max_model_len: 32768
max_num_seqs: 32
data_parallel_size: 4
tensor_parallel_size: 1
enable_expert_parallel: True
enable_prefix_caching: False
splitwise_role: decode
cache_transfer_protocol: "rdma"
rdma_comm_ports: "7671,7672,7673,7674"
pd_comm_port: "2335"
engine_worker_queue_port: "4582,4583,4584,4585"
graph_optimization_config:
  use_cudagraph: False