[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
@@ -2,14 +2,19 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import importlib
from unittest.mock import patch
from transformers import processing_utils
from transformers.processing_utils import ProcessingKwargs
from typing_extensions import Unpack
from vllm.tokenizers import get_tokenizer
from vllm.transformers_utils.processor import (
get_processor,
get_processor_kwargs_keys,
get_processor_kwargs_type,
)
from vllm.transformers_utils.repo_utils import _snapshot_non_weight_files
class _FakeProcessorKwargs(ProcessingKwargs, total=False): # type: ignore
@@ -65,3 +70,26 @@ def test_get_processor_kwargs_from_processor_module_scan_returns_full_union():
proc = _ProcWithoutUnpack()
keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))
_assert_has_all_expected(keys)
def test_hub_processor_loads_from_local_snapshot():
"""Transformers makes a Hub request per file when given a repo ID."""
with patch.object(
processing_utils,
"list_repo_templates",
wraps=processing_utils.list_repo_templates,
) as list_repo_templates:
processor = get_processor("Qwen/Qwen2-VL-2B-Instruct")
assert processor.tokenizer.name_or_path == "Qwen/Qwen2-VL-2B-Instruct"
assert processor.chat_template is not None
list_repo_templates.assert_not_called()
def test_processor_reuses_tokenizer_snapshot():
"""The chat template lookup must not snapshot the repo a second time."""
model = "Qwen/Qwen2-VL-2B-Instruct"
_snapshot_non_weight_files.cache_clear()
get_tokenizer(model)
get_processor(model)
assert _snapshot_non_weight_files.cache_info().misses == 1