SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Remote code improvements (#23959)

Browse Source 
* Fix model load when it has both code on the Hub and locally

* Add input check with timeout

* Add tests

* Apply suggestions from code review

Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>

* Some non-saved stuff

* Add feature extractors

* Add image processor

* Add model

* Add processor and tokenizer

* Reduce timeout

---------

Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>
This commit is contained in:
Sylvain Gugger
2023-06-06 14:31:14 -04:00
committed by GitHub
parent 60825f2c6e
commit f1660d7e23
13 changed files with 434 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

64
tests/models/auto/test_tokenization_auto.py
View File

@@ -22,6 +22,7 @@ from pathlib import Path
import pytest
import transformers
from transformers import (
BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -65,6 +66,9 @@ if is_tokenizers_available():
class AutoTokenizerTest(unittest.TestCase):
def setUp(self):
transformers.dynamic_module_utils.TIME_OUT_REMOTE_CODE = 0
@slow
def test_tokenizer_from_pretrained(self):
for model_name in (x for x in BERT_PRETRAINED_CONFIG_ARCHIVE_MAP.keys() if "japanese" not in x):
@@ -298,6 +302,15 @@ class AutoTokenizerTest(unittest.TestCase):
del TOKENIZER_MAPPING._extra_content[CustomConfig]
def test_from_pretrained_dynamic_tokenizer(self):
# If remote code is not set, we will time out when asking whether to load the model.
with self.assertRaises(ValueError):
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/test_dynamic_tokenizer")
# If remote code is disabled, we can't load this config.
with self.assertRaises(ValueError):
tokenizer = AutoTokenizer.from_pretrained(
"hf-internal-testing/test_dynamic_tokenizer", trust_remote_code=False
)
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/test_dynamic_tokenizer", trust_remote_code=True)
self.assertTrue(tokenizer.special_attribute_present)
# Test tokenizer can be reloaded.
@@ -326,6 +339,57 @@ class AutoTokenizerTest(unittest.TestCase):
self.assertEqual(tokenizer.__class__.__name__, "NewTokenizer")
self.assertEqual(reloaded_tokenizer.__class__.__name__, "NewTokenizer")
@require_tokenizers
def test_from_pretrained_dynamic_tokenizer_conflict(self):
class NewTokenizer(BertTokenizer):
special_attribute_present = False
class NewTokenizerFast(BertTokenizerFast):
slow_tokenizer_class = NewTokenizer
special_attribute_present = False
try:
AutoConfig.register("custom", CustomConfig)
AutoTokenizer.register(CustomConfig, slow_tokenizer_class=NewTokenizer)
AutoTokenizer.register(CustomConfig, fast_tokenizer_class=NewTokenizerFast)
# If remote code is not set, the default is to use local
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/test_dynamic_tokenizer")
self.assertEqual(tokenizer.__class__.__name__, "NewTokenizerFast")
self.assertFalse(tokenizer.special_attribute_present)
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/test_dynamic_tokenizer", use_fast=False)
self.assertEqual(tokenizer.__class__.__name__, "NewTokenizer")
self.assertFalse(tokenizer.special_attribute_present)
# If remote code is disabled, we load the local one.
tokenizer = AutoTokenizer.from_pretrained(
"hf-internal-testing/test_dynamic_tokenizer", trust_remote_code=False
)
self.assertEqual(tokenizer.__class__.__name__, "NewTokenizerFast")
self.assertFalse(tokenizer.special_attribute_present)
tokenizer = AutoTokenizer.from_pretrained(
"hf-internal-testing/test_dynamic_tokenizer", trust_remote_code=False, use_fast=False
)
self.assertEqual(tokenizer.__class__.__name__, "NewTokenizer")
self.assertFalse(tokenizer.special_attribute_present)
# If remote is enabled, we load from the Hub
tokenizer = AutoTokenizer.from_pretrained(
"hf-internal-testing/test_dynamic_tokenizer", trust_remote_code=True
)
self.assertEqual(tokenizer.__class__.__name__, "NewTokenizerFast")
self.assertTrue(tokenizer.special_attribute_present)
tokenizer = AutoTokenizer.from_pretrained(
"hf-internal-testing/test_dynamic_tokenizer", trust_remote_code=True, use_fast=False
)
self.assertEqual(tokenizer.__class__.__name__, "NewTokenizer")
self.assertTrue(tokenizer.special_attribute_present)
finally:
if "custom" in CONFIG_MAPPING._extra_content:
del CONFIG_MAPPING._extra_content["custom"]
if CustomConfig in TOKENIZER_MAPPING._extra_content:
del TOKENIZER_MAPPING._extra_content[CustomConfig]
def test_from_pretrained_dynamic_tokenizer_legacy_format(self):
tokenizer = AutoTokenizer.from_pretrained(
"hf-internal-testing/test_dynamic_tokenizer_legacy", trust_remote_code=True