ai-agents-model-saving-shared--when-fine-tuning-a-model-with-shared-tensor-weight-0aa34389

THEN Disable safe_serialization by setting save_safetensors=False in TrainingArguments, or manually save using model.save_pretrained with safe_serialization=False. Alternatively, ensure that shared tensors are handled by removing sharing before saving, or use the save_model method instead. The root cause is that Trainer's default save logic for safetensors does not correctly detect and handle shared tensors in models like Gemma 2.

from transformers import TrainingArguments

# Workaround: disable safe serialization
training_args = TrainingArguments(
    output_dir="./results",
    save_safetensors=False,
    ...
)