_ml.models.base import InferenceEntry, InferenceModel
 from immich_ml.schemas import Options
 
@@ -15,7 +16,8 @@ class ModelCache:
         self._expiries: dict[Key, asyncio.TimerHandle] = {}
 
     def get[O: Options](self, entry: InferenceEntry[O], ttl: int | None = None) -> InferenceModel[O]:
-        key = (entry.model, entry.name, entry.model.graph(entry.options))
+        # names that differ only in what cleaning drops, like an `immich-app/` prefix, are the same model
+        key = (entry.model, clean_name(entry.name), entry.model.graph(entry.options))
         if key not in self._models:
             self._models[key] = entry.model.create(entry.name, entry.options)
         elif key not in self._expiries:

machine-learning/test_main.py

@@ -1856,6 +1856,14 @@ class TestCache:
         assert len(model_cache._models) == 1
         model.create.assert_called_once()
 
+    async def test_shares_an_instance_across_names_of_the_same_model(self) -> None:
+        model_cache = ModelCache()
+        bare = model_cache.get(InferenceEntry(OpenClipTextualEncoder, "ViT-B-32__openai", TextualOptions()))
+        prefixed = model_cache.get(
+            InferenceEntry(OpenClipTextualEncoder, "immich-app/ViT-B-32__openai", TextualOptions(language="en-US"))
+        )
+        assert bare is prefixed
+
     async def test_separate_instances_per_graph_option(self) -> None:
         # the real TextDetector, so that keying it on anything but its resolution fails this
         model_cache = ModelCache()