Transformers: Segmentation fault when trying to load models

[michaelcapizzi]

We are using Azure ML pipelines to train our transformers models. We have had it working for a few weeks, and then recently (just noticed it a few days ago), when trying to initialize a model, we are getting Segmentation fault.

I tried just loading the models locally this morning and have the same issues. See snippet below.

config = config_class.from_pretrained(model_name, num_labels=10)
tokenizer = tokenizer_class.from_pretrained(model_name, do_lower_case=False)
model = model_class.from_pretrained("distilroberta-base", from_tf=False, config=config)

I also tried to download the *_model.bin and pass a local path instead of the model name and also got a Segmentation fault. I also tried to use bert-base-uncased instead of distilroberta-base and had the same issue.

I am running on Ubuntu, with the following package versions:

torch==1.3.0
tokenizers=0.0.11
transformers==2.4.1

UPDATE:

I hacked some example scripts and had success, so I think the issue is that our code uses...

    "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
    "mroberta": (RobertaConfig, RobertaForMultiLabelTokenClassification, RobertaTokenizer),    # our custom multilabel class

instead of what the example scripts use...

    AutoConfig,
    AutoModelForTokenClassification,
    AutoTokenizer,

Was there a breaking change to model files recently that would mean that our use of the "non-auto" classes are no longer usable?

UPDATE 2:

Our original code does not cause a Segmentation fault on Windows.

[michaelcapizzi]

Bumping to torch==1.5.1 fixes this issue. But it's still unclear why.

[jsw-zorro]

I have also met the same issue and upgrading to torch1.5.1 also solves my problem.

[LysandreJik]

Possibly related to https://github.com/huggingface/transformers/issues/4857

[chmod-700]

Downgrade to sentencepiece==0.1.91 solve it.
I am using PyTorch 1.2.0 + transformers3.0.0

[drevicko]

Downgrade to sentencepiece==0.1.91 solve it.
I am using PyTorch 1.2.0 + transformers3.0.0

Also PyTorch 1.4.0 + transformers 3.0.2

[LysandreJik]

Closing this as solved by #5418. Feel free to re-open if you still face an issue.

[pramitchoudhary]

For me either adding sentencepiece==0.1.91 + torch==1.3.1 + transformers==2.4.1
or torch==1.5.1 + transformers==2.4.1 worked.