Suggested title

[Bug] ModelScope 1.37.1 fails to load BERT/T5 NLP models with transformers 5.x because transformers.onnx was removed

Checklist

• I have searched the tutorial on modelscope doc-site
• I have searched related issues but cannot get the expected help.
• The bug has not been fixed in the latest version.

Describe the bug

ModelScope 1.37.1 still imports ONNX config classes from transformers.onnx in several NLP configuration modules. In an environment where transformers==5.8.1 is installed, transformers.onnx no longer exists, so common NLP pipelines such as BERT fill-mask fail during model import.

This blocks normal non-plugin models, for example iic/nlp_bert_fill-mask_chinese-base, before inference starts:

modelscope/models/nlp/bert/configuration.py:21: in <module>
    from transformers.onnx import OnnxConfig
E   ModuleNotFoundError: No module named 'transformers.onnx'

E   ModuleNotFoundError: BertForMaskedLM: No module named 'transformers.onnx'
E   ModuleNotFoundError: FillMaskPipeline: BertForMaskedLM: No module named 'transformers.onnx'

The direct imports still exist in the 1.37.1 source tree:

modelscope/models/nlp/bert/configuration.py:21
modelscope/models/nlp/megatron_bert/configuration.py:22
modelscope/models/nlp/xlm_roberta/configuration.py:21
modelscope/models/nlp/T5/configuration.py:19

requirements/framework.txt currently leaves transformers unconstrained, so a fresh install can resolve to transformers 5.x and hit this failure.

Expected behavior: BERT/T5/XLM-RoBERTa/MegatronBERT model loading should either work with supported transformers versions, or ModelScope should declare an upper bound that prevents incompatible transformers 5.x installs.

To Reproduce

• What command or script did you run?

docker run --rm \
  -v /mnt/nas2/hzh/modelscope:/workspace \
  -w /workspace \
  modelscope-registry.cn-hangzhou.cr.aliyuncs.com/modelscope-repo/modelscope:ubuntu22.04-cuda12.8.1-py312-torch2.10.0-1.37.1 \
  bash -lc 'pip install pytest pytest-ordering pytest-assume jsonpath-ng filelock -q && cd /workspace/test_release1.37 && python -m pytest test_r37_trust_remote_code.py::TestR37TrustRemoteCode::test_safe_model_not_blocked_without_trust test_r37_trust_remote_code.py::TestR37TrustRemoteCode::test_normal_model_compatible_with_136 -v --tb=short'

A minimal script reproduces the same failure:

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

pipeline(
    task=Tasks.fill_mask,
    model='iic/nlp_bert_fill-mask_chinese-base',
)

• Did you make any modifications on the code or config? Did you understand what you have modified?

No source-code modification was made to ModelScope for this repro. The SDK was installed from the release/1.37 source tree at commit 121f4150fa9615882414bf424199e15be09c8ae5. The pytest file only exercises normal public model loading.

• What dataset did you use?

No dataset is required. The public model iic/nlp_bert_fill-mask_chinese-base was used.

Your Environments (required)

• OS: test report records Linux Ubuntu 22.04 host running the docker image below; exact uname -a was not captured by the automation log.
• CPU: exact lscpu output was not captured by the automation log. Host also has 2 x NVIDIA A100-SXM4-80GB GPUs, driver 570.133.20, CUDA 12.8.
• Commit id (e.g. a3ffc7d8): 121f4150fa9615882414bf424199e15be09c8ae5 (release/1.37, ModelScope 1.37.1)
• You may add addition that may be helpful for locating the problem, such as
  • Docker image: modelscope-registry.cn-hangzhou.cr.aliyuncs.com/modelscope-repo/modelscope:ubuntu22.04-cuda12.8.1-py312-torch2.10.0-1.37.1
  • Python: 3.12.13
  • pytest: 9.0.3
  • transformers: 5.8.1
  • modelscope: 1.37.1, installed from source
  • Related requirement: requirements/framework.txt contains unconstrained transformers

Please @ corresponding people according to your problem:

issue_01_transformers_onnx_removed.md

02c_trust_final.log