GLM4.5V部署

conda create -n glm45v python=3.10
conda activate glm45v

pip install -U vllm --pre --extra-index-url https://wheels.vllm.ai/nightly -i https://pypi.tuna.tsinghua.edu.cn/simple
          
pip install transformers-v4.55.0-GLM-4.5V-preview -i https://pypi.tuna.tsinghua.edu.cn/simple

启动脚本

#!/bin/bash

export VLLM_USE_MODELSCOPE="true"
export VLLM_WORKER_MULTIPROC_METHOD="spawn"
export NCCL_SOCKET_IFNAME="bond0"
export GLOO_SOCKET_IFNAME="bond0"
export GLOO_DEVICE_TRANSPORT="TCP"
export NCCL_P2P_DISABLE="1"
export NCCL_NVLS_DISABLE="1"
export NCCL_SHM_DISABLE="1"
export NCCL_ALGO="Ring"
export NCCL_C4A_DISABLE="1"
export NCCL_PROTO="LL,LL128"
export CUDA_VISIBLE_DEVICES=2,3,4,5
python -m vllm.entrypoints.openai.api_server  \
    --host 0.0.0.0 \
    --port 8088 \
    --model /data/modelscope/models/ZhipuAI/GLM-4___5V/ \
    --served-model-name glm45v \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.80 \
    --enable-auto-tool-choice \
    --tool-call-parser glm45

nohup sh /opt/vllm/vllm-glm45v.sh  > /var/log/vllm-glm45v.log 2>&1 &

GLM4.6V部署

conda create -n glm46vrun python=3.12
conda activate glm46vrun 


pip install vllm==0.13.0 --root-user-action=ignore
pip install transformers==5.0.0rc0 --root-user-action=ignore

pip install modelscope>=1.18.1 --root-user-action=ignore

# 安装官方仓库下面的所有依赖
https://github.com/zai-org/GLM-V/blob/main/requirements.txt

启动脚本

#!/bin/bash

export VLLM_USE_MODELSCOPE="true"
export VLLM_WORKER_MULTIPROC_METHOD="spawn"
export NCCL_SOCKET_IFNAME="bond0"
export GLOO_SOCKET_IFNAME="bond0"
export GLOO_DEVICE_TRANSPORT="TCP"
export NCCL_P2P_DISABLE="1"
export NCCL_NVLS_DISABLE="1"
export NCCL_SHM_DISABLE="1"
export NCCL_ALGO="Ring"
export NCCL_C4A_DISABLE="1"
export NCCL_PROTO="LL,LL128"
export CUDA_VISIBLE_DEVICES=2,3,4,5
python -m vllm.entrypoints.openai.api_server  \
    --host 0.0.0.0 \
    --port 8088 \
    --model /data/modelscope/models/ZhipuAI/GLM-4___5V/ \
    --served-model-name glm45v \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.80 \
    --enable-auto-tool-choice \
    --tool-call-parser glm45

qwen3-coder-480b 部署

pip install -U vllm --pre --extra-index-url https://wheels.vllm.ai/nightly -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install modelscope>=1.18.1
nohup sh /opt/vllm/vllm-coder.sh  > /var/log/vllm-coder.log 2>&1 &

tail -400f /var/log/vllm-coder.log