huggingface/transformers · error · ImportError

Unable to convert output to PyTorch tensors format, PyTorch

Error message

Unable to convert output to PyTorch tensors format, PyTorch is not installed.

What it means

BatchFeature.convert_to_tensors resolves the as_tensor function per TensorType. For TensorType.PYTORCH it first checks is_torch_available(); if transformers was installed without torch (CPU-only extras, minimal dep set, or broken torch install), conversion cannot proceed and ImportError is raised.

Source

Thrown at src/transformers/feature_extraction_utils.py:118

    def __getstate__(self):
        return {"data": self.data}

    def __setstate__(self, state):
        if "data" in state:
            self.data = state["data"]

    def _get_is_as_tensor_fns(self, tensor_type: str | TensorType | None = None):
        if tensor_type is None:
            return None, None

        # Convert to TensorType
        if not isinstance(tensor_type, TensorType):
            tensor_type = TensorType(tensor_type)

        if tensor_type == TensorType.PYTORCH:
            if not is_torch_available():
                raise ImportError("Unable to convert output to PyTorch tensors format, PyTorch is not installed.")
            import torch

            def as_tensor(value):
                if torch.is_tensor(value):
                    return value

                # stack list of tensors if tensor_type is PyTorch (# torch.tensor() does not support list of tensors)
                if isinstance(value, (list, tuple)) and len(value) > 0 and torch.is_tensor(value[0]):
                    return torch.stack(value)

                # convert list of numpy arrays to numpy array (stack) if tensor_type is Numpy
                if isinstance(value, (list, tuple)) and len(value) > 0:
                    if isinstance(value[0], np.ndarray):
                        value = np.array(value)
                    elif (
                        isinstance(value[0], (list, tuple))
                        and len(value[0]) > 0
                        and isinstance(value[0][0], np.ndarray)

View on GitHub (pinned to a597f97485)

Solutions

Install torch in the environment (pip install torch) or use the appropriate transformers extra
If torch is installed, verify 'import torch' works — a broken install can also trip is_torch_available()
If you don't need tensors, call with return_tensors=None to keep numpy/python objects

Example fix

# before (env without torch)
fe(audio, return_tensors="pt")

# after
pip install torch   # then:
fe(audio, return_tensors="pt")
# or without torch:
fe(audio)  # numpy output

Defensive patterns

Strategy: validation

Validate before calling

from transformers.utils import is_torch_available

if not is_torch_available():
    raise ImportError("install torch before requesting return_tensors='pt'")

Type guard

def can_return_pt() -> bool:
    from transformers.utils import is_torch_available
    return is_torch_available()

Try / catch

try:
    batch = fe(audio, return_tensors="pt")
except ImportError as e:
    if "PyTorch is not installed" in str(e):
        batch = fe(audio, return_tensors="np")
    else:
        raise

Prevention

Pin torch in requirements for any tensor-returning path
Check is_torch_available() at startup, not mid-pipeline
Have a numpy fallback path for slim environments

When it happens

Trigger: Calling a feature extractor (or BatchFeature(..., tensor_type='pt') / return_tensors='pt') in an environment where import torch fails or torch is not installed.

Common situations: Deploying in slim containers/docker images without the torch dependency; CI lint/type-check jobs with transformers but no torch; installing transformers via a meta-package that omits torch.

Related errors

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/5cc4fad3f8f8a38d. Report an issue: GitHub.