max_model_len: 32768
max_num_seqs: 56
gpu_memory_utilization: 0.9
kv_cache_ratio: 0.8
tensor_parallel_size: 8
quantization: wint4
limit_mm_per_prompt: '{"image": 100, "video": 100}'
reasoning_parser: ernie-45-vl
graph_optimization_config:
  use_cudagraph: True
  graph_opt_level: 2