Honor trust_remote_code for custom tokenizers (#28854)

* pass through trust_remote_code for dynamically loading unregistered tokenizers specified by config
add test

* change directories back to previous directory after test

* fix ruff check

* Add a note to that block for future in case we want to remove it later

---------

Co-authored-by: Matt <rocketknight1@gmail.com>

src/transformers/models/auto/tokenization_auto.py

@@ -800,7 +800,9 @@ class AutoTokenizer:
             _ = kwargs.pop("code_revision", None)
             if os.path.isdir(pretrained_model_name_or_path):
                 tokenizer_class.register_for_auto_class()
-            return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+            return tokenizer_class.from_pretrained(
+                pretrained_model_name_or_path, *inputs, trust_remote_code=trust_remote_code, **kwargs
+            )
         elif config_tokenizer_class is not None:
             tokenizer_class = None
             if use_fast and not config_tokenizer_class.endswith("Fast"):

src/transformers/tokenization_utils_base.py

@@ -1810,6 +1810,7 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
         local_files_only: bool = False,
         token: Optional[Union[str, bool]] = None,
         revision: str = "main",
+        trust_remote_code=False,
         **kwargs,
     ):
         r"""
@@ -1853,6 +1854,10 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
                 facebook/rag-token-base), specify it here.
             inputs (additional positional arguments, *optional*):
                 Will be passed along to the Tokenizer `__init__` method.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
             kwargs (additional keyword arguments, *optional*):
                 Will be passed to the Tokenizer `__init__` method. Can be used to set special tokens like `bos_token`,
                 `eos_token`, `unk_token`, `sep_token`, `pad_token`, `cls_token`, `mask_token`,
@@ -2036,6 +2041,7 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
             local_files_only=local_files_only,
             _commit_hash=commit_hash,
             _is_local=is_local,
+            trust_remote_code=trust_remote_code,
             **kwargs,
         )
 
@@ -2051,6 +2057,7 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
         local_files_only=False,
         _commit_hash=None,
         _is_local=False,
+        trust_remote_code=False,
         **kwargs,
     ):
         # We instantiate fast tokenizers based on a slow tokenizer if we don't have access to the tokenizer.json
@@ -2099,6 +2106,10 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
             )
 
         if config_tokenizer_class is None:
+            # Matt: This entire block is only used to decide if the tokenizer class matches the class in the repo.
+            #       If not, it raises a warning, but otherwise continues. Since we mostly load tokenizers with
+            #       AutoTokenizer these days, it seems like a lot of work (and a source of bugs) for little gain.
+            #       Maybe we can just remove this entirely?
             from .models.auto.configuration_auto import AutoConfig  # tests_ignore
 
             # Second attempt. If we have not yet found tokenizer_class, let's try to use the config.
@@ -2108,6 +2119,7 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
                     token=token,
                     cache_dir=cache_dir,
                     local_files_only=local_files_only,
+                    trust_remote_code=trust_remote_code,
                     _commit_hash=_commit_hash,
                 )
                 config_tokenizer_class = config.tokenizer_class

tests/models/auto/test_tokenization_auto.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import json
 import os
 import shutil
 import sys
@@ -429,3 +430,73 @@ class AutoTokenizerTest(unittest.TestCase):
         self.assertEqual(counter["GET"], 0)
         self.assertEqual(counter["HEAD"], 1)
         self.assertEqual(counter.total_calls, 1)
+
+    def test_init_tokenizer_with_trust(self):
+        nop_tokenizer_code = """
+import transformers
+
+class NopTokenizer(transformers.PreTrainedTokenizer):
+    def get_vocab(self):
+        return {}
+"""
+
+        nop_config_code = """
+from transformers import PretrainedConfig
+
+class NopConfig(PretrainedConfig):
+    model_type = "test_unregistered_dynamic"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+"""
+
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            fake_model_id = "hf-internal-testing/test_unregistered_dynamic"
+            fake_repo = os.path.join(tmp_dir, fake_model_id)
+            os.makedirs(fake_repo)
+
+            tokenizer_src_file = os.path.join(fake_repo, "tokenizer.py")
+            with open(tokenizer_src_file, "w") as wfp:
+                wfp.write(nop_tokenizer_code)
+
+            model_config_src_file = os.path.join(fake_repo, "config.py")
+            with open(model_config_src_file, "w") as wfp:
+                wfp.write(nop_config_code)
+
+            config = {
+                "model_type": "test_unregistered_dynamic",
+                "auto_map": {"AutoConfig": f"{fake_model_id}--config.NopConfig"},
+            }
+
+            config_file = os.path.join(fake_repo, "config.json")
+            with open(config_file, "w") as wfp:
+                json.dump(config, wfp, indent=2)
+
+            tokenizer_config = {
+                "auto_map": {
+                    "AutoTokenizer": [
+                        f"{fake_model_id}--tokenizer.NopTokenizer",
+                        None,
+                    ]
+                }
+            }
+
+            tokenizer_config_file = os.path.join(fake_repo, "tokenizer_config.json")
+            with open(tokenizer_config_file, "w") as wfp:
+                json.dump(tokenizer_config, wfp, indent=2)
+
+            prev_dir = os.getcwd()
+            try:
+                # it looks like subdir= is broken in the from_pretrained also, so this is necessary
+                os.chdir(tmp_dir)
+
+                # this should work because we trust the code
+                _ = AutoTokenizer.from_pretrained(fake_model_id, local_files_only=True, trust_remote_code=True)
+                try:
+                    # this should fail because we don't trust and we're not at a terminal for interactive response
+                    _ = AutoTokenizer.from_pretrained(fake_model_id, local_files_only=True, trust_remote_code=False)
+                    self.fail("AutoTokenizer.from_pretrained with trust_remote_code=False should raise ValueException")
+                except ValueError:
+                    pass
+            finally:
+                os.chdir(prev_dir)