近期尝试使用flagos，发现推理效率相较海光自有的vllm0.11.0提升3倍多。命令如下： ```language docker run \ --name flagos-minimax \ --network=host \ --ipc=host \ --device=/dev/kfd \ --device=/dev/mkfd \ --device=/dev/dri \ -v /opt/hyhal:/opt/hyhal \ -v /{path_to}/workspace:/workspace \ -v /{path_to}/models:/models \ --group-add video \ --cap-add=SYS_PTRACE \ --security-opt seccomp=unconfined \ -itd \ harbor.baai.ac.cn/flagrelease-public/flagrelease-hygon-minimax:202604120035 \ /bin/bash -c ' HIP_VISIBLE_DEVICES="2,3,4,5" \ OMP_NUM_THREADS=32 \ NCCL_MIN_NCHANNELS=16 \ NCCL_MAX_NCHANNELS=16 \ VLLM_V1_FAST_TOKEN_ID_COPY=1 \ NCCL_P2P_LEVEL=LOC \ HSA_FORCE_FINE_GRAIN_PCIE=1 \ HIP_PERF=1 \ SAFETENSORS_FAST_GPU=1 \ vllm serve /models/QuantTrio/MiniMax-M2.5-AWQ \ --trust-remote-code \ --served-model-name minimax-m2.5 \ --max-model-len 196608 \ --max-num-batched-tokens 32768 \ --dtype bfloat16 \ --gpu_memory_utilization 0.90 \ --max-num-seqs 10 \ --tensor-parallel-size 4 \ --port 8000 \ --enable-auto-tool-choice \ --tool-call-parser minimax_m2 \ --reasoning-parser minimax_m2 \ --disable-custom-all-reduce ' ```