max_model_len: 32768
max_num_seqs: 256
gpu_memory_utilization: 0.9
kv_cache_ratio: 0.75
quantization: wint8
tensor_parallel_size: 1