stash commit

src/transformers/tokenization_utils_base.py

@@ -1604,6 +1604,7 @@ def __init__(self, **kwargs):
 
         # Stores a Jinja template that formats chat histories into tokenizable strings
         self.chat_template = kwargs.pop("chat_template", None)
+        self.inverse_chat_template = kwargs.pop("inverse_chat_template", None)
         if isinstance(self.chat_template, (list, tuple)):
             # Chat templates are stored as lists of dicts with fixed key names,
             # we reconstruct that into a single dict while loading them.
@@ -1881,6 +1882,33 @@ def apply_chat_template(
         else:
             return rendered
 
+    def apply_inverse_chat_template(self, formatted_chat, inverse_chat_template=None):
+        """
+        This method performs the inverse of the apply_chat_template method. In other words, it converts a formatted
+        string back into the arguments to apply_chat_template that would have created it: The chat history as a list
+        of dicts, and optionally other inputs like tools.
+
+        Note that inverting a chat template is not trivial, and so we require a separate Jinja inverse template
+        to be defined. This template takes a string as input, and yield the formatted args in JSON format.
+
+        Args:
+            formatted_chat (`str`): The formatted chat history to convert back into arguments.
+            inverse_chat_template (`str`, *optional*): A Jinja template to use for this conversion. If not provided, the
+                tokenizer's inverse_chat_template attribute will be used.
+        """
+
+        if inverse_chat_template is None:
+            if self.inverse_chat_template is not None:
+                inverse_chat_template = self.inverse_chat_template
+            else:
+                raise ValueError("No inverse chat template set, cannot use apply_inverse_chat_template!")
+
+        # Compilation function uses a cache to avoid recompiling the same template
+        compiled_template = self._compile_jinja_template(inverse_chat_template)
+
+        return compiled_template.render(formatted_chat)
+
+
     @lru_cache
     def _compile_jinja_template(self, chat_template):
         try: