max_model_len: 32768
max_num_seqs: 128
kv_cache_ratio: 0.75
tensor_parallel_size: 1
quantization: wint8