* tokenize inputs directly in apply_chat_template * refactor processing * revert changes processing llava * Update docs * fix issue with str being iterable * add test chat text only * change function name
qwen2_5_omni
TrainingArguments.torch_empty_cache_steps