RuntimeError: from_dlpack received an invalid capsule.

[john-a-m]

How to reproduce the behaviour

I'm trying to run python -m spacy train config.cfg --output ./output --paths.train ./train.spacy --paths.dev ./dev.spacy --gpu-id 0 and as soon as I see

ℹ Saving to output directory: output
ℹ Using GPU: 0

=========================== Initializing pipeline ===========================
Some weights of RobertaModel were not initialized from the model checkpoint at roberta-base and are newly initialized: ['roberta.pooler.dense.bias', 'roberta.pooler.dense.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
✔ Initialized pipeline

============================= Training pipeline =============================
ℹ Pipeline: ['transformer', 'ner']
ℹ Initial learn rate: 0.0

I get the stacktrace

C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\thinc\util.py:395: VisibleDeprecationWarning: This function is deprecated and will be removed in a future release. Use the cupy.from_dlpack() array constructor instead.  dlpack_tensor = xp_tensor.toDlpack()  # type: ignore⚠ Aborting and saving the final best model. Encountered exception:
RuntimeError('from_dlpack received an invalid capsule. Note that DLTensor
capsules can be consumed only once, so you might have already constructed a
tensor from it once.')
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\__main__.py", line 4, in <module>
    setup_cli()
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\cli\_util.py", line 87, in setup_cli
    command(prog_name=COMMAND)
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\click\core.py", line 1485, in __call__
    return self.main(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\typer\core.py", line 795, in main
    return _main(
           ^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\typer\core.py", line 188, in _main
    rv = self.invoke(ctx)
         ^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\click\core.py", line 1873, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\click\core.py", line 1269, in invoke
    return ctx.invoke(self.callback, **ctx.params)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\click\core.py", line 824, in invoke
    return callback(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\typer\main.py", line 1514, in wrapper
    return callback(**use_params)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\cli\train.py", line 54, in train_cli
    train(config_path, output_path, use_gpu=use_gpu, overrides=overrides)
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\cli\train.py", line 84, in train
    train_nlp(nlp, output_path, use_gpu=use_gpu, stdout=sys.stdout, stderr=sys.stderr)
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\training\loop.py", line 135, in train
    raise e
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\training\loop.py", line 118, in train
    for batch, info, is_best_checkpoint in training_step_iterator:
                                           ^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\training\loop.py", line 220, in train_while_improving
    nlp.update(
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy\language.py", line 1197, in update    proc.update(examples, sgd=None, losses=losses, **component_cfg[name])  # type: ignore
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "spacy/pipeline/transition_parser.pyx", line 439, in spacy.pipeline.transition_parser.Parser.update
    backprop_tok2vec(golds)
  File "spacy/ml/parser_model.pyx", line 334, in spacy.ml.parser_model.ParserStepModel.finish_steps
    self.bp_tokvecs(d_tokvecs[:-1])
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\thinc\layers\chain.py", line 60, in backprop
    dX = callback(dY)
         ^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\thinc\layers\chain.py", line 60, in backprop
    dX = callback(dY)
         ^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy_transformers\layers\listener.py", line 41, in backprop_and_clear
    result = self._backprop(*args, **kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy_transformers\pipeline_component.py", line 316, in backprop
    d_trf_full = trf_full.unsplit_by_doc(d_tensors)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\spacy_transformers\data_classes.py", line 315, in unsplit_by_doc
    model_output[f"output_{i}"] = xp2torch(xp.vstack(x))
                                  ^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\thinc\util.py", line 396, in xp2torch
    torch_tensor = torch.utils.dlpack.from_dlpack(dlpack_tensor)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\john\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\utils\dlpack.py", line 121, in from_dlpack
    return _from_dlpack(dlpack)
           ^^^^^^^^^^^^^^^^^^^^
RuntimeError: from_dlpack received an invalid capsule. Note that DLTensor capsules can be consumed only once, so you might have already constructed a tensor from it once.

I've found a few other issues that suggest this is a version mismatch between pytorch and numpy, however I can't seem to find the winning version combination. Any ideas on how I can resolve this issue?

#12221
cupy/cupy#8070

Your Environment

• Operating System: Windows 10
• Python Version Used: 3.12.10
• spaCy Version Used: 3.8.11
• Environment Information: here is the output of pip list

Package            Version
------------------ ------------
annotated-doc      0.0.4
annotated-types    0.7.0
blis               1.3.3
catalogue          2.0.10
certifi            2026.2.25
charset-normalizer 3.4.6
click              8.3.1
cloudpathlib       0.23.0
colorama           0.4.6
confection         0.1.5
cuda-pathfinder    1.4.3
cuda-toolkit       13.2.0
cupy-cuda12x       14.0.1
cymem              2.0.13
fastrlock          0.8.3
filelock           3.25.2
fsspec             2026.2.0
huggingface_hub    0.36.2
idna               3.11
Jinja2             3.1.6
markdown-it-py     4.0.0
MarkupSafe         3.0.3
mdurl              0.1.2
mpmath             1.3.0
murmurhash         1.0.15
networkx           3.6.1
numpy              2.4.3
packaging          26.0
pillow             12.0.0
pip                26.0.1
preshed            3.0.12
pydantic           2.12.5
pydantic_core      2.41.5
Pygments           2.19.2
PyYAML             6.0.3
regex              2026.2.28
requests           2.32.5
rich               14.3.3
safetensors        0.7.0
setuptools         82.0.1
shellingham        1.5.4
smart_open         7.5.1
spacy              3.8.11
spacy-alignments   0.9.2
spacy-legacy       3.0.12
spacy-loggers      1.0.5
spacy-transformers 1.3.9
srsly              2.5.2
sympy              1.13.1
thinc              8.3.10
tokenizers         0.21.4
torch              2.5.1+cu121
torchaudio         2.5.1+cu121
torchvision        0.20.1+cu121
tqdm               4.67.3
transformers       4.49.0
typer              0.24.1
typer-slim         0.24.0
typing_extensions  4.15.0
typing-inspection  0.4.2
urllib3            2.6.3
wasabi             1.1.3
weasel             0.4.3
wrapt              2.1.2

[rehan243]

fwiw, we ran into this with thinc (pretty sure you’re on spaCy v3.x?) and it’s almost always the DLPack capsule getting double-consumed. The stacktrace about from_dlpack received an invalid capsule is super misleading — turns out, cupy and thinc don’t play nice with each other if any intermediate tensor op grabs the capsule twice. The VisibleDeprecationWarning you got is legit: thinc’s old util stuff is still calling toDlpack() and not cupy’s recommended from_dlpack() constructor.

We got past this by making sure no ops chained on the tensor before spaCy tried to use it for training. Also, if you’re using Python 3.12, that’s actually bitten us a couple times since cupy (v12+) has some quirks with DLPack — rolling back to 3.10 or 3.11 fixed it for us without needing to hack thinc. Honestly, if you can’t downgrade, you’d need to patch thinc to avoid consuming the capsule more than once, but last I checked, upstream’s dragging their feet on merging that fix.

So yeah, it’s a combo of thinc’s legacy code, cupy changes, and Python version. Not gonna lie, took us a few weeks to find this, since the error’s just cryptic as hell.