glm-4.7-flash本地部署

本地环境情况：

GPU：4090（48GB）* 2

RAM：512GB

使用docker运行多个模型实例，目标为了显存的复用

使用的镜像

vllm/vllm-openai:nightly

运行命令：

nginx 复制代码

docker run -d --name glm-4.7-flash-vllm \
  --gpus '"device=0,1"' \
  -v /home/ls/.cache/modelscope/hub/models/ZhipuAI/GLM-4.7-Flash:/app/models \
  --ipc=host \
  -p 8003:8000 \
  vllm/vllm-openai:nightly \
  --model /app/models \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.95 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 1 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --enable-auto-tool-choice \
  --max-model-len=32768 \
  --max_num_seqs=8 \
  --served-model-name glm-4.7-flash

不出意外肯定要报错了：

(APIServer pid=1) You can update Transformers with the command pip install --upgrade transformers. If this does not work, and the checkpoint is very new, then there may not be a release version that supports this model yet. In this case, you can get the most up-to-date code by installing Transformers from source with the command pip install git+https://github.com/huggingface/transformers.git [type=value_error, input_value=ArgsKwargs((), {'model': ...rocessor_plugin': None}), input_type=ArgsKwargs]

(APIServer pid=1) For further information visit https://errors.pydantic.dev/2.12/v/value_error

意思是：nightly预览版里面的transformers的版本还是太低了。。。

升级transformers版本

构建本地docker镜像

nginx 复制代码

mkdir -p ~/vllm-glm47 && cd ~/vllm-glm47

cat > Dockerfile << 'EOF'
FROM vllm/vllm-openai:nightly

RUN pip install --upgrade transformers accelerate -q

ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
EOF

docker build -t vllm-glm47:latest .

运行命令改成：

nginx 复制代码

docker run -d --name glm-4.7-flash-vllm \
  --gpus '"device=0,1"' \
  -v /home/ls/.cache/modelscope/hub/models/ZhipuAI/GLM-4.7-Flash:/app/models \
  --ipc=host \
  -p 8003:8000 \
  vllm-glm47:latest \
  --model /app/models \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.95 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 1 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --enable-auto-tool-choice \
  --max-model-len=32768 \
  --max_num_seqs=8 \
  --served-model-name glm-4.7-flash

当然如果想在docker bash升级测试也可以

nginx 复制代码

# 0. 进入bash环境
docker run -it --rm \
  --gpus '"device=1"' \
  -v /home/ls/.cache/modelscope/hub/models/ZhipuAI/GLM-4.7-Flash:/app/models \
  --ipc=host \
  -p 8003:8000 \
  --entrypoint /bin/bash \
  vllm/vllm-openai:nightly



# 1. 升级依赖
pip install --upgrade transformers accelerate

# 2. 验证版本
python -c "import transformers; print(transformers.__version__)"

# 3. 手动启动服务
python -m vllm.entrypoints.openai.api_server \
  --model /app/models \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 1 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --enable-auto-tool-choice \
  --max-model-len=32768 \
  --max_num_seqs=8 \
  --served-model-name glm-4.7-flash

如果运行爆显存

调整相关参数：--max-model-len=32768、--max_num_seqs=8