infrastructure-model-saving-runtimeerror-some-tensors-share-memory-when-saving-73d39cba

THEN When saving a fine-tuned model that has shared tensors (common in models like Gemma 2 where embedding and lm_head weigh the same tensor), either use the `save_model` method instead of the default safetensors serialization, or set `save_safetensors=False` in TrainingArguments to fall back to PyTorch's native saving format. Ensure the workaround does not affect inference or downstream use.

training_args = TrainingArguments(..., save_safetensors=False, ...)  # Disables safetensors to avoid shared tensor error