NameError: name 'nn' is not defined when importing sentence-transformers with latest transformers

[Alan-Jowett]

System Info

transformers version: latest (installed via pip today, 2026-02-05)
sentence-transformers version: latest
torch version: 2.x (from pytorch/pytorch Docker image)
Python version: 3.11+
OS: Linux (Docker container)

Who can help?

@ArthurZucker @Rocketknight1

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder
• My own task or dataset

Reproduction

Minimal reproduction:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")

Or directly:

from transformers import AutoModel

Results in:

NameError: name 'nn' is not defined

The error occurs during the import chain when transformers attempts to load a module that references nn.Module or similar torch.nn constructs without first importing torch.nn as nn.

Expected behavior

The import should succeed without errors. This worked with previous versions of transformers.

Additional context

This appears to be a similar regression to #38269 ("NameError: name 'Replicate' is not defined"), which was caused by a missing import in transformers/integrations/tensor_parallel.py.

The current error (nn not defined) suggests another module has a similar missing import issue, likely referencing nn.Module without the corresponding import torch.nn as nn statement.

Workaround: Pinning transformers<4.53.0 (or to the last known working version) resolves the issue.

[Alan-Jowett]

Update: This appears to be related to the transformers v5.0.0 release (January 26, 2026).

The last working CI build was on February 4, 2026, which likely used a cached Docker layer with transformers 4.x. Fresh builds now pull v5.0.0, which causes this import error.

The error occurs when sentence-transformers (or other downstream packages) imports from transformers, and a module in v5.0.0 references nn.Module without the proper import torch.nn as nn statement.

Temporary workaround: Pin transformers<5.0.0 until this is resolved.

[ArthurZucker]

cc @tomaarsen as well!

[tomaarsen]

Hello @Alan-Jowett!

Thanks for raising this. I'm sadly unable to reproduce this right now. I created a fresh virtualenvironment with (only) the latest sentence-transformers and transformers, and the code you linked seems to work for me:

>>> import transformers
>>> from transformers import AutoModel
>>> from sentence_transformers import SentenceTransformer
>>>
>>> model = SentenceTransformer("all-MiniLM-L6-v2")
Loading weights: 100%|████████████████████| 103/103 [00:00<00:00, 3319.53it/s, Materializing param=pooler.dense.weight]
BertModel LOAD REPORT from: sentence-transformers/all-MiniLM-L6-v2
Key                     | Status     |  |
------------------------+------------+--+-
embeddings.position_ids | UNEXPECTED |  |

Notes:
- UNEXPECTED    :can be ignored when loading from different task/architecture; not ok if you expect identical arch.

Could you perhaps find which line in transformers uses an un-imported nn? Perhaps the import is locked behind an if is_..._available(): from torch import nn, causing your issue.

• Tom Aarsen

[Alan-Jowett]

Thank you for taking a look into this issue. I will try stripping down my container to produce a minimal reproduction of the issue.

Alan Jowett

[Alan-Jowett]

Seems like it's triggered by an old version of PyTorch?

# Minimal reproduction for transformers v5.0.0 NameError: name 'nn' is not defined
# Issue: https://github.com/huggingface/transformers/issues/43784
#
# The bug occurs when:
# 1. PyTorch version is < 2.4 (e.g., 2.2.1 from pytorch/pytorch base image)
# 2. transformers >= 5.0.0 is installed
# 3. transformers detects old PyTorch and "disables" it
# 4. transformers/integrations/accelerate.py line 62 uses `nn.Module` without importing it
#
# Build: docker build -t transformers-nn-bug .
# Run:   docker run --rm transformers-nn-bug
#
# Expected output:
#   Disabling PyTorch because PyTorch >= 2.4 is required but found 2.2.1
#   ...
#   File ".../transformers/integrations/accelerate.py", line 62, in <module>
#     ) -> tuple[int, list[str], list[nn.Module]]:
#   NameError: name 'nn' is not defined

FROM pytorch/pytorch@sha256:11691e035a3651d25a87116b4f6adc113a27a29d8f5a6a583f8569e0ee5ff897

# This base image has PyTorch 2.2.1, which is < 2.4 required by transformers 5.0.0

# Install latest transformers (v5.0.0+) - the bug is in transformers, not sentence-transformers
RUN pip install --no-cache-dir "transformers>=5.0.0" sentence-transformers

# Minimal test - just importing sentence_transformers triggers the bug
# because it imports from transformers.integrations.peft which chains to accelerate.py
CMD ["python", "-c", "from sentence_transformers import SentenceTransformer; print('Success!')"]

Disabling PyTorch because PyTorch >= 2.4 is required but found 2.2.1
PyTorch was not found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
Python: 3.10.13 (main, Sep 11 2023, 13:44:35) [GCC 11.2.0]
PyTorch: 2.2.1
Transformers: 5.1.0
Traceback (most recent call last):
  File "/test.py", line 7, in <module>
    from sentence_transformers import SentenceTransformer
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/__init__.py", line 15, in <module>
    from sentence_transformers.cross_encoder import (
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/cross_encoder/__init__.py", line 3, in <module>
    from .CrossEncoder import CrossEncoder
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/cross_encoder/CrossEncoder.py", line 35, in <module>
    from sentence_transformers.cross_encoder.fit_mixin import FitMixin
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/cross_encoder/fit_mixin.py", line 20, in <module>
    from sentence_transformers.datasets.NoDuplicatesDataLoader import NoDuplicatesDataLoader
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/datasets/__init__.py", line 13, in <module>
    from .ParallelSentencesDataset import ParallelSentencesDataset
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/datasets/ParallelSentencesDataset.py", line 19, in <module>
    from sentence_transformers import SentenceTransformer
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/SentenceTransformer.py", line 45, in <module>
    from .peft_mixin import PeftAdapterMixin
  File "/opt/conda/lib/python3.10/site-packages/sentence_transformers/peft_mixin.py", line 5, in <module>
    from transformers.integrations.peft import PeftAdapterMixin as PeftAdapterMixinTransformers
  File "/opt/conda/lib/python3.10/site-packages/transformers/integrations/peft.py", line 22, in <module>
    from ..conversion_mapping import (
  File "/opt/conda/lib/python3.10/site-packages/transformers/conversion_mapping.py", line 20, in <module>
    from .core_model_loading import (
  File "/opt/conda/lib/python3.10/site-packages/transformers/core_model_loading.py", line 34, in <module>
    from .integrations.accelerate import get_device, offload_weight
  File "/opt/conda/lib/python3.10/site-packages/transformers/integrations/accelerate.py", line 62, in <module>
    ) -> tuple[int, list[str], list[nn.Module]]:
NameError: name 'nn' is not defined