常用命令

2026-04-28

查看进程使用时间

ps -efp xxxx -o lstart,etime

在线检测ModelID

DS FAQ

挂代理

export http_proxy=http://agent.baidu.com:8891
export https_proxy=http://agent.baidu.com:8891

curl - AS

curl http://10.178.29.70:80/v2/chat/completions  -H "Content-Type: application/json" -H "Authorization: Bearer bce-v3/ALTAK-seG05tfjoB4V83dl1I60P/df902ea834c2c21559ac5feb3a5720dc175c9733"  -d '{"model": "aiak_bzz2_deepseek_v32_rd_test2", "safety_level":"none", "stream_options": {"include_usage": true}, "max_tokens": 1, "temperature": 0 ,"stream": true, "messages": [{"content":"why we say 42 is the end answer?", "role": "user"}]}'

curl 机内

curl http://127.0.0.1:8999/v1/completions -H 'Authorization: Bearer EMPTY' -H 'Content-Type: application/json' -d '{ "model": "deepseek_v32", "prompt": "今天几号？", "max_tokens": 1, "temperature": 0}'

get_models

kubectl get feddeployment | awk '{print $1}' | sed '1d' | xargs -I {} sh -c "echo {}; kubectl get feddeployment -oyaml {} | sed -n '/MODEL_ID/{n;p;}' | grep 'value'"

kubectl -n pd-test get feddeployment | awk '{ print $1 }' | sed '1d' | xargs -I {} sh -c "echo {}; kubectl -n pd-test get feddeployment -oyaml {} | sed -n '/MODEL_ID/{n;p;}' | grep 'value'"

docker 登录

docker login -u liyanzhen01 -p uuap@Onmyown599 iregistry.baidu-int.com

pip 内网 install

pip install -i http://mirrors.baidubce.com/pypi/simple/ --trusted-host=mirrors.baidubce.com

SGLANG 官方启动 PD 分离

nohup python3 -m sglang.launch_server --model-path /ssd1/models/Qwen3-30B-A3B --disaggregation-mode prefill --port 12345  --disaggregation-ib-device mlx5_bond_0 --disable-radix-cache > /tmp/prefill_sglang.log 2>&1 &

nohup python3 -m sglang.launch_server --model-path /ssd1/models/Qwen3-30B-A3B --disaggregation-mode decode --port 12346  --base-gpu 1 --disaggregation-ib-device mlx5_bond_0  --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}}' --context-length 131072 > /tmp/decode_sglang.log 2>&1 &

nohup python3 -m sglang.srt.disaggregation.mini_lb --prefill http://127.0.0.1:12345 --decode http://127.0.0.1:12346 --host 0.0.0.0 --port 8081 > /tmp/sglang.log 2>&1 &

macOS 的docker build

docker buildx build --platform linux/amd64 \
-t iregistry.baidu-int.com/lyz/gpu-dev-env:latest \
--push \
-f ./scripts/Dockerfile .

循环测试不同输入的 P 的性能

for ifv in 131070 98300 32700 65530 32000 16380 13310 10240 5120; do
  echo "Running with -if $ifv ..."
  aiakperf -r openai_chat \
    -a http://10.215.110.12:8000 \
    -M aiak_deepseek_h200_v3_rd_test1 \
    -m /ssd3/models/huggingface.co/deepseek-ai/DeepSeek-V3/DeepSeek-V3.2-Exp \
    -w 1 -n 100 \
    -if $ifv \
    -of 1 \
    -d chat \
    -D ~/DATASET/aiak_dataset/144k.jsonl \
    > ~/result_if_${ifv}.log 2>&1
done

占 GPU 算力避免被超发任务

git clone https://github.com/wilicc/gpu-burn.git
cd gpu-burn
# 2. 编译
make
# 3. 跑 10 分钟，把所有 GPU 都打满
./gpu_burn

导出冲突的DB二进制文件

git show :3:data.db > data_local.db
git show :2:data.db > data_remote.db

切换cuda-toolkit 版本

ls -ld /usr/local/cuda*
# 强制指定 CUDA 根目录为 12.9 
export CUDA_HOME="/usr/local/cuda-12.9" 
# 将 12.9 的编译器路径临时提到 PATH 的最前面 
export PATH="${CUDA_HOME}/bin:${PATH}" 
# 将 12.9 的库路径临时提到 LD_LIBRARY_PATH 的最前面 
export LD_LIBRARY_PATH="${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}"