RuntimeError: Error(s) in loading state_dict for RobertaModel: Unexpected key(s) in state_dict: "embeddings.position_ids".

[dzenilee]

How to reproduce the behaviour

import spacy 
nlp = spacy.load('en_core_web_trf')

Full traceback:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[7], line 1
----> 1 nlp = spacy.load('en_core_web_trf')

File /opt/conda/lib/python3.8/site-packages/spacy/__init__.py:51, in load(name, vocab, disable, enable, exclude, config)
     27 def load(
     28     name: Union[str, Path],
     29     *,
   (...)
     34     config: Union[Dict[str, Any], Config] = util.SimpleFrozenDict(),
     35 ) -> Language:
     36     """Load a spaCy model from an installed package or a local path.
     37 
     38     name (str): Package name or model path.
   (...)
     49     RETURNS (Language): The loaded nlp object.
     50     """
---> 51     return util.load_model(
     52         name,
     53         vocab=vocab,
     54         disable=disable,
     55         enable=enable,
     56         exclude=exclude,
     57         config=config,
     58     )

File /opt/conda/lib/python3.8/site-packages/spacy/util.py:465, in load_model(name, vocab, disable, enable, exclude, config)
    463     return get_lang_class(name.replace("blank:", ""))()
    464 if is_package(name):  # installed as package
--> 465     return load_model_from_package(name, **kwargs)  # type: ignore[arg-type]
    466 if Path(name).exists():  # path to model data directory
    467     return load_model_from_path(Path(name), **kwargs)  # type: ignore[arg-type]

File /opt/conda/lib/python3.8/site-packages/spacy/util.py:501, in load_model_from_package(name, vocab, disable, enable, exclude, config)
    484 """Load a model from an installed package.
    485 
    486 name (str): The package name.
   (...)
    498 RETURNS (Language): The loaded nlp object.
    499 """
    500 cls = importlib.import_module(name)
--> 501 return cls.load(vocab=vocab, disable=disable, enable=enable, exclude=exclude, config=config)

File /opt/conda/lib/python3.8/site-packages/en_core_web_trf/__init__.py:10, in load(**overrides)
      9 def load(**overrides):
---> 10     return load_model_from_init_py(__file__, **overrides)

File /opt/conda/lib/python3.8/site-packages/spacy/util.py:682, in load_model_from_init_py(init_file, vocab, disable, enable, exclude, config)
    680 if not model_path.exists():
    681     raise IOError(Errors.E052.format(path=data_path))
--> 682 return load_model_from_path(
    683     data_path,
    684     vocab=vocab,
    685     meta=meta,
    686     disable=disable,
    687     enable=enable,
    688     exclude=exclude,
    689     config=config,
    690 )

File /opt/conda/lib/python3.8/site-packages/spacy/util.py:547, in load_model_from_path(model_path, meta, vocab, disable, enable, exclude, config)
    538 config = load_config(config_path, overrides=overrides)
    539 nlp = load_model_from_config(
    540     config,
    541     vocab=vocab,
   (...)
    545     meta=meta,
    546 )
--> 547 return nlp.from_disk(model_path, exclude=exclude, overrides=overrides)

File /opt/conda/lib/python3.8/site-packages/spacy/language.py:2155, in Language.from_disk(self, path, exclude, overrides)
   2152 if not (path / "vocab").exists() and "vocab" not in exclude:  # type: ignore[operator]
   2153     # Convert to list here in case exclude is (default) tuple
   2154     exclude = list(exclude) + ["vocab"]
-> 2155 util.from_disk(path, deserializers, exclude)  # type: ignore[arg-type]
   2156 self._path = path  # type: ignore[assignment]
   2157 self._link_components()

File /opt/conda/lib/python3.8/site-packages/spacy/util.py:1392, in from_disk(path, readers, exclude)
   1389 for key, reader in readers.items():
   1390     # Split to support file names like meta.json
   1391     if key.split(".")[0] not in exclude:
-> 1392         reader(path / key)
   1393 return path

File /opt/conda/lib/python3.8/site-packages/spacy/language.py:2149, in Language.from_disk.<locals>.<lambda>(p, proc)
   2147     if not hasattr(proc, "from_disk"):
   2148         continue
-> 2149     deserializers[name] = lambda p, proc=proc: proc.from_disk(  # type: ignore[misc]
   2150         p, exclude=["vocab"]
   2151     )
   2152 if not (path / "vocab").exists() and "vocab" not in exclude:  # type: ignore[operator]
   2153     # Convert to list here in case exclude is (default) tuple
   2154     exclude = list(exclude) + ["vocab"]

File /opt/conda/lib/python3.8/site-packages/spacy_transformers/pipeline_component.py:416, in Transformer.from_disk(self, path, exclude)
    409         self.model.attrs["set_transformer"](self.model, hf_model)
    411 deserialize = {
    412     "vocab": self.vocab.from_disk,
    413     "cfg": lambda p: self.cfg.update(deserialize_config(p)),
    414     "model": load_model,
    415 }
--> 416 util.from_disk(path, deserialize, exclude)  # type: ignore
    417 return self

File /opt/conda/lib/python3.8/site-packages/spacy/util.py:1392, in from_disk(path, readers, exclude)
   1389 for key, reader in readers.items():
   1390     # Split to support file names like meta.json
   1391     if key.split(".")[0] not in exclude:
-> 1392         reader(path / key)
   1393 return path

File /opt/conda/lib/python3.8/site-packages/spacy_transformers/pipeline_component.py:390, in Transformer.from_disk.<locals>.load_model(p)
    388 try:
    389     with open(p, "rb") as mfile:
--> 390         self.model.from_bytes(mfile.read())
    391 except AttributeError:
    392     raise ValueError(Errors.E149) from None

File /opt/conda/lib/python3.8/site-packages/thinc/model.py:638, in Model.from_bytes(self, bytes_data)
    636 msg = srsly.msgpack_loads(bytes_data)
    637 msg = convert_recursive(is_xp_array, self.ops.asarray, msg)
--> 638 return self.from_dict(msg)

File /opt/conda/lib/python3.8/site-packages/thinc/model.py:676, in Model.from_dict(self, msg)
    674         node.set_param(param_name, value)
    675     for i, shim_bytes in enumerate(msg["shims"][i]):
--> 676         node.shims[i].from_bytes(shim_bytes)
    677 return self

File /opt/conda/lib/python3.8/site-packages/spacy_transformers/layers/hf_shim.py:120, in HFShim.from_bytes(self, bytes_data)
    118     filelike.seek(0)
    119     device = get_torch_default_device()
--> 120     self._model.load_state_dict(torch.load(filelike, map_location=device))
    121     self._model.to(device)
    122 else:

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:2041, in Module.load_state_dict(self, state_dict, strict)
   2036         error_msgs.insert(
   2037             0, 'Missing key(s) in state_dict: {}. '.format(
   2038                 ', '.join('"{}"'.format(k) for k in missing_keys)))
   2040 if len(error_msgs) > 0:
-> 2041     raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
   2042                        self.__class__.__name__, "\n\t".join(error_msgs)))
   2043 return _IncompatibleKeys(missing_keys, unexpected_keys)

RuntimeError: Error(s) in loading state_dict for RobertaModel:
	Unexpected key(s) in state_dict: "embeddings.position_ids". 

Also:

~$ conda list torch
# packages in environment at /opt/conda:
#
# Name                    Version                   Build  Channel
efficientnet-pytorch      0.7.1              pyhd8ed1ab_1    conda-forge
pytorch                   2.0.1           py3.8_cuda11.7_cudnn8.5.0_0    pytorch
pytorch-cuda              11.7                 h67b0de4_0    pytorch
pytorch-lightning         2.0.1.post0              pypi_0    pypi
pytorch-mutex             1.0                        cuda    pytorch
rotary-embedding-torch    0.2.1                    pypi_0    pypi
torchaudio                2.0.2                py38_cu117    pytorch
torchmetrics              0.11.4                   pypi_0    pypi
torchtriton               2.0.0                      py38    pytorch
torchvision               0.15.2               py38_cu117    pytorch
torchviz                  0.0.2                    pypi_0    pypi

Your Environment

• Operating System:
• Python Version Used:
• spaCy Version Used:
• Environment Information:

spaCy version    3.6.1
Location         /opt/conda/lib/python3.8/site-packages/spacy
Platform         Linux-5.13.0-1023-aws-x86_64-with-glibc2.17
Python version   3.8.17
Pipelines        en_core_web_trf (3.6.1)

[adrianeboyd]

Most likely you have a version of transformers that is too new for the current spacy-transformers release. Can you show the versions of spacy-transformers and transformers?

It can also be helpful to run pip check to see if it finds some conflicting package versions.

[adrianeboyd]

This is related to #12900, but spacy-transformers v1.3.0 was yanked from both PyPI and conda-forge, so it's hopefully unlikely that you have that installed.

[dzenilee]

It can also be helpful to run pip check to see if it finds some conflicting package versions.

^ Thanks for this. Downgrading to transformers==4.30.2 resolved it.

[seanaedmiston]

Is there any indication when spacy-transformers will be compatible with hf transformers>=4.31? 4.31 is the minimum for running Llama 2 models.

[adrianeboyd]

We are working on it (explosion/spacy-transformers#398, explosion/spacy-transformers#401) but haven't made a final decision about extending the transformers requirements beyond the current (and therefore tested) release.

[github-actions]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.