offline mode for firewalled envs (part 2)

examples/research_projects/bertabs/run_summarization.py

@@ -9,8 +9,10 @@
 from torch.utils.data import DataLoader, SequentialSampler
 from tqdm import tqdm
 
+from filelock import FileLock
 from modeling_bertabs import BertAbs, build_predictor
 from transformers import BertTokenizer
+from transformers.file_utils import is_offline_mode
 
 from .utils_summarization import (
     CNNDMDataset,
@@ -48,7 +50,16 @@ def evaluate(args):
 
         import rouge
 
-        nltk.download("punkt")
+        try:
+            nltk.data.find("tokenizers/punkt")
+        except (LookupError, OSError):
+            if is_offline_mode():
+                raise LookupError(
+                    "Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
+                )
+            with FileLock(".lock") as lock:  # noqa
+                nltk.download("punkt", quiet=True)
+
         rouge_evaluator = rouge.Rouge(
             metrics=["rouge-n", "rouge-l"],
             max_n=2,

src/transformers/feature_extraction_utils.py

@@ -35,6 +35,7 @@
     cached_path,
     hf_bucket_url,
     is_flax_available,
+    is_offline_mode,
     is_remote_url,
     is_tf_available,
     is_torch_available,
@@ -342,6 +343,10 @@ def get_feature_extractor_dict(
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
 
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
         pretrained_model_name_or_path = str(pretrained_model_name_or_path)
         if os.path.isdir(pretrained_model_name_or_path):
             feature_extractor_file = os.path.join(pretrained_model_name_or_path, FEATURE_EXTRACTOR_NAME)

src/transformers/file_utils.py

@@ -1105,6 +1105,10 @@ def cached_path(
     if isinstance(cache_dir, Path):
         cache_dir = str(cache_dir)
 
+    if is_offline_mode() and not local_files_only:
+        logger.info("Offline mode: forcing local_files_only=True")
+        local_files_only = True
+
     if is_remote_url(url_or_filename):
         # URL, so get it from the cache (downloading if necessary)
         output_path = get_from_cache(

src/transformers/modeling_flax_utils.py

@@ -28,7 +28,7 @@
 from jax.random import PRNGKey
 
 from .configuration_utils import PretrainedConfig
-from .file_utils import FLAX_WEIGHTS_NAME, WEIGHTS_NAME, cached_path, hf_bucket_url, is_remote_url
+from .file_utils import FLAX_WEIGHTS_NAME, WEIGHTS_NAME, cached_path, hf_bucket_url, is_offline_mode, is_remote_url
 from .utils import logging
 
 
@@ -229,6 +229,10 @@ def from_pretrained(
         use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
 
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
         # Load config if we don't provide a configuration
         if not isinstance(config, PretrainedConfig):
             config_path = config if config is not None else pretrained_model_name_or_path

src/transformers/modeling_tf_utils.py

@@ -36,6 +36,7 @@
     ModelOutput,
     cached_path,
     hf_bucket_url,
+    is_offline_mode,
     is_remote_url,
 )
 from .generation_tf_utils import TFGenerationMixin
@@ -1151,6 +1152,10 @@ def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
         revision = kwargs.pop("revision", None)
         mirror = kwargs.pop("mirror", None)
 
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
         # Load config if we don't provide a configuration
         if not isinstance(config, PretrainedConfig):
             config_path = config if config is not None else pretrained_model_name_or_path

tests/test_offline.py

@@ -27,20 +27,37 @@ def test_offline_mode(self):
         # while running an external program
 
         # python one-liner segments
-        load = "from transformers import BertConfig, BertModel, BertTokenizer;"
-        run = "mname = 'lysandre/tiny-bert-random'; BertConfig.from_pretrained(mname) and BertModel.from_pretrained(mname) and BertTokenizer.from_pretrained(mname);"
-        mock = 'import socket; exec("def offline_socket(*args, **kwargs): raise socket.error(\\"Offline mode is enabled.\\")"); socket.socket = offline_socket;'
+
+        # this must be loaded before socket.socket is monkey-patched
+        load = """
+from transformers import BertConfig, BertModel, BertTokenizer
+        """
+
+        run = """
+mname = "lysandre/tiny-bert-random"
+BertConfig.from_pretrained(mname)
+BertModel.from_pretrained(mname)
+BertTokenizer.from_pretrained(mname)
+print("success")
+        """
+
+        mock = """
+import socket
+def offline_socket(*args, **kwargs): raise socket.error("Offline mode is enabled")
+socket.socket = offline_socket
+        """
 
         # baseline - just load from_pretrained with normal network
-        cmd = [sys.executable, "-c", f"{load} {run}"]
+        cmd = [sys.executable, "-c", "\n".join([load, run])]
 
         # should succeed
         env = self.get_env()
         result = subprocess.run(cmd, env=env, check=False, capture_output=True)
         self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertIn("success", result.stdout.decode())
 
         # next emulate no network
-        cmd = [sys.executable, "-c", f"{load} {mock} {run}"]
+        cmd = [sys.executable, "-c", "\n".join([load, mock, run])]
 
         # should normally fail as it will fail to lookup the model files w/o the network
         env["TRANSFORMERS_OFFLINE"] = "0"
@@ -51,3 +68,4 @@ def test_offline_mode(self):
         env["TRANSFORMERS_OFFLINE"] = "1"
         result = subprocess.run(cmd, env=env, check=False, capture_output=True)
         self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertIn("success", result.stdout.decode())