Your current environment

Four servers, each equipped with two 80GB A800 GPUs, running vLLM version 0.17.0. The image startup script is as follows:

`# 1 docker run -itd --entrypoint /bin/bash --network host --ipc=host
--name ray_vllm_node --shm-size 10.24g --gpus all
-v /nfsai/ai:/mllm_models
-e GLOO_SOCKET_IFNAME=bond0 -e NCCL_SOCKET_IFNAME=bond0
-e TP_SOCKET_IFNAME=bond0 -e NCCL_DEBUG=TRACE
-e VLLM_LOG_LEVEL=DEBUG
swr.cn-north-4.myhuaweicloud.com/ddn-k8s/docker.io/vllm/vllm-openai:v0.17.0

2、header

ray start --head --port=6379 --dashboard-host='0.0.0.0' --disable-usage-stats

3、worker

ray start --address=headerip:6379`

vllm server script is as follows: nohup python3 -m vllm.entrypoints.openai.api_server \ --model /mllm_models/MiniMax/MiniMax-M2.5 \ --served-model-name MiniMax-M2.5 \ --tensor-parallel-size 2 \ --pipeline-parallel-size 4 \ --enable-expert-parallel \ --enable-auto-tool-choice \ --max-model-len 32000 \ --max_num_seqs 32 \ --enable-chunked-prefill \ --enable-prefix-caching \ --trust-remote-code \ --host 0.0.0.0 \ --port 8001 \ --dtype bfloat16 \ --quantization fp8 \ --gpu-memory-utilization 0.87 \ --tool-call-parser minimax_m2 \ --chat-template /mllm_models/MiniMax/MiniMax-M2.5/chat_template.jinja \ --reasoning-parser minimax_m2_append_think \ --distributed-executor-backend ray > ./minimax-m2.5-20260316.log 2>&1 &

🐛 Describe the bug

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.