max_model_len: 32768
max_num_seqs: 50
gpu_memory_utilization: 0.8
kv_cache_ratio: 0.75
quantization: wint4
tensor_parallel_size: 1