vllm==0.10.0 breaks skip_tokenizer_init=True #818
Description
After upgrade vllm to v0.10.0 in #766, skip_tokenizer_init=True will be broken.
Current we always set skip_tokenizer_init=False as a workaround, need to revert back to self.cfg["vllm_cfg"]["skip_tokenizer_init"] once we upgrade vllm to a version vllm-project/vllm#21846 is resolved.