Return assistant generated tokens mask in apply_chat_template (#30650)

return assistant generated tokens mask in apply_chat_template

src/transformers/tokenization_utils_base.py

@@ -1697,6 +1697,7 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
         max_length: Optional[int] = None,
         return_tensors: Optional[Union[str, TensorType]] = None,
         return_dict: bool = False,
+        return_assistant_tokens_mask: bool = False,
         tokenizer_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs,
     ) -> Union[str, List[int], List[str], List[List[int]], BatchEncoding]:
@@ -1747,6 +1748,10 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
             return_dict (`bool`, defaults to `False`):
                 Whether to return a dictionary with named outputs. Has no effect if tokenize is `False`.
             tokenizer_kwargs (`Dict[str: Any]`, *optional*): Additional kwargs to pass to the tokenizer.
+            return_assistant_tokens_mask (`bool`, defaults to `False`):
+                Whether to return a mask of the assistant generated tokens. For tokens generated by the assistant,
+                the mask will contain 1. For user and system tokens, the mask will contain 0.
+                This functionality is only available for chat templates that support it via the `{% generation %}` keyword.
             **kwargs: Additional kwargs to pass to the template renderer. Will be accessible by the chat template.
 
         Returns:
@@ -1761,6 +1766,9 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
                 "of tokenizer outputs to return."
             )
 
+        if return_assistant_tokens_mask and not return_dict:
+            raise ValueError("`return_assistant_tokens_mask=True` is incompatible with `return_dict=False`")
+
         if tokenizer_kwargs is None:
             tokenizer_kwargs = {}
 
@@ -1813,6 +1821,11 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
                 "then to ensure that this model continues working without issues."
             )
 
+        if return_assistant_tokens_mask and not re.search(r"\{\%-?\s*generation\s*-?\%\}", chat_template):
+            logger.warning_once(
+                "return_assistant_tokens_mask==True but chat template does not contain `{% generation %}` keyword."
+            )
+
         # Compilation function uses a cache to avoid recompiling the same template
         compiled_template = self._compile_jinja_template(chat_template)
 
@@ -1847,11 +1860,23 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
                     raise TypeError("Documents should be a list of dicts with 'title' and 'text' keys!")
 
         rendered = []
+        all_generation_indices = []
         template_kwargs = {**self.special_tokens_map, **kwargs}  # kwargs overwrite special tokens if both are present
         for chat in conversations:
             if hasattr(chat, "messages"):
                 # Indicates it's a Conversation object
                 chat = chat.messages
+            if return_assistant_tokens_mask:
+                rendered_chat, generation_indices = self._render_with_assistant_indices(
+                    compiled_template=compiled_template,
+                    messages=chat,
+                    tools=tool_schemas,
+                    documents=documents,
+                    add_generation_prompt=add_generation_prompt,
+                    **template_kwargs,
+                )
+                all_generation_indices.append(generation_indices)
+            else:
                 rendered_chat = compiled_template.render(
                     messages=chat,
                     tools=tool_schemas,
@@ -1875,17 +1900,54 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
                 **tokenizer_kwargs,
             )
             if return_dict:
+                if return_assistant_tokens_mask:
+                    assistant_masks = []
+                    if is_batched or return_tensors:
+                        input_ids = out["input_ids"]
+                    else:
+                        input_ids = [out["input_ids"]]
+                    for i in range(len(input_ids)):
+                        current_mask = [0] * len(input_ids[i])
+                        for assistant_start_char, assistant_end_char in all_generation_indices[i]:
+                            start_token = out.char_to_token(i, assistant_start_char)
+                            end_token = out.char_to_token(i, assistant_end_char - 1)
+                            if start_token is None:
+                                # start_token is out of bounds maybe due to truncation.
+                                break
+                            for token_id in range(start_token, end_token + 1 if end_token else len(input_ids)):
+                                current_mask[token_id] = 1
+                        assistant_masks.append(current_mask)
+                    out["assistant_masks"] = assistant_masks if is_batched else assistant_masks[0]
                 return out
             else:
                 return out["input_ids"]
         else:
             return rendered
 
+    def _render_with_assistant_indices(
+        self, compiled_template, messages, tools, documents, add_generation_prompt, **template_kwargs
+    ):
+        rendered_blocks = []
+        generation_indices = []
+        with compiled_template.environment.activate_tracker(rendered_blocks, generation_indices):
+            for block in compiled_template.generate(
+                messages=messages,
+                tools=tools,
+                documents=documents,
+                add_generation_prompt=add_generation_prompt,
+                **template_kwargs,
+            ):
+                rendered_blocks.append(block)
+            rendered_chat = "".join(rendered_blocks)
+        return rendered_chat, generation_indices
+
     @lru_cache
     def _compile_jinja_template(self, chat_template):
         try:
             import jinja2
+            from jinja2 import nodes
             from jinja2.exceptions import TemplateError
+            from jinja2.ext import Extension
             from jinja2.sandbox import ImmutableSandboxedEnvironment
         except ImportError:
             raise ImportError("apply_chat_template requires jinja2 to be installed.")
@@ -1903,7 +1965,49 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
             # We also expose some options like custom indents and separators
             return json.dumps(x, ensure_ascii=ensure_ascii, indent=indent, separators=separators, sort_keys=sort_keys)
 
-        jinja_env = ImmutableSandboxedEnvironment(trim_blocks=True, lstrip_blocks=True)
+        class AssistantTracker(Extension):
+            # This extension is used to track the indices of assistant-generated tokens in the rendered chat
+            tags = {"generation"}
+
+            def __init__(self, environment: ImmutableSandboxedEnvironment):
+                # The class is only initiated by jinja.
+                super().__init__(environment)
+                environment.extend(activate_tracker=self.activate_tracker)
+                self._rendered_blocks = None
+                self._generation_indices = None
+
+            def parse(self, parser: jinja2.parser.Parser) -> jinja2.nodes.CallBlock:
+                lineno = next(parser.stream).lineno
+                body = parser.parse_statements(["name:endgeneration"], drop_needle=True)
+                return nodes.CallBlock(self.call_method("_generation_support"), [], [], body).set_lineno(lineno)
+
+            @jinja2.pass_eval_context
+            def _generation_support(self, context: jinja2.nodes.EvalContext, caller: jinja2.runtime.Macro) -> str:
+                rv = caller()
+                if self.is_active():
+                    # Only track generation indices if the tracker is active
+                    start_index = len("".join(self._rendered_blocks))
+                    end_index = start_index + len(rv)
+                    self._generation_indices.append((start_index, end_index))
+                return rv
+
+            def is_active(self) -> bool:
+                return self._rendered_blocks or self._generation_indices
+
+            @contextmanager
+            def activate_tracker(self, rendered_blocks: list[int], generation_indices: list[int]):
+                try:
+                    if self.is_active():
+                        raise ValueError("AssistantTracker should not be reused before closed")
+                    self._rendered_blocks = rendered_blocks
+                    self._generation_indices = generation_indices
+
+                    yield
+                finally:
+                    self._rendered_blocks = None
+                    self._generation_indices = None
+
+        jinja_env = ImmutableSandboxedEnvironment(trim_blocks=True, lstrip_blocks=True, extensions=[AssistantTracker])
         jinja_env.filters["tojson"] = tojson
         jinja_env.globals["raise_exception"] = raise_exception
         return jinja_env.from_string(chat_template)

tests/models/layoutlmv2/test_tokenization_layoutlmv2.py

@@ -2483,3 +2483,7 @@ class LayoutLMv2TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     @unittest.skip(reason="Chat is not supported")
     def test_chat_template(self):
         pass
+
+    @unittest.skip("Chat is not supported")
+    def test_chat_template_return_assistant_tokens_mask(self):
+        pass

tests/models/layoutlmv3/test_tokenization_layoutlmv3.py

@@ -2436,3 +2436,7 @@ class LayoutLMv3TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     @unittest.skip(reason="Chat is not supported")
     def test_chat_template(self):
         pass
+
+    @unittest.skip("Chat is not supported")
+    def test_chat_template_return_assistant_tokens_mask(self):
+        pass

tests/models/layoutxlm/test_tokenization_layoutxlm.py

@@ -1977,3 +1977,7 @@ class LayoutXLMTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     @unittest.skip(reason="Chat is not supported")
     def test_chat_template(self):
         pass
+
+    @unittest.skip("Chat is not supported")
+    def test_chat_template_return_assistant_tokens_mask(self):
+        pass

tests/models/markuplm/test_tokenization_markuplm.py

@@ -2316,3 +2316,7 @@ class MarkupLMTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     @unittest.skip(reason="The model tested fails `Hub -> Fast == Hub -> Slow`, nothing much we can do")
     def test_added_tokens_serialization(self):
         pass
+
+    @unittest.skip("Chat is not supported")
+    def test_chat_template_return_assistant_tokens_mask(self):
+        pass

tests/models/tapas/test_tokenization_tapas.py

@@ -1277,3 +1277,7 @@ class TapasTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     @unittest.skip(reason="Chat is not supported")
     def test_chat_template(self):
         pass
+
+    @unittest.skip("Chat is not supported")
+    def test_chat_template_return_assistant_tokens_mask(self):
+        pass

tests/models/udop/test_tokenization_udop.py

@@ -1157,6 +1157,10 @@ class UdopTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     def test_chat_template(self):
         pass
 
+    @unittest.skip(reason="Chat template tests don't play well with table/layout models.")
+    def test_chat_template_return_assistant_tokens_mask(self):
+        pass
+
     @unittest.skip(reason="Chat template tests don't play well with table/layout models.")
     def test_chat_template_batched(self):
         pass

tests/test_tokenization_common.py

@@ -1153,6 +1153,135 @@ class TokenizerTesterMixin:
                     dummy_conversations, chat_template=dummy_template, tokenize=True
                 )  # Check that no error raised
 
+    @require_jinja
+    def test_chat_template_return_assistant_tokens_mask(self):
+        dummy_template = (
+            "{% for message in messages %}"
+            "{% if (message['role'] != 'assistant') %}"
+            "{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}"
+            "{% elif (message['role'] == 'assistant')%}"
+            "{{'<|im_start|>' + message['role'] + '\n'}}"
+            "{% generation %}"
+            "{{message['content'] + '<|im_end|>'}}"
+            "{% endgeneration %}"
+            "{{'\n'}}"
+            "{% endif %}"
+            "{% endfor %}"
+        )
+        conversations = [
+            [
+                {"role": "system", "content": "system message"},
+                {"role": "user", "content": "user message"},
+                {"role": "assistant", "content": "start turn 1 assistant message. end turn 1"},
+                {"role": "user", "content": "user message 2"},
+                {"role": "assistant", "content": "start turn 2 assistant message. end turn 2"},
+            ],
+            [
+                {"role": "system", "content": "system message 3"},
+                {"role": "user", "content": "user message 3"},
+                {"role": "assistant", "content": "start turn 3 assistant message. end turn 3"},
+                {"role": "user", "content": "user message 4"},
+                {"role": "assistant", "content": "start turn 4 assistant message. end turn 4"},
+            ],
+        ]
+
+        # These are the prefix and suffix strings of all the assistant messages. Used to find the assistant substring
+        # in the entire chat string, and then find the corresponding tokens in the tokenized output.
+        assistant_prefix_suffix = [
+            [("start turn 1", "end turn 1<|im_end|>"), ("start turn 2", "end turn 2<|im_end|>")],
+            [("start turn 3", "end turn 3<|im_end|>"), ("start turn 4", "end turn 4<|im_end|>")],
+        ]
+        for tokenizer, pretrained_name, _ in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                if not self.test_rust_tokenizer:
+                    self.skipTest(reason="No fast tokenizer defined")
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name)
+
+                # check batched
+                output = tokenizer_r.apply_chat_template(
+                    conversations,
+                    chat_template=dummy_template,
+                    tokenize=True,
+                    return_assistant_tokens_mask=True,
+                    return_dict=True,
+                )
+                for i, conv in enumerate(conversations):
+                    chat_string = tokenizer_r.apply_chat_template(
+                        conversations[i], tokenize=False, chat_template=dummy_template
+                    )
+                    assistant_start = output.char_to_token(i, chat_string.index(assistant_prefix_suffix[i][0][0]))
+                    assistant_end = output.char_to_token(
+                        i,
+                        chat_string.index(assistant_prefix_suffix[i][0][1])
+                        + len(assistant_prefix_suffix[i][0][1])
+                        - 1,
+                    )
+
+                    assistant_start2 = output.char_to_token(i, chat_string.index(assistant_prefix_suffix[i][1][0]))
+                    assistant_end2 = output.char_to_token(
+                        i,
+                        chat_string.index(assistant_prefix_suffix[i][1][1])
+                        + len(assistant_prefix_suffix[i][1][1])
+                        - 1,
+                    )
+
+                    # assert 1 in first assistant message
+                    self.assertEqual(
+                        output["assistant_masks"][i][assistant_start : assistant_end + 1],
+                        [1] * (assistant_end - assistant_start + 1),
+                    )
+                    # assert 1 second assistant message
+                    self.assertEqual(
+                        output["assistant_masks"][i][assistant_start2 : assistant_end2 + 1],
+                        [1] * (assistant_end2 - assistant_start2 + 1),
+                    )
+
+                    # assert 0 in user/system indices
+                    self.assertEqual(output["assistant_masks"][i][:assistant_start], [0] * assistant_start)
+                    self.assertEqual(
+                        output["assistant_masks"][i][assistant_end + 1 : assistant_start2],
+                        [0] * (assistant_start2 - assistant_end - 1),
+                    )
+
+                # check not batched
+                output = tokenizer_r.apply_chat_template(
+                    conversations[0],
+                    chat_template=dummy_template,
+                    tokenize=True,
+                    return_assistant_tokens_mask=True,
+                    return_dict=True,
+                )
+
+                chat_string = tokenizer_r.apply_chat_template(
+                    conversations[0], tokenize=False, chat_template=dummy_template
+                )
+                assistant_start = output.char_to_token(0, chat_string.index(assistant_prefix_suffix[0][0][0]))
+                assistant_end = output.char_to_token(
+                    0, chat_string.index(assistant_prefix_suffix[0][0][1]) + len(assistant_prefix_suffix[0][0][1]) - 1
+                )
+                assistant_start2 = output.char_to_token(0, chat_string.index(assistant_prefix_suffix[0][1][0]))
+                assistant_end2 = output.char_to_token(
+                    0, chat_string.index(assistant_prefix_suffix[0][1][1]) + len(assistant_prefix_suffix[0][1][1]) - 1
+                )
+
+                # assert 1 in assistant indices
+                self.assertEqual(
+                    output["assistant_masks"][assistant_start : assistant_end + 1],
+                    [1] * (assistant_end - assistant_start + 1),
+                )
+                self.assertEqual(
+                    output["assistant_masks"][assistant_start2 : assistant_end2 + 1],
+                    [1] * (assistant_end2 - assistant_start2 + 1),
+                )
+
+                # assert 0 in user/system indices
+                self.assertEqual(output["assistant_masks"][:assistant_start], [0] * assistant_start)
+                self.assertEqual(
+                    output["assistant_masks"][assistant_end + 1 : assistant_start2],
+                    [0] * (assistant_start2 - assistant_end - 1),
+                )
+
     @require_jinja
     def test_chat_template_dict(self):
         dummy_template_1 = "{{'a'}}"