fix: change tokenizer_class from TokenizersBackend to Qwen2Tokenizer for vLLM compat 0be270c verified alphaXiv commited on 3 days ago