Cleanup some huggingface_hub-related stuff (#32788)

vllm/lora/utils.py

@@ -5,12 +5,7 @@ import os
 from typing import TYPE_CHECKING, Optional
 
 import huggingface_hub
-from huggingface_hub.utils import (
-    EntryNotFoundError,
-    HfHubHTTPError,
-    HFValidationError,
-    RepositoryNotFoundError,
-)
+from huggingface_hub.utils import HfHubHTTPError, HFValidationError
 from torch import nn
 from transformers import PretrainedConfig
 
@@ -243,12 +238,7 @@ def get_adapter_absolute_path(lora_path: str) -> str:
     # If the path does not exist locally, assume it's a Hugging Face repo.
     try:
         local_snapshot_path = huggingface_hub.snapshot_download(repo_id=lora_path)
-    except (
-        HfHubHTTPError,
-        RepositoryNotFoundError,
-        EntryNotFoundError,
-        HFValidationError,
-    ):
+    except (HfHubHTTPError, HFValidationError):
         # Handle errors that may occur during the download
         # Return original path instead of throwing error here
         logger.exception("Error downloading the HuggingFace model")

vllm/transformers_utils/config.py

@@ -33,7 +33,6 @@ from .gguf_utils import (
     split_remote_gguf,
 )
 from .repo_utils import (
-    _get_hf_token,
     file_or_path_exists,
     get_hf_file_to_dict,
     list_repo_files,
@@ -135,7 +134,6 @@ class HFConfigParser(ConfigParserBase):
             revision=revision,
             code_revision=code_revision,
             trust_remote_code=trust_remote_code,
-            token=_get_hf_token(),
             **kwargs,
         )
         # Use custom model class if it's in our registry
@@ -157,7 +155,6 @@ class HFConfigParser(ConfigParserBase):
                 revision=revision,
                 code_revision=code_revision,
                 trust_remote_code=trust_remote_code,
-                token=_get_hf_token(),
                 **kwargs,
             )
         else:
@@ -168,7 +165,6 @@ class HFConfigParser(ConfigParserBase):
                     trust_remote_code=trust_remote_code,
                     revision=revision,
                     code_revision=code_revision,
-                    token=_get_hf_token(),
                     **kwargs,
                 )
             except ValueError as e:
@@ -218,7 +214,6 @@ class MistralConfigParser(ConfigParserBase):
                 model,
                 revision=revision,
                 code_revision=code_revision,
-                token=_get_hf_token(),
                 **kwargs,
             )
         except OSError:  # Not found
@@ -529,7 +524,6 @@ def maybe_override_with_speculators(
         model if gguf_model_repo is None else gguf_model_repo,
         revision=revision,
         trust_remote_code=trust_remote_code,
-        token=_get_hf_token(),
         **kwargs,
     )
     speculators_config = config_dict.get("speculators_config")
@@ -871,9 +865,7 @@ def get_sentence_transformer_tokenizer_config(
     if not encoder_dict and not Path(model).is_absolute():
         try:
             # If model is on HuggingfaceHub, get the repo files
-            repo_files = list_repo_files(
-                model, revision=revision, token=_get_hf_token()
-            )
+            repo_files = list_repo_files(model, revision=revision)
         except Exception:
             repo_files = []
 
@@ -1042,10 +1034,7 @@ def try_get_safetensors_metadata(
     revision: str | None = None,
 ):
     get_safetensors_metadata_partial = partial(
-        get_safetensors_metadata,
-        model,
-        revision=revision,
-        token=_get_hf_token(),
+        get_safetensors_metadata, model, revision=revision
     )
 
     try:

vllm/transformers_utils/repo_utils.py

@@ -12,10 +12,7 @@ from pathlib import Path
 from typing import TypeVar
 
 import huggingface_hub
-from huggingface_hub import (
-    hf_hub_download,
-    try_to_load_from_cache,
-)
+from huggingface_hub import hf_hub_download, try_to_load_from_cache
 from huggingface_hub import list_repo_files as hf_list_repo_files
 from huggingface_hub.utils import (
     EntryNotFoundError,
@@ -31,21 +28,6 @@ from vllm.logger import init_logger
 logger = init_logger(__name__)
 
 
-def _get_hf_token() -> str | None:
-    """
-    Get the HuggingFace token from environment variable.
-
-    Returns None if the token is not set, is an empty string,
-    or contains only whitespace.
-    This follows the same pattern as huggingface_hub library which
-    treats empty string tokens as None to avoid authentication errors.
-    """
-    token = os.getenv("HF_TOKEN")
-    if token and token.strip():
-        return token
-    return None
-
-
 _R = TypeVar("_R")
 
 
@@ -153,6 +135,8 @@ def file_exists(
     revision: str | None = None,
     token: str | bool | None = None,
 ) -> bool:
+    # `list_repo_files` is cached and retried on error, so this is more efficient than
+    # huggingface_hub.file_exists default implementation when looking for multiple files
     file_list = list_repo_files(
         repo_id, repo_type=repo_type, revision=revision, token=token
     )
@@ -178,9 +162,7 @@ def file_or_path_exists(
     # hf_hub. This will fail in offline mode.
 
     # Call HF to check if the file exists
-    return file_exists(
-        str(model), config_name, revision=revision, token=_get_hf_token()
-    )
+    return file_exists(str(model), config_name, revision=revision)
 
 
 def get_model_path(model: str | Path, revision: str | None = None):
@@ -209,9 +191,7 @@ def get_hf_file_bytes(
     file_path = try_get_local_file(model=model, file_name=file_name, revision=revision)
 
     if file_path is None:
-        hf_hub_file = hf_hub_download(
-            model, file_name, revision=revision, token=_get_hf_token()
-        )
+        hf_hub_file = hf_hub_download(model, file_name, revision=revision)
         file_path = Path(hf_hub_file)
 
     if file_path is not None and file_path.is_file():