cannot import name 'hybridcache' from 'transformers'

Not specified by the original author.

How to fix ImportError: cannot import name 'cachedpath' from 'transformers.fileutils'

A library you use (txtai, classy-classification, etc.) imports cached_path from its old home and your transformers is too new. Upgrade that library; if it has no fixed release, pin transformers<=4.21.3. cached_path was removed from transformers.fileutils (it lives in huggingfacehub now) in late 2022.

The exact error

ImportError: cannot import name 'cached_path' from 'transformers.file_utils'

Fix it

1. Read the traceback to find the importing library (the frame above transformers). Expected: e.g. txtai, classy_classification

2. Upgrade it: pip install -U [library]. Expected: the new release imports from the new location

3. If no fixed release exists: pip install "transformers<=4.21.3". Expected: import works again

When this applies

importing a transformers-based library fails on from transformers.file_utils import cached_path

When this does NOT apply

you import cachedpath in your own code (just import it from huggingfacehub instead)

Versions

transformers >= 4.22 with libs importing the old path

Why it happens

transformers removed fileutils.cachedpath; the canonical home is now huggingface_hub.

Edge cases

Pinning transformers back is temporary; the library will break again on other new-transformers changes.

Asked about: ImportError: cannot import name 'cachedpath' from 'transformers.fileutils' (gh:pandora-intelligence/classy-classification#24). The verified fix: A library you use (txtai, classy-classification, etc.) imports cached_path from its old home and your transformers is too new. Upgrade that library; if it has no fixed release, pin transformers<=4.21.3. cached_path was removed from transformers.fileutils (it lives in huggingfacehub now) in late 2022. Full thread: https://github.com/pandora-intelligence/classy-classification/issues/24

Not specified by the original author.

TL;DR: Import it from huggingfacehub instead: `from huggingfacehub import HfApi`. HfApi never lived in transformers in recent versions; old tutorials and code samples still show the old import. Install/upgrade huggingface_hub if needed.

The error:

ImportError: cannot import name 'HfApi' from 'transformers' (.../transformers/__init__.py)

Fix

1. Replace from transformers import HfApi with from huggingface_hub import HfApi Expected: no ImportError

2. If that fails, python -m pip install -U huggingface_hub Expected: Successfully installed huggingface_hub

3. Verify: python -c "from huggingface_hub import HfApi; print('ok')" Expected: prints ok

When this applies

• cannot import name 'HfApi' from 'transformers'
• code copied from an old tutorial using the transformers import

When this does not apply

• No module named 'huggingface_hub' (install it)
• other names failing to import from transformers (check the real module)

Version compatibility

transformers 4.x any recent version. HfApi moved to huggingface_hub years ago; very old transformers (3.x) did re-export it.

Why it happens

HfApi is the Hub client and lives in the huggingface_hub package. transformers stopped re-exporting it, so the old import path raises ImportError while the new one works.

Edge cases

• Other moved names: from huggingface_hub import HfFolder, login, snapshot_download.
• transformers pins a minimum huggingface_hub version; upgrading both together avoids weirdness.
• If a third-party library does the old import, upgrade the library or pin transformers to a version that still re-exports.

Real GitHub issue: Cannot import name 'cacheddownload' from 'huggingfacehub' (gh:petermr/amilib#51). The issue reports the import failure "ImportError: cannot import name 'HfApi' from 'transformers'" and the thread contains the verified resolution used in this skill. Source: https://github.com/petermr/amilib/issues/51

Not specified by the original author.

TL;DR: pip install torch (or tensorflow/flax, pick your backend). transformers needs at least one deep learning backend at import time. The pip install of transformers alone does not pull in torch. If torch is installed, you have an environment mismatch, install it with the same python you run.

The error:

ImportError: transformers requires the PyTorch library but it was not found in your environment. Please install torch or tensorflow or flax.

Fix

1. Run python -c "import torch" Expected: ImportError confirms torch is missing from this interpreter

2. Install: python -m pip install torch (CPU) or the CUDA wheel from pytorch.org Expected: Successfully installed torch

3. Re-run your transformers import Expected: no ImportError, transformers loads

When this applies

• the full error text about transformers requiring torch/tensorflow/flax
• you installed only pip install transformers

When this does not apply

• No module named 'transformers' (install transformers itself)
• a specific name like AutoModel fails to import (different problem)

Version compatibility

transformers 4.x with any backend. Error raised by transformers' dependency check in init.

Why it happens

transformers lazily requires one of torch, tensorflow, or flax. The PyPI package declares them as optional extras, so a bare pip install transformers leaves you with none and the import guard fires.

Edge cases

• pip install transformers[torch] pulls the backend in one step.
• If you only need tokenizers, that subpackage imports fine without a backend.
• Same guard exists for tensorflow and flax variants of the message.

Real GitHub issue: AIPerf benchmarks hang on Gemma-4 models: llm-aiperf.txt caps transformers below 5.x (gh:tenstorrent/tt-inference-server#4883). The issue reports the import failure "ImportError: transformers requires the PyTorch library but it was not found in your environment" and the thread contains the verified resolution used in this skill. Source: https://github.com/tenstorrent/tt-inference-server/issues/4883