GroundingDINO crashes with _path_isfile NoneType on Python 3.12 / transformers ≥4.x (BERT tokenizer incompatibility)

[zhouchiyan82]

环境信息（Environment）
OS: Windows 10 x64
Python: 3.12.10
ComfyUI: 0.7.0
PyTorch: 2.8.0
CUDA: 12.9
transformers: 4.x (latest)
Node: ComfyUI_LayerStyle_Advance
Feature: SegmentAnythingUltra / GroundingDINO

问题描述（Problem）
When running SegmentAnythingUltra / GroundingDINO, ComfyUI crashes with:

TypeError: _path_isfile: path should be string, bytes, os.PathLike or integer, not NoneType

Stack trace points to:

os.path.isfile(vocab_file)

inside transformers/models/bert/tokenization_bert.py.

关键现象（Important Findings）

Internet access is working

HuggingFace Hub is accessible

The following works correctly in the same Python environment:

from transformers import AutoTokenizer
AutoTokenizer.from_pretrained("bert-base-uncased")

However, GroundingDINO initialization fails inside the node

🧠 根本原因分析（Root Cause）

The current local_groundingdino/util/get_tokenlizer.py assumes:

AutoTokenizer.from_pretrained("bert-base-uncased")

will always return a tokenizer with a valid vocab_file path

This assumption breaks with modern transformers versions:

AutoTokenizer often returns BertTokenizerFast

Fast tokenizers do not expose vocab_file

vocab_file becomes None

GroundingDINO later calls os.path.isfile(vocab_file) → crash

This issue is not related to network, cache, or missing models, but to:

incompatible assumptions between old GroundingDINO code and new transformers behavior

✅ 解决方案（Working Fix）

Force GroundingDINO to use local HF snapshot paths and explicitly initialize:

BertTokenizer with a real vocab.txt

BertModel from the same local snapshot

Patched get_tokenlizer.py

import os
import huggingface_hub
from transformers import BertTokenizer, BertModel


def _get_local_bert_path(text_encoder_type: str):
    repo_path = huggingface_hub.snapshot_download(
        repo_id=text_encoder_type,
        allow_patterns=[
            "vocab.txt",
            "tokenizer_config.json",
            "config.json",
            "pytorch_model.bin",
        ],
    )
    return repo_path


def get_tokenlizer(text_encoder_type):
    repo_path = _get_local_bert_path(text_encoder_type)

    vocab_file = os.path.join(repo_path, "vocab.txt")
    if not os.path.isfile(vocab_file):
        raise RuntimeError(f"vocab.txt not found: {vocab_file}")

    tokenizer = BertTokenizer(
        vocab_file=vocab_file,
        do_lower_case=True,
    )
    return tokenizer


def get_pretrained_language_model(text_encoder_type):
    repo_path = _get_local_bert_path(text_encoder_type)

    model = BertModel.from_pretrained(
        repo_path,
        local_files_only=True,
    )
    return model

结果（Result）

GroundingDINO loads correctly

SegmentAnythingUltra works as expected

No _path_isfile NoneType error

Compatible with:

Python 3.12

transformers ≥ 4.x

Fast / slow tokenizer differences

This is not a user environment issue, but a compatibility issue caused by newer transformers behavior.
Explicitly loading BERT tokenizer/model from local HF snapshots resolves the problem reliably.