max_model_len: 32768
max_num_seqs: 256
kv_cache_ratio: 0.75
tensor_parallel_size: 4
gpu_memory_utilization: 0.9