Issue with NPU recovery from Linux NPU Driver v1.13.0 release

[jwludzik]

Issue

We let a bug into v1.13.0 release where the intel_vpu kernel module driver is unable to recover NPU from broken workload because of unhandled error code from firmware. The issue might only occur when user runs inference that ends with ZE_RESUTL_ERROR_DEVICE_LOST return code

Solutions

1. Reload the driver to recover NPU

sudo rmmod intel_vpu
sudo modprobe intel_vpu

2. Or patch the kernel - https://lore.kernel.org/stable/20250401110807.415197-1-jacek.lawrynowicz@linux.intel.com/T/#m8426e2b416b4cfaee9fcbb5d344de5306b68265b
3. Or use "intel_vpu" from drm-tip - https://gitlab.freedesktop.org/drm/tip
4. Or use older firmware release than v1.13.0

Additional

In case of repeating ZE_RESUTL_ERROR_DEVICE_LOST error code from inference run despite the applying above workarounds, the model might not be supported by NPU. Please, check issues in OpenVINO repository - https://github.com/openvinotoolkit/openvino/issues

In the meantime my team works to patch the kernel. Sorry for inconvenience

[koush]

I assume you mean the bug was let into v1.16.0

[jwludzik]

The issue is in both versions - v1.13.0, v1.16.0 and every new release. The proposed solution is to patch the kernel or use the latest "intel_vpu" from drm-tip. The vanilla kernel patching is in progress