FIX Honor HF_HUB_OFFLINE mode if set by user

src/peft/auto.py

@@ -18,8 +18,6 @@
 import os
 from typing import Optional
 
-from huggingface_hub import file_exists
-from huggingface_hub.utils import HfHubHTTPError, HFValidationError
 from transformers import (
     AutoModel,
     AutoModelForCausalLM,
@@ -42,6 +40,7 @@
     PeftModelForTokenClassification,
 )
 from .utils.constants import TOKENIZER_CONFIG_NAME
+from .utils.other import check_file_exists_on_hf_hub
 
 
 class _BaseAutoPeftModel:
@@ -112,16 +111,13 @@ def from_pretrained(
             if token is None:
                 token = kwargs.get("use_auth_token", None)
 
-            try:
-                tokenizer_exists = file_exists(
-                    repo_id=pretrained_model_name_or_path,
-                    filename=TOKENIZER_CONFIG_NAME,
-                    revision=kwargs.get("revision", None),
-                    repo_type=kwargs.get("repo_type", None),
-                    token=token,
-                )
-            except (HfHubHTTPError, HFValidationError):  # not on the Hub, so probably local repo
-                pass
+            tokenizer_exists = check_file_exists_on_hf_hub(
+                repo_id=pretrained_model_name_or_path,
+                filename=TOKENIZER_CONFIG_NAME,
+                revision=kwargs.get("revision", None),
+                repo_type=kwargs.get("repo_type", None),
+                token=token,
+            )
 
         if tokenizer_exists:
             tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path)

src/peft/utils/other.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 import copy
 import inspect
+import os
 import warnings
 from contextlib import nullcontext
 from typing import Optional, Tuple
@@ -21,6 +22,8 @@
 import torch
 from accelerate.hooks import add_hook_to_module, remove_hook_from_module
 from accelerate.utils import is_npu_available, is_xpu_available
+from huggingface_hub import file_exists
+from huggingface_hub.utils import EntryNotFoundError, HFValidationError
 from safetensors.torch import storage_ptr, storage_size
 
 from ..import_utils import is_auto_gptq_available, is_torch_tpu_available
@@ -537,3 +540,39 @@ def cast_mixed_precision_params(model, dtype):
             p.data = p.to(dtype)
         else:
             p.data = p.to(torch.float32)
+
+
+def str_to_bool(value: str) -> int:
+    """
+    Converts a string representation of truth to `True` (1) or `False` (0).
+
+    True values are `y`, `yes`, `t`, `true`, `on`, and `1`; False value are `n`, `no`, `f`, `false`, `off`, and `0`;
+    """
+    # same as function as in accelerate.utils, which replaces the deprecated distutils.util.strtobool
+    value = value.lower()
+    if value in ("y", "yes", "t", "true", "on", "1"):
+        return 1
+    elif value in ("n", "no", "f", "false", "off", "0"):
+        return 0
+    else:
+        raise ValueError(f"invalid truth value {value}")
+
+
+def check_file_exists_on_hf_hub(repo_id: str, filename: str, **kwargs) -> Optional[bool]:
+    """Check if a file exists on HF Hub, if check was not successful returns None instead of erroring.
+
+    Respect offline mode if set.
+
+    """
+    exists: Optional[bool] = None
+    if str_to_bool(os.environ.get("HF_HUB_OFFLINE", "0")):
+        # user set offline mode, cannot check
+        return exists
+
+    try:
+        exists = file_exists(repo_id, filename, **kwargs)
+    except (HFValidationError, EntryNotFoundError):
+        # error, exists stays None
+        pass
+
+    return exists

src/peft/utils/save_and_load.py

@@ -17,10 +17,16 @@
 
 import torch
 from huggingface_hub import file_exists, hf_hub_download
-from huggingface_hub.utils import EntryNotFoundError, HFValidationError
+from huggingface_hub.utils import EntryNotFoundError
 from safetensors.torch import load_file as safe_load_file
 
-from .other import EMBEDDING_LAYER_NAMES, SAFETENSORS_WEIGHTS_NAME, WEIGHTS_NAME, infer_device
+from .other import (
+    EMBEDDING_LAYER_NAMES,
+    SAFETENSORS_WEIGHTS_NAME,
+    WEIGHTS_NAME,
+    check_file_exists_on_hf_hub,
+    infer_device,
+)
 from .peft_types import PeftType
 
 
@@ -140,14 +146,17 @@ def get_peft_model_state_dict(
         # we need to make sure we can download that config.
         has_remote_config = False
 
+        # ensure that this check is not performed in HF offline mode, see #1452
         if model_id is not None:
-            try:
-                has_remote_config = file_exists(model_id, "config.json")
-            except (HFValidationError, EntryNotFoundError):
+            exists = check_file_exists_on_hf_hub(model_id, "config.json")
+            if exists is None:
+                # check failed, could not determine if it exists or not
                 warnings.warn(
                     f"Could not find a config file in {model_id} - will assume that the vocabulary was not modified."
                 )
                 has_remote_config = False
+            else:
+                has_remote_config = exists
 
         # check if the vocab size of the base model is different from the vocab size of the finetuned model
         if (