Troubleshoot Ascend NPU in a Virtual Machine
Use this page when an Ascend NPU is exposed to a virtual machine or a virtualized node. The checks here describe limitations observed in a 310P AlaudaOS test environment. They do not define the capabilities of a physical Ascend node.
TOC
Recognize the environmentKnown VM limitationsWhat to verify before opening a product issueWhen a vNPU workload failsEscalate with the right evidenceRecognize the environment
Confirm that the workload node is a virtual machine through the platform's node administration path. Then check the guest NPU health and the resources reported to Kubernetes:
If the guest reports a healthy NPU and Kubernetes reports an allocatable resource, whole-card runtime workloads may still be usable even when vNPU operations are unavailable.
Known VM limitations
In the validated 310P AlaudaOS VM environment:
npu-smican report a healthy NPU while vNPU creation is rejected by the guest driver;device-sharecannot be changed from inside the VM;- static vNPU creation can fail before a virtual device is created;
- dynamic vNPU workloads can remain Pending or fail during virtual-device creation;
- a device-plugin enumeration error can be specific to the guest environment and does not by itself prove a physical-node driver failure.
These observations are environment-specific. Do not use them to declare static or dynamic vNPU unsupported on physical 310P, 910B, or other Ascend hardware.
What to verify before opening a product issue
Check the layers separately:
- The guest reports the expected NPU and health state.
- The node reports the resource key requested by the workload.
- The device-plugin Pod is Ready and its logs show the resource registration.
- A whole-card workload using the Ascend runtime can complete an ACL memory allocation and a small NPU operation.
- The failure is reproducible after the test Pod is deleted and the device is released.
Do not treat Pod Running, a /dev/davinci* device, or npu-smi info alone as proof that a workload can use the NPU.
When a vNPU workload fails
Record the exact allocation mode, resource key, admitted RuntimeClass, scheduler, Pod events, and the guest NPU status. If the failure occurs only in the VM, repeat the same workload on a supported physical node before reporting a general hardware or plugin defect.
Keep the following cases separate:
- whole-card allocation through the Ascend runtime;
- static vNPU creation and discovery;
- dynamic vNPU creation and cleanup;
- HAMi soft slicing and device-share requirements.
A successful whole-card workload does not prove that every vNPU mode is available in the VM.
Escalate with the right evidence
Include the product versions, chip model, guest OS and kernel, resource key, Pod events, device-plugin logs, npu-smi info, and the ACL or application result. State explicitly whether the test ran on a physical node or a VM.
For physical-node installation and driver preparation, return to Get Started with NPU Operator. For HAMi owner switching and Ascend slicing, follow the HAMi Ascend Device Plugin documentation.