[Bug] Reference to depreciated torch_dtype.

[JonathanRaines]

1. Did you update? pip install --upgrade unsloth unsloth_zoo: Yes
2. Colab or Kaggle or local / cloud: local
3. Number GPUs used, use nvidia-smi: n/a
4. Which notebook? Please link!: n/a
5. Which Unsloth version, TRL version, transformers version, PyTorch version? Unsloth: 2025.7.2, TRL: 0.19.1, tranformers 4.56.1, pytorch 2.8.0
6. Which trainer? SFTTrainer, GRPOTrainer etc: n/a

model, tokenizer = unsloth.FastModel.from_pretrained(
        model_name="unsloth/gemma-3-1b-it-unsloth-bnb-4bit",
        max_seq_length=512,
        load_in_4bit = True,
    )

I get:
File "[...]/.venv/lib/python3.13/site-packages/unsloth/models/loader.py", line 830, in from_pretrained
"bnb_4bit_compute_dtype" : model.config.to_dict()["torch_dtype"],
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
KeyError: 'torch_dtype'
FastModel appears to have been updated to use "dtype" instead of "torch_dtype" but this line still tries to access "torch_dtype" as a key.
Manually changing it to model.config.to_dict()["dtype"], fixes the issue

[Erland366]

I am pretty sure this is already been fixed. I tested in colab, my local using pypi, and my local using main. All are working.

By the way, right now we are pinning on this version which throughly tested already :

transformers                      4.55.4
trl                               0.22.2
unsloth                           2025.9.7
unsloth-zoo                       2025.9.9

[rolandtannous]

He might need to install from the main-repo.
Make sure to update your installation as follows:

pip install --force-reinstall --no-deps git+https://github.com/unslothai/unsloth-zoo git+https://github.com/unslothai/unsloth

[JonathanRaines]

Many thanks