intel-gpu-device-selection

name	intel-gpu-device-selection
description	Select the Intel GPU device to use when a system has multiple Intel GPU devices. Use this skill when the user wants to run a workload on Intel GPU, mentions device selection, ZE_AFFINITY_MASK, or when multiple Level Zero GPU devices are detected and one must be chosen.
license	MIT
metadata	{"XPU":"Intel GPU"}

Intel GPU Device Selection

Select a single Intel GPU device via ZE_AFFINITY_MASK when multiple Level Zero GPU devices are present.

This skill does NOT set ZE_FLAT_DEVICE_HIERARCHY. On PVC, the default is FLAT (each tile is a separate device). On other platforms, the default is COMPOSITE (tiles merged into one device). If the user needs to override this, they must set it explicitly before running this skill.

Instructions

Step 1: Detect Level Zero GPU devices

sycl-ls | grep -i level_zero

If sycl-ls is not available, use the "source-oneapi" skill first.

Only consider lines matching [level_zero:gpu][level_zero:<index>].

Step 2: Select the target device

• 0 devices found: Stop. Report: "No Level Zero GPU device was found. Check GPU driver, Level Zero runtime, and oneAPI environment."
• 1 device found: Use its index. Proceed to Step 3.
• Multiple devices found: If the list contains both discrete and integrated GPUs, select the discrete GPU by default. If there are multiple discrete GPUs, ask the user to select. Do NOT proceed until a valid index is determined.

To distinguish dGPU from iGPU, run the bundled script:

python3 .claude/skills/action/intel-gpu-device-selection/scripts/l0_igpu_check.py

Output labels each GPU as iGPU or dGPU based on the Level Zero ZE_DEVICE_PROPERTY_FLAG_INTEGRATED flag.

Prompt format when user selection is needed:

Multiple Level Zero GPU devices found:
<device list from sycl-ls>

Which device should be used? Enter the device index (e.g., 0):

Step 3: Set ZE_AFFINITY_MASK

export ZE_AFFINITY_MASK=<selected index>