Is OpenVINO running in a heterogeneous/hybrid mode when GPU.0 is set on unsupported hardware like Iris Plus — splitting work between CPU and iGPU

[VK-10]

OpenVINO backend reports "using device CPU" even when OPENVINO_DEVICE=GPU.0 is set

Note: I am aware that Intel Iris Plus is not officially supported by OpenVINO for GPU inference.
This is an exploratory test to understand the actual device behaviour when GPU.0 is forced.

Hardware: Intel Iris Plus Graphics
OS: Windows 11 x64
OpenVINO version: 2026.1.0

What I did

Built llama.cpp with the OpenVINO backend and ran llama-bench with $env:OPENVINO_DEVICE="GPU.0" set before execution.

Observed behaviour

The log prints OpenVINO: using device CPU regardless of whether OPENVINO_DEVICE is set or not. However the benchmark numbers change significantly:

OPENVINO_DEVICE	pp128 (t/s)	tg128 (t/s)
unset (default)	94.52	15.78
GPU.0	172.48	18.93

Prompt processing nearly doubled when GPU.0 was set, which suggests the iGPU is being used in some capacity — but the log doesn't reflect this.