Hi,
Multiple AMD cards (7900xtx and 890m at least what I have) are again having the same old problem, 1 cpu core per gpu are having 100% utilization even when idle.
100,0% VLLM::EngineCor
I have the -e VLLM_SLEEP_WHEN_IDLE=1 \ in the docker run command.
The docker container is the latest: vllm/vllm-openai-rocm:latest as of today.
And vLLM is what it has, 0.29.
The host is Ubuntu 24.04 VM inside Proxmox host.
How to make it not consume electricity when doing nothing? (cpu)
This issue has been reported before: on some AMD/ROCm setups, vLLM’s EngineCore process can consume 100% CPU per GPU even when idle, despite setting VLLM_SLEEP_WHEN_IDLE=1. The root cause is often a ROCm/HSA async queue bug, not vLLM itself. A reliable workaround is to set the environment variable GPU_MAX_HW_QUEUES=1 in your Docker run command, which resolves the busy-wait and high idle CPU usage on affected AMD GPUs and ROCm versions. Example:
docker run ... -e VLLM_SLEEP_WHEN_IDLE=1 -e GPU_MAX_HW_QUEUES=1 ...
This workaround is confirmed by multiple users and is linked to ROCm issue #5706. If you still see high GPU utilization at idle, that is a separate ROCm bug, but the CPU busy-wait should be fixed with the above environment variable. Would you like more detail on the technical background or further troubleshooting steps?
Sources:
Hint: Mention @Herald in the post for followups.
Sorry but my docker run command already contained -e GPU_MAX_HW_QUEUES=1 \ and -e VLLM_SLEEP_WHEN_IDLE=1 \
so no help with that
@Herald
Setting GPU_MAX_HW_QUEUES=1 and VLLM_SLEEP_WHEN_IDLE=1 addresses known ROCm idle CPU/GPU utilization issues, but if you’re still experiencing high usage, the root cause may be unresolved in your vLLM/ROCm version or related to other ROCm bugs. Upgrading to vLLM ≥ 0.21.0 has resolved the 100% CPU idle issue for many users, but persistent GPU load at idle may be due to ROCm driver issues (see ROCm/ROCm#5706), not vLLM itself. Workarounds like GPU_MAX_HW_QUEUES=1 help, but may not fully resolve all hardware/driver combinations according to this issue.
If you are already on vLLM 0.21.0+ and still see the problem, further mitigation may require ROCm driver updates or additional environment tweaks. Would you like more detail on troubleshooting steps or environment diagnostics?
Sources:
Hint: Mention @Herald in the post for followups.
I use the latest docker, so I suppose it has the latest rocm. OR does it matter what the host vm has in terms @Heraldf drivers? @Herald
Yes, it matters what ROCm drivers are installed on the host VM. The vLLM Docker image includes the ROCm userland and libraries, but the host must have a compatible ROCm kernel driver and GPU support for the container to access the hardware. The container alone is not sufficient—host drivers must match or be compatible with the ROCm version in the container for proper GPU access and stability. See official docs and ROCm installation guide.
Would you like more detail on how to check or match host and container ROCm versions?
Sources:
Hint: Mention @Herald in the post for followups.
@Herald
yes how to check the host drivers and compare them to the docker ones and then reinstall if needed on host
It seems likely that you can check your host NVIDIA driver version using nvidia-smi on the host machine, which will display the installed driver and supported CUDA version. Inside the Docker container, you can also run nvidia-smi (if the container is started with the correct NVIDIA runtime) to verify that the container sees the host drivers. If the versions do not match or you encounter errors like “unsupported display driver / cuda driver combination,” you may need to update or reinstall the host drivers.
To reinstall or update the host NVIDIA drivers, follow the official NVIDIA instructions for your OS and GPU model (see NVIDIA’s driver download page). After installation, reboot and verify with nvidia-smi. For more on resolving driver/container mismatches, see the troubleshooting steps in this vLLM issue and official vLLM Docker docs.
Would you like more detailed step-by-step instructions for your specific OS?
Sources:
Hint: Mention @Herald in the post for followups.
Solved by downgrading vLLM from 0.29 to 0.26
vllm/vllm-openai-rocm:v0.26.0
Not anymore 100% cpu 1 core utilization while idling.
Hope vLLM team will fix this issue in the future vLLM versions.