Revert VLM support in parse_response

tests/test_chat_template_utils.py

@@ -716,7 +716,8 @@ def test_parse_response(self, model_name):
             prefix = prefix[0]
             text = text[0]
         response = text[len(prefix) :]
-        parsed = parse_response(processing_class, response)
+        tokenizer = processing_class.tokenizer if self.is_vlm else processing_class
+        parsed = parse_response(tokenizer, response)
         assert parsed == expected
 
     def test_parse_response_with_reasoning_content(self, model_name):
@@ -746,7 +747,8 @@ def test_parse_response_with_reasoning_content(self, model_name):
             prefix = prefix[0]
             text = text[0]
         response = text[len(prefix) :]
-        parsed = parse_response(processing_class, response)
+        tokenizer = processing_class.tokenizer if self.is_vlm else processing_class
+        parsed = parse_response(tokenizer, response)
         assert parsed == expected
 
     def test_parse_response_tool_call(self, model_name):
@@ -772,7 +774,8 @@ def test_parse_response_tool_call(self, model_name):
             prefix = prefix[0]
             text = text[0]
         response = text[len(prefix) :]
-        parsed = parse_response(processing_class, response)
+        tokenizer = processing_class.tokenizer if self.is_vlm else processing_class
+        parsed = parse_response(tokenizer, response)
         assert parsed == expected
 
     def test_parse_response_tool_call_with_content(self, model_name):
@@ -797,7 +800,8 @@ def test_parse_response_tool_call_with_content(self, model_name):
             prefix = prefix[0]
             text = text[0]
         response = text[len(prefix) :]
-        parsed = parse_response(processing_class, response)
+        tokenizer = processing_class.tokenizer if self.is_vlm else processing_class
+        parsed = parse_response(tokenizer, response)
         assert parsed == expected
 
     def test_parse_response_tool_call_without_arguments(self, model_name):
@@ -823,7 +827,8 @@ def test_parse_response_tool_call_without_arguments(self, model_name):
             prefix = prefix[0]
             text = text[0]
         response = text[len(prefix) :]
-        parsed = parse_response(processing_class, response)
+        tokenizer = processing_class.tokenizer if self.is_vlm else processing_class
+        parsed = parse_response(tokenizer, response)
         assert parsed == expected
 
     def test_parse_response_multiple_tool_calls(self, model_name):
@@ -858,7 +863,8 @@ def test_parse_response_multiple_tool_calls(self, model_name):
             prefix = prefix[0]
             text = text[0]
         response = text[len(prefix) :]
-        parsed = parse_response(processing_class, response)
+        tokenizer = processing_class.tokenizer if self.is_vlm else processing_class
+        parsed = parse_response(tokenizer, response)
         assert parsed == expected
 
     def test_parse_response_malformed_tool_call(self, model_name):

trl/chat_template_utils.py

@@ -685,7 +685,7 @@ def _validate_tool_calls(tool_calls: list | None) -> None:
                 tool_call["arguments"] = {}
 
 
-def parse_response(processing_class: PreTrainedTokenizerBase | ProcessorMixin, ids: list[int]) -> dict:
+def parse_response(tokenizer: PreTrainedTokenizerBase, ids: list[int]) -> dict:
     r"""
     Parse a token sequence into structured response dictionaries with fallback handling.
 
@@ -695,11 +695,9 @@ def parse_response(processing_class: PreTrainedTokenizerBase | ProcessorMixin, i
     Also removes incorrectly appended EOS tokens from tool call content when present, and validates tool_calls to
     ensure all required fields exist.
 
-    For VLM processors, automatically uses the inner tokenizer for parsing.
-
     Args:
-        processing_class (`PreTrainedTokenizerBase` or VLM processor):
-            Tokenizer or processor with a `parse_response()` method (directly or via inner tokenizer).
+        tokenizer (`PreTrainedTokenizerBase`):
+            Tokenizer with a `parse_response()` method.
         ids (`list[int]`):
             List of token sequences.
 
@@ -720,8 +718,6 @@ def parse_response(processing_class: PreTrainedTokenizerBase | ProcessorMixin, i
     {'role': 'assistant', 'content': '', 'tool_calls': [{'type': 'function', 'function': {'name': 'multiply', 'arguments': {'a': 3, 'b': 4}}}]}
     ```
     """
-    # VLM processors don't have parse_response directly; use the inner tokenizer
-    tokenizer = getattr(processing_class, "tokenizer", processing_class)
     try:
         parsed = tokenizer.parse_response(ids)
         # Hotfix: remove incorrectly appended EOS token from tool calls