Lora Training & TrainingPRO transformers training cuttoff error.

[Daniel23stack]

Training on the Undi95_MXLewd-L2-20B and other models with the transformer loader. On a new installation, I attempted transformers training with and without exllama and received the following error, causing the training to terminate immediately. Loaded into 4bit, auto devices. With and without use_double_quant. Lora training uses raw text files.

It should be noted that the identical data and settings were previously used to create a lora.

When trying to load without exllama I get the following error:

Traceback (most recent call last):
  File "G:\Character_AI\ModelLoader\text-generation-webui\modules\models.py", line 236, in huggingface_loader
    bits=config.quantization_config.get('bits', 4),
         ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "G:\Character_AI\ModelLoader\text-generation-webui\installer_files\env\Lib\site-packages\transformers\configuration_utils.py", line 263, in __getattribute__
    return super().__getattribute__(key)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'LlamaConfig' object has no attribute 'quantization_config'

The error in question:

Step:     32 {'loss': 8.2581, 'grad_norm': 0.10365231335163116, 'learning_rate': 0.028000000000000004, 'epoch': 0.0006699555393142091}
Step:     33 {'loss': 7.6102, 'grad_norm': 0.08936744183301926, 'learning_rate': 0.028999999999999998, 'epoch': 0.0006902572223237306}
Step:     34 Stop Loss 0 reached.
{'loss': 0.0, 'grad_norm': nan, 'learning_rate': 0.028999999999999998, 'epoch': 0.0007105589053332521}
Step:     34 {'train_runtime': 145.5746, 'train_samples_per_second': 1353.437, 'train_steps_per_second': 338.363, 'train_loss': 10.109378821509225, 'epoch': 0.0007105589053332521}
G:\Character_AI\ModelLoader\text-generation-webui\installer_files\env\Lib\site-packages\peft\utils\save_and_load.py:148: UserWarning: Could not find a config file in models\Undi95_MXLewd-L2-20B - will assume that the vocabulary was not modified.

[Daniel23stack]

You can only train with base models such as llama-2 or llama-3 in transformers.