max_model_len: 32768
max_num_seqs: 96
gpu_memory_utilization: 0.85
kv_cache_ratio: 0.71
tensor_parallel_size: 4
quantization: wint4