Preserve standalone model chat templates

engine_compat.py

@@ -81,7 +81,14 @@ def repair_tokenizer_config(
         patched["tokenizer_class"] = "PreTrainedTokenizerFast"
         repairs.append("TokenizersBackend -> PreTrainedTokenizerFast")
 
-    if not patched.get("chat_template"):
+    # Transformers can load a standalone template with higher precedence than
+    # tokenizer_config.json. Do not synthesize one when the repository already
+    # supplies a template in that format.
+    has_standalone_template = any(
+        (model_dir / name).is_file()
+        for name in ("chat_template.jinja", "chat_template.json")
+    )
+    if not patched.get("chat_template") and not has_standalone_template:
         config_path = model_dir / "config.json"
         if config_path.is_file():
             model_config = _json_object(config_path)