[Performance] Load HF tokenizers and processors from a local snapshot (#60825)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
Harry Mellor
2026-10-12 13:45:54 +00:00
committed by GitHub
co-authored by Claude Opus 5.5 Kevin H. Luu
parent 092c35dd15
commit 3facf822b7
5 changed files with 146 additions and 4 deletions
+15 -2
View File
@@ -22,7 +22,10 @@ from typing_extensions import TypeVar
from vllm.logger import init_logger
from vllm.transformers_utils import processors
from vllm.transformers_utils.repo_utils import get_hf_file_to_dict
from vllm.transformers_utils.repo_utils import (
get_hf_file_to_dict,
get_non_weight_snapshot_path,
)
from vllm.transformers_utils.utils import convert_model_repo_to_path
from vllm.utils.func_utils import get_allowed_kwarg_only_overrides
@@ -214,7 +217,12 @@ def get_processor(
if revision is None:
revision = "main"
try:
processor_name = convert_model_repo_to_path(processor_name)
processor_name = requested_name = convert_model_repo_to_path(processor_name)
# A snapshot pinned to `revision` would shadow remote code at `code_revision`
if kwargs.get("code_revision") in (None, revision):
processor_name = get_non_weight_snapshot_path(
processor_name, revision=revision, token=kwargs.get("token")
)
registered_cls_name = get_processor_cls_name_from_config(
processor_name, revision=revision
)
@@ -269,6 +277,11 @@ def get_processor(
f"found type: {type(processor)}"
)
# Report the Hub ID rather than the local snapshot it was loaded from
tokenizer = getattr(processor, "tokenizer", processor)
if hasattr(tokenizer, "name_or_path"):
tokenizer.name_or_path = requested_name
return processor
+74
View File
@@ -304,6 +304,80 @@ def get_model_path(model: str | Path, revision: str | None = None):
)
_WEIGHT_FILE_PATTERNS = [
"*.safetensors",
"*.bin",
"*.pt",
"*.pth",
"*.ckpt",
"*.gguf",
"*.h5",
"*.msgpack",
"*.onnx",
"*.onnx_data",
# Large non-weight assets that tokenizers and processors never read
"*.png",
"*.jpg",
"*.jpeg",
"*.gif",
"*.webp",
"*.mp3",
"*.mp4",
"*.wav",
"*.flac",
"*.parquet",
"*.arrow",
"*.ipynb",
"*.zip",
"*.tar*",
]
@cache
def _snapshot_non_weight_files(
repo_id: str,
revision: str | None,
cache_dir: str | None,
token: str | bool | None,
) -> str:
# Avoid circular import
from vllm.model_executor.model_loader.weight_utils import get_lock
with get_lock(repo_id, cache_dir):
return hf_api().snapshot_download(
repo_id=repo_id,
revision=revision,
cache_dir=cache_dir,
token=token,
local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
ignore_patterns=_WEIGHT_FILE_PATTERNS,
)
def get_non_weight_snapshot_path(
repo_id: str,
revision: str | None = None,
cache_dir: str | None = None,
token: str | bool | None = None,
) -> str:
"""Download all non-weight files of a Hub repo and return the local snapshot.
Loading tokenizers and processors from a local directory stops Transformers
from making a Hub request per file (including the `additional_chat_templates`
listing). Returns `repo_id` unchanged if it is local or cannot be fetched, so
Transformers handles it as before. Failures are not cached, so they are retried.
"""
if envs.VLLM_USE_MODELSCOPE or Path(repo_id).exists():
return repo_id
# Normalise so that `None` and "main" share a cache entry
revision = revision or huggingface_hub.constants.DEFAULT_REVISION
try:
return _snapshot_non_weight_files(repo_id, revision, cache_dir, token)
except (HfHubHTTPError, OSError, ValueError):
logger.debug("Failed to snapshot %s", repo_id, exc_info=True)
return repo_id
def _try_download_from_hf_hub(
model: str | Path, file_name: str, revision: str | None
) -> Path | None: