tensor_parallel_size: 8
max_model_len: 32768
max_num_seqs: 32
num_gpu_blocks_override: 4096
kv_cache_ratio: 0.5
reasoning_parser: ernie-x1