Fix LFM2.5 pythonic tool parser auto-detection

mlx_lm/tokenizer_utils.py

@@ -559,7 +559,9 @@ def _infer_tool_parser(chat_template):
         return "longcat"
     elif "<arg_key>" in chat_template:
         return "glm47"
-    elif "<|tool_list_start|>" in chat_template:
+    elif (
+        "<|tool_list_start|>" in chat_template or "<|tool_call_start|>" in chat_template
+    ):
         return "pythonic"
     elif (
         "<tool_call>\\n<function=" in chat_template

tests/test_tokenizers.py

@@ -9,6 +9,7 @@
     BPEStreamingDetokenizer,
     NaiveStreamingDetokenizer,
     SPMStreamingDetokenizer,
+    _infer_tool_parser,
 )
 from mlx_lm.utils import load_tokenizer
 
@@ -109,6 +110,30 @@ def test_thinking(self):
         self.assertIsNone(tokenizer.think_start_id)
         self.assertIsNone(tokenizer.think_end_id)
 
+    def test_infer_tool_parser(self):
+        # LFM2 (original) uses <|tool_list_start|> ... <|tool_list_end|>.
+        lfm2_template = (
+            "{%- if tools %}<|tool_list_start|>{{ tools | tojson }}"
+            "<|tool_list_end|>{% endif -%}"
+        )
+        self.assertEqual(_infer_tool_parser(lfm2_template), "pythonic")
+
+        # LFM2.5 renamed the tokens to <|tool_call_start|> ... <|tool_call_end|>
+        # while keeping the same pythonic call format inside.
+        # See https://huggingface.co/LiquidAI/LFM2.5-8B-A1B for the template.
+        lfm25_template = (
+            "{%- macro render_tool_calls(tool_calls) -%}"
+            "{{- '<|tool_call_start|>[' + "
+            "(tool_calls_ns.tool_calls | join(', ')) + "
+            "']<|tool_call_end|>' -}}"
+            "{%- endmacro -%}"
+        )
+        self.assertEqual(_infer_tool_parser(lfm25_template), "pythonic")
+
+        # Sanity: a template with neither token should not match pythonic.
+        self.assertIsNone(_infer_tool_parser("plain template, no tool tokens"))
+        self.assertIsNone(_infer_tool_parser(None))
+
 
 if __name__ == "__main__":
     unittest.main()