[BUG] Deepspeed / transformers mismatch

[pascal-pfeiffer]

🐛 Bug

When training with deepspeed there is an error that appears to be due to changes in transformers and a resulting mismatch between deepspeed and transformers:

  File "/workspace/.venv/lib/python3.10/site-packages/deepspeed/runtime/config_utils.py", line 57, in __init__
    super().__init__(**data)
  File "/workspace/.venv/lib/python3.10/site-packages/pydantic/main.py", line 250, in __init__
    validated_self = self.__pydantic_validator__.validate_python(data, self_instance=self)
pydantic_core._pydantic_core.ValidationError: 1 validation error for DeepSpeedBF16Config
loss_scale_window
  Extra inputs are not permitted [type=extra_forbidden, input_value=100, input_type=int]
    For further information visit https://errors.pydantic.dev/2.12/v/extra_forbidden

To Reproduce

• Train a model with deepspeed (default settings and bf16).

LLM Studio version

f1cb8b6