[vllm dump] need to toggle all the subprocesses before criu dump, it's too slow

i just create a vllm service with 2 A10 GPU and use cuda-checkpoint & criu to dump/restore. the model i use is Qwen2.5-72B.

i use the command to start-up:
python3 -m vllm.entrypoints.openai.api_server --model /home/yunzhu.lyq/model14b --host 0.0.0.0 --port 8000 --max-num-seqs 64 --swap-space 4 --gpu-memory-utilization 0.90 --tensor-parallel-size 2 --trust-remote-code --disable-custom-all-reduce >> output.log 2>&1

well, the service contains 5 sub process:

but when i try to dump it, i have to toggle all the processs, which wastes a lot of time, like:
time ./cuda-checkpoint --toggle --pid 1874873
time ./cuda-checkpoint --toggle --pid 1875769
time ./cuda-checkpoint --toggle --pid 1875770
time ./cuda-checkpoint --toggle --pid 1876114
time ./cuda-checkpoint --toggle --pid 1876115

then i can dump them successfully:
time criu dump --shell-job --images-dir images --tree 1874873 --ghost-limit $((10 * 1024 * 1024)) --link-remap --tcp-established

Is this the expected behavior? how can i get faster?

Start by investigating how cuda-checkpoint --toggle --pid handles a process and whether the command or its process traversal supports descendants. Reproduce the five-process vLLM setup with the commands shown, then verify that one toggle operation covers all subprocesses before the existing criu dump command runs.