[Unsloth Docker] Can't import 'FastSentenceTransformer'

[river folio]

I'm running Docker (WSL2; local; Nvidia 3080 GPU) with the unsloth/unsloth:stable image. When trying to run the embedding fine-tuning notebook (EmbeddingGemma_(300M).ipynb ), I get this error:

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
Cell In[1], line 1
----> 1 from unsloth import FastSentenceTransformer
      3 fourbit_models = [
      4     "unsloth/all-MiniLM-L6-v2",
      5     "unsloth/embeddinggemma-300m",
   (...)     11 
     12 ] # More models at https://huggingface.co/unsloth
     14 model = FastSentenceTransformer.from_pretrained(
     15     model_name = "unsloth/embeddinggemma-300m",
     16     max_seq_length = 1024,   # Choose any for long context!
     17     full_finetuning = False, # [NEW!] We have full finetuning now!
     18 )

ImportError: cannot import name 'FastSentenceTransformer' from 'unsloth' (/opt/conda/lib/python3.11/site-packages/unsloth/__init__.py)

Since the top-level module comes from unsloth, I'm puzzled by the error; surely the docker image has unsloth and its dependencies correctly installed by default? Any suggestions?

Things I'm certain of

• Nvidia GPU correctly detected and used in Docker

[rain lintel]

can you check unsloth version?

pip show unsloth

[river folio]

Name: unsloth
Version: 2026.1.2
Summary: 2-5X faster training, reinforcement learning & finetuning
Home-page: http://www.unsloth.ai
Author: Unsloth AI team
Author-email: [email protected]
License-Expression: Apache-2.0
Location: /opt/conda/lib/python3.11/site-packages
Requires: accelerate, bitsandbytes, datasets, diffusers, hf_transfer, huggingface_hub, numpy, packaging, peft, protobuf, psutil, sentencepiece, torch, torchvision, tqdm, transformers, triton, trl, tyro, unsloth_zoo, wheel, xformers
Required-by: 

[rain lintel]

oh maybe docker is not updated

latest is 2026.1.4

comes with embedding support

[river folio]

Just pulled the latest image, now seems to be ollama acting up:

unsloth  | 2026-01-23 21:35:45,457 INFO spawned: 'ollama' with pid 357
unsloth  | 2026-01-23 21:35:45,457 INFO success: sshd entered RUNNING state, process has stayed up for > than 1 seconds (startsecs)                                                       
unsloth  | 2026-01-23 21:35:45,462 INFO exited: ollama (exit status 1; not expected)
unsloth  | 2026-01-23 21:35:48,255 INFO spawned: 'ollama' with pid 360                       
unsloth  | 2026-01-23 21:35:48,259 INFO exited: ollama (exit status 1; not expected)
unsloth  | 2026-01-23 21:35:51,262 INFO spawned: 'ollama' with pid 361
unsloth  | 2026-01-23 21:35:51,266 INFO exited: ollama (exit status 1; not expected)
unsloth  | 2026-01-23 21:35:52,266 INFO gave up: ollama entered FATAL state, too many start retries too quickly

gonna see if it persists or not

Oh, this line has also been persistent, although the notebook server etc don't seem affected:

unsloth  | 2026-01-23 21:35:40,752 CRIT could not write pidfile /var/run/supervisord.pid

Might be good to know if it's just me and/or whether it's a missed bug in the image

Re-running the embedding notebook:

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
File /opt/conda/lib/python3.11/site-packages/unsloth_zoo/temporary_patches/utils.py:107
    106 try:
--> 107     from transformers.processing_utils import Unpack
    108     assert \
    109         type(Unpack) is type(t_Unpack), \
    110         "Unsloth: Unpack type changed! Please file a bug report asap!"

File /opt/conda/lib/python3.11/site-packages/transformers/processing_utils.py:75
     74 if is_torch_available():
---> 75     from .modeling_utils import PreTrainedAudioTokenizerBase
     78 logger = logging.get_logger(__name__)

[river folio]

yeah the ollama error persists

[craggy crater]

The docker container hasn't yet been updated to support embedding model training that's why

it will be today

usually it takes a couple of days after a pypi release is made because we gotta test and resolve dependency conflicts inside the container before we release it

[craggy crater]

latest in docker is 2026.1.3
you need to do docker pull unsloth/unsloth to update your local image
but still 2026.1.3 doesn't support embedding models yet. this will come in the next few days

[river folio]

Indeed, I already pulled the latest image as soon as etherl mentioned it - but as you said, it didn't resolve the error since the image fundamentally lacks support for the embedding models

[craggy crater]

yes no worries. todays' image will

[craggy crater]

@river folio pull the latest (i just updated the tag):

docker pull unsloth/unsloth:latest

should work for you and can do embeddings

also fixed ollama issue