Fix local repos with remote code not registering for pipelines (#33100)

* Extremely experimental fix!

* Try removing the clause entirely

* Add test

* make fixup

* stash commit

* Remove breakpoint

* Add anti-regression test

* make fixup

* Move repos to hf-internal-testing!

src/transformers/models/auto/auto_factory.py

@@ -17,7 +17,6 @@
 import copy
 import importlib
 import json
-import os
 import warnings
 from collections import OrderedDict
 
@@ -427,9 +426,6 @@ class _BaseAutoModelClass:
             else:
                 repo_id = config.name_or_path
             model_class = get_class_from_dynamic_module(class_ref, repo_id, **kwargs)
-            if os.path.isdir(config._name_or_path):
-                model_class.register_for_auto_class(cls.__name__)
-            else:
             cls.register(config.__class__, model_class, exist_ok=True)
             _ = kwargs.pop("code_revision", None)
             return model_class._from_config(config, **kwargs)
@@ -552,9 +548,6 @@ class _BaseAutoModelClass:
                 class_ref, pretrained_model_name_or_path, code_revision=code_revision, **hub_kwargs, **kwargs
             )
             _ = hub_kwargs.pop("code_revision", None)
-            if os.path.isdir(pretrained_model_name_or_path):
-                model_class.register_for_auto_class(cls.__name__)
-            else:
             cls.register(config.__class__, model_class, exist_ok=True)
             return model_class.from_pretrained(
                 pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs

tests/models/auto/test_modeling_auto.py

@@ -21,6 +21,7 @@ from collections import OrderedDict
 from pathlib import Path
 
 import pytest
+from huggingface_hub import Repository
 
 import transformers
 from transformers import BertConfig, GPT2Model, is_safetensors_available, is_torch_available
@@ -529,3 +530,12 @@ class AutoModelTest(unittest.TestCase):
         _MODEL_MAPPING_NAMES = OrderedDict([("bert", "GPT2Model")])
         _MODEL_MAPPING = _LazyAutoMapping(_CONFIG_MAPPING_NAMES, _MODEL_MAPPING_NAMES)
         self.assertEqual(_MODEL_MAPPING[BertConfig], GPT2Model)
+
+    def test_dynamic_saving_from_local_repo(self):
+        with tempfile.TemporaryDirectory() as tmp_dir, tempfile.TemporaryDirectory() as tmp_dir_out:
+            _ = Repository(local_dir=tmp_dir, clone_from="hf-internal-testing/tiny-random-custom-architecture")
+            model = AutoModelForCausalLM.from_pretrained(tmp_dir, trust_remote_code=True)
+            model.save_pretrained(tmp_dir_out)
+            _ = AutoModelForCausalLM.from_pretrained(tmp_dir_out, trust_remote_code=True)
+            self.assertTrue((Path(tmp_dir_out) / "modeling_fake_custom.py").is_file())
+            self.assertTrue((Path(tmp_dir_out) / "configuration_fake_custom.py").is_file())

tests/pipelines/test_pipelines_common.py

@@ -22,7 +22,7 @@ from pathlib import Path
 
 import datasets
 import numpy as np
-from huggingface_hub import HfFolder, delete_repo
+from huggingface_hub import HfFolder, Repository, delete_repo
 from requests.exceptions import HTTPError
 
 from transformers import (
@@ -226,6 +226,14 @@ class CommonPipelineTest(unittest.TestCase):
         pipe.model = None
         self.assertIsNone(pipe.torch_dtype)
 
+    @require_torch
+    def test_auto_model_pipeline_registration_from_local_dir(self):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            _ = Repository(local_dir=tmp_dir, clone_from="hf-internal-testing/tiny-random-custom-architecture")
+            pipe = pipeline("text-generation", tmp_dir, trust_remote_code=True)
+
+            self.assertIsInstance(pipe, TextGenerationPipeline)  # Assert successful load
+
 
 @is_pipeline_test
 class PipelineScikitCompatTest(unittest.TestCase):