[Bug]: cpu core 100%

[nvliajia]

Your current environment

In vllm version 0.8.3, I launched the deepseek-32B model using four 4090 cards. Even without inference, the CPU usage rate was at 100%. Is this normal?

🐛 Describe the bug

In vllm version 0.8.3, I launched the deepseek-32B model using four 4090 cards. Even without inference, the CPU usage rate was at 100%. Is this normal?

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.

[Aveline67]

same for me

  0[|  0.6%]  4[   0.0%]  8[|100.0%] 12[   0.0%] 16[|  0.6%] 20[   0.0%] 24[   0.0%]              26[   0.0%] 30[   0.0%] 34[   0.0%] 38[   0.0%] 42[   0.0%] 46[   0.0%] 50[   0.0%]
    1[   0.0%]  5[   0.0%]  9[   0.0%] 13[   0.0%] 17[   0.0%] 21[   0.0%] 25[   0.0%]              27[   0.0%] 31[   0.0%] 35[|  0.6%] 39[   0.0%] 43[   0.0%] 47[   0.0%] 51[   0.0%]
    2[   0.0%]  6[   0.0%] 10[|  0.6%] 14[   0.0%] 18[   0.0%] 22[   0.0%]                          28[   0.0%] 32[   0.0%] 36[   0.0%] 40[   0.0%] 44[   0.0%] 48[   0.0%]
    3[|100.0%]  7[   0.0%] 11[   0.0%] 15[   0.0%] 19[|  0.6%] 23[   0.0%]                          29[   0.0%] 33[   0.0%] 37[   0.0%] 41[   0.0%] 45[   0.0%] 49[   0.0%]
  Mem[||||||||||||||||||||||||||||||||||||||||||                                      6.32G/94.0G] Tasks: 29, 269 thr, 0 kthr; 3 running
  Swp[                                                                                   0K/24.0G] Load average: 2.08 2.04 2.00
                                                                                                   Uptime: 02:35:25

  [Main] [I/O]
    PID USER       PRI  NI  VIRT   RES   SHR S  CPU%▽MEM%   TIME+  Command
    627 avlin       20   0  137G 2759M  649M R  91.6  2.9 50:09.24 /home/avlin/.env/bin/python3 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=32, pipe_handle=25) --multipro    628 avlin       20   0  137G 2766M  649M R  91.0  2.9 50:10.21 /home/avlin/.env/bin/python3 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=32, pipe_handle=28) --multipro      1 root        20   0 21968 12896  9984 S   0.0  0.0  0:00.97 /sbin/init                                                                                                                            127 root        19  -1 49360 13268 12436 S   0.0  0.0  0:03.12 /usr/lib/systemd/systemd-journald
    663 avlin       20   0  137G 2766M  649M S   0.0  2.9  0:04.76 /home/avlin/.env/bin/python3 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=32, pipe_handle=28) --multipro    668 avlin       20   0  137G 2759M  649M S   0.0  2.9  0:04.82 /home/avlin/.env/bin/python3 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=32, pipe_handle=25) --multipro    701 avlin       20   0  137G 2759M  649M S   0.6  2.9  0:04.65 /home/avlin/.env/bin/python3 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=32, pipe_handle=25) --multipro    705 avlin       20   0  137G 2766M  649M S   0.0  2.9  0:04.87 /home/avlin/.env/bin/python3 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=32, pipe_handle=28) --multipro    711 avlin       20   0  137G 2766M  649M S   0.0  2.9  0:04.76 /home/avlin/.env/bin/python3 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=32, pipe_handle=28) --multipro   1973 avlin       20   0  6540  4992  3328 R   0.6  0.0  0:00.09 htop
      2 root        20   0  2776  1664  1664 S   0.0  0.0  0:00.00 /init
      7 root        20   0  2776  1664  1664 S   0.0  0.0  0:00.00 plan9 --control-socket 7 --log-level 4 --server-fd 8 --pipe-fd 10 --log-truncate
      8 root        20   0  2776  1664  1664 S   0.0  0.0  0:00.00 plan9 --control-socket 7 --log-level 4 --server-fd 8 --pipe-fd 10 --log-truncate
      9 root        20   0  2776  1664  1664 S   0.0  0.0  0:00.00 /init
     19 root        20   0  2784   416   416 S   0.0  0.0  0:00.00 /init
     20 root        20   0  2784   832   832 S   0.0  0.0  0:00.01 /init
     21 avlin       20   0  6872  5408  3328 S   0.0  0.0  0:00.03 -bash
     22 root        20   0  7752  4992  4576 S   0.0  0.0  0:00.00 /bin/login -f
     72 avlin       20   0  6836  5408  3328 S   0.0  0.0  0:00.01 -bash
    173 root        20   0 25596  6664  4992 S   0.0  0.0  0:00.13 /usr/lib/systemd/systemd-udevd
    306 systemd-re  20   0 21044 12480 10400 S   0.0  0.0  0:00.12 /usr/lib/systemd/systemd-resolved
F1Help  F2Setup F3SearchF4FilterF5Tree  F6SortByF7Nice -F8Nice +F9Kill  F10Quit

vLLM API server version 0.8.5.post1

[Aveline67]

only if i set tensor parrelelism

[Aveline67]

linked to #16660

[p12tic]

I believe this bug has been fixed by #16226. The PR contains patch and instructions how to apply the fix locally in docker.

[p12tic]

Fixed in #16226, can be closed. When a new version of vllm is released, add VLLM_SLEEP_WHEN_IDLE=1 environment variable to reduce power consumption.

[p12tic]

cc @njhill this issue can be closed

[njhill]

This should now be resolved by #16226.