From c1a88dd9a33f746f434ce982dfe6d0221c39dbb8 Mon Sep 17 00:00:00 2001
From: Aditya Raj Singh <aditya@bncw.in>
Date: Fri, 28 Aug 2026 08:34:40 +0530
Subject: [PATCH] fix(ml): read CLIP model configs as UTF-8 (#31075)

---
 .../immich_ml/models/clip/textual.py          |  6 ++--
 .../immich_ml/models/clip/visual.py           |  4 +--
 machine-learning/test_main.py                 | 32 +++++++++++++++++++
 3 files changed, 37 insertions(+), 5 deletions(-)

diff --git a/machine-learning/immich_ml/models/clip/textual.py b/machine-learning/immich_ml/models/clip/textual.py
index 989f46a2a..421459fc0 100644
--- a/machine-learning/immich_ml/models/clip/textual.py
+++ b/machine-learning/immich_ml/models/clip/textual.py
@@ -58,7 +58,7 @@ class BaseCLIPTextualEncoder(InferenceModel):
     @cached_property
     def model_cfg(self) -> dict[str, Any]:
         log.debug(f"Loading model config for CLIP model '{self.model_name}'")
-        model_cfg: dict[str, Any] = json.load(self.model_cfg_path.open())
+        model_cfg: dict[str, Any] = json.load(self.model_cfg_path.open(encoding="utf-8"))
         log.debug(f"Loaded model config for CLIP model '{self.model_name}'")
         return model_cfg
 
@@ -70,14 +70,14 @@ class BaseCLIPTextualEncoder(InferenceModel):
     @cached_property
     def tokenizer_file(self) -> dict[str, Any]:
         log.debug(f"Loading tokenizer file for CLIP model '{self.model_name}'")
-        tokenizer_file: dict[str, Any] = json.load(self.tokenizer_file_path.open())
+        tokenizer_file: dict[str, Any] = json.load(self.tokenizer_file_path.open(encoding="utf-8"))
         log.debug(f"Loaded tokenizer file for CLIP model '{self.model_name}'")
         return tokenizer_file
 
     @cached_property
     def tokenizer_cfg(self) -> dict[str, Any]:
         log.debug(f"Loading tokenizer config for CLIP model '{self.model_name}'")
-        tokenizer_cfg: dict[str, Any] = json.load(self.tokenizer_cfg_path.open())
+        tokenizer_cfg: dict[str, Any] = json.load(self.tokenizer_cfg_path.open(encoding="utf-8"))
         log.debug(f"Loaded tokenizer config for CLIP model '{self.model_name}'")
         return tokenizer_cfg
 
diff --git a/machine-learning/immich_ml/models/clip/visual.py b/machine-learning/immich_ml/models/clip/visual.py
index 16993f9c0..1d46d8fe1 100644
--- a/machine-learning/immich_ml/models/clip/visual.py
+++ b/machine-learning/immich_ml/models/clip/visual.py
@@ -46,14 +46,14 @@ class BaseCLIPVisualEncoder(InferenceModel):
     @cached_property
     def model_cfg(self) -> dict[str, Any]:
         log.debug(f"Loading model config for CLIP model '{self.model_name}'")
-        model_cfg: dict[str, Any] = json.load(self.model_cfg_path.open())
+        model_cfg: dict[str, Any] = json.load(self.model_cfg_path.open(encoding="utf-8"))
         log.debug(f"Loaded model config for CLIP model '{self.model_name}'")
         return model_cfg
 
     @cached_property
     def preprocess_cfg(self) -> dict[str, Any]:
         log.debug(f"Loading visual preprocessing config for CLIP model '{self.model_name}'")
-        preprocess_cfg: dict[str, Any] = json.load(self.preprocess_cfg_path.open())
+        preprocess_cfg: dict[str, Any] = json.load(self.preprocess_cfg_path.open(encoding="utf-8"))
         log.debug(f"Loaded visual preprocessing config for CLIP model '{self.model_name}'")
         return preprocess_cfg
 
diff --git a/machine-learning/test_main.py b/machine-learning/test_main.py
index 748a8840c..9e3fe50cb 100644
--- a/machine-learning/test_main.py
+++ b/machine-learning/test_main.py
@@ -617,6 +617,38 @@ class TestCLIP:
         assert len(embedding) == clip_model_cfg["embed_dim"]
         mocked.run.assert_called_once()
 
+    def test_reads_model_configs_as_utf8(self, mocker: MockerFixture, tmp_path: Path) -> None:
+        original_open = Path.open
+
+        def locale_default_is_ascii(self: Path, mode: str = "r", *args: Any, **kwargs: Any) -> Any:
+            if "b" not in mode and kwargs.get("encoding") is None:
+                kwargs["encoding"] = "ascii"
+            return original_open(self, mode, *args, **kwargs)
+
+        mocker.patch.object(OpenClipTextualEncoder, "download")
+        mocker.patch.object(OpenClipVisualEncoder, "download")
+
+        textual = OpenClipTextualEncoder("ViT-B-32__openai", cache_dir=tmp_path)
+        visual = OpenClipVisualEncoder("ViT-B-32__openai", cache_dir=tmp_path)
+        paths = [
+            textual.model_cfg_path,
+            textual.tokenizer_file_path,
+            textual.tokenizer_cfg_path,
+            visual.model_cfg_path,
+            visual.preprocess_cfg_path,
+        ]
+        for path in paths:
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_text(orjson.dumps({"eos_token": "<|café|>"}).decode(), encoding="utf-8")
+
+        mocker.patch.object(Path, "open", locale_default_is_ascii)
+
+        assert textual.model_cfg["eos_token"] == "<|café|>"
+        assert textual.tokenizer_file["eos_token"] == "<|café|>"
+        assert textual.tokenizer_cfg["eos_token"] == "<|café|>"
+        assert visual.model_cfg["eos_token"] == "<|café|>"
+        assert visual.preprocess_cfg["eos_token"] == "<|café|>"
+
     def test_openclip_tokenizer(
         self,
         mocker: MockerFixture,