菜单
Inventory Intel GPUs (Arc, Arc Pro, Data Center GPU Max) on a Linux host. Detect devices, check driver health, list processes using each XPU, run a quick diagnostic, and read live utilisation.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/intel/gpu-ai-skills --skill xpu-discover命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Recommend how to configure vLLM-XPU for a Hugging Face decoder-only LLM on Intel Arc B-series GPUs: choose quantization, KV dtype, DP/TP layout, max concurrency, and max context using roofline math against published hardware specs. Use when the user asks "How should I configure vLLM?", requests the best vLLM configuration, or asks how to use one or multiple Arc cards. Must run recommend.py; predictions are physics-bounded ranges, not measured throughput. Use after xpu-discover and before vllm-xpu-run.
Run an arbitrary Hugging Face safetensors model on an Intel GPU using **upstream PyTorch** (>= 2.8) with the built-in `torch.xpu` device. Covers loading from the Hub, picking the right dtype, autocast, multi-GPU with accelerate's `device_map`, and the CUDA -> XPU code translation a user has to do once. Use for the Transformers / Accelerate / Diffusers path. Not for OpenAI-compatible serving (use vllm-xpu-run); explicitly not via intel-extension-for-pytorch (ipex) or ipex-llm — those paths are end-of-life and upstream PyTorch supersedes them.
Benchmark a **running vLLM-XPU OpenAI-compatible server** on an Intel GPU using `vllm bench`. Measures TTFT (time-to-first-token), TPOT (time-per-output-token), ITL (inter-token latency), end-to-end latency, and throughput under concurrency. Covers online (`vllm bench serve`) and offline (`vllm bench throughput`) modes; concurrency sweeps and quant comparison live in `references/sweep-and-compare.md`. Use after **vllm-xpu-run** when the user asks "how fast is this?".
基于 SOC 职业分类
正在显示 SKILL.md
| name | xpu-discover |
| description | Inventory Intel GPUs (Arc, Arc Pro, Data Center GPU Max) on a Linux host. Detect devices, check driver health, list processes using each XPU, run a quick diagnostic, and read live utilisation. |
xpu-smi is Intel's nvidia-smi. Sees only Intel GPUs (Arc, Arc
Pro, Battlemage, Flex, Max).
Run in order. If step 1 is empty, stop.
xpu-smi discovery # 1. inventory
xpu-smi diag --precheck # 2. driver/firmware health
xpu-smi ps # 3. processes using each GPU
xpu-smi diag -d 0 -l 1 # 4. quick functional test
xpu-smi stats -d 0 # 5. utilisation snapshot
xpu-smi dump -d 0 -m 0,5,18 -i 1 # 6. live CSV stream (Ctrl-C)
All commands accept -j for JSON output (use when parsing).
| CUDA | Intel |
|---|---|
nvidia-smi | xpu-smi discovery |
nvidia-smi -L | xpu-smi discovery -j |
nvidia-smi pmon -c 1 | xpu-smi ps |
nvidia-smi dmon | xpu-smi dump -d <id> -m 0,5,18 -i 1 |
nvidia-smi --query-gpu=... | xpu-smi stats -d <id> -j |
nvidia-smi topo -m | xpu-smi topology -m |
CUDA_VISIBLE_DEVICES=0 | ZE_AFFINITY_MASK=0 |
cuda-memcheck | xpu-smi diag -d 0 -l 1 |
CUDA refugee footgun: CUDA_VISIBLE_DEVICES=99 silently hides
all GPUs; ZE_AFFINITY_MASK=99 crashes the Level Zero loader
with an assertion. Always check xpu-smi discovery for valid IDs
(start at 0) before setting the mask.
discovery — inventoryOne stanza per Intel GPU. Key fields:
Device ID — small integer, used as -d and as
ZE_AFFINITY_MASK value.
PCI BDF Address — stable across reboots (e.g. 0000:36:00.0).
DRM Device — /dev/dri/card0, used in --device for Docker.
Device Name — Battlemage shows Intel(R) Graphics [0xe2XX]
rather than the marketing name; driver quirk, not a problem.
Map the PCI device ID in brackets to the product SKU:
| PCI device ID | Product SKU | Confirmed |
|---|---|---|
0xe20b | Arc B580 | yes (lspci on hardware) |
0xe211 | Arc Pro B60 | yes (pci.ids) |
0xe220 | Arc Pro B50 | yes (pci.ids) |
0xe221 | Arc Pro B65 | yes (pci.ids) |
0xe223 | Arc Pro B70 | yes (lspci on hardware) |
Full table provided above. Cross-check with lspci -d 8086: -nn (prints [8086:XXXX]).
Empty output -> kernel didn't enumerate any Intel GPU. See "Troubleshooting".
diag --precheck — driver healthScans journalctl for known Intel-GPU error categories (GuC/HuC
firmware, IOMMU, PCIe, DRM, i915/Xe, Level Zero init).
Scope with --since today / --since yesterday /
--listtypes (show every error category).
ps — what's using each GPULists processes holding Level Zero handles + shared/device memory
in MiB. Desktop processes (plasmashell, xauth_*) are normal on
a workstation; only worry about a stale model server still
holding memory.
diag -d <id> -l <level> — does it compute| Level | Time | Impact |
|---|---|---|
-l 1 | seconds | safe on a busy box |
-l 2 | medium | impacts other workloads |
-l 3 | minutes | impacts performance |
Reading on a workstation: the Software Permission sub-test
fails when other processes already hold the device, so the
overall line says Fail even when Computation Check: Pass.
Read per-sub-test rows. For a clean run, log out of the GUI
session and run from TTY, or run inside a privileged container.
Pick individual tests with --singletest:
| ID | Test | When |
|---|---|---|
| 1 | Computation | does it compute |
| 2 | Memory Error | suspected ECC / bit-flip |
| 3 | Memory Bandwidth | sanity-check HBM/GDDR |
| 4 | Media Codec | video pipelines |
| 5 | PCIe Bandwidth | suspected slot/cable issue |
| 6 | Power | thermal/TDP investigation |
| 7 | Computation functional | quick sanity (lighter than 1) |
| 8 | Media Codec functional | lighter media check |
| 9 | Xe Link Throughput | multi-GPU peer link |
| 10 | Xe Link all-to-all | multi-GPU; needs -d -1 |
Example: xpu-smi diag -d 0 --singletest 1,3 -j.
stats -d <id> — utilisation snapshotMany fields show N/A on consumer Battlemage drivers (Arc Pro
B70 included) — counter-wiring limitation, not a bug. Memory
(column 5) and power (column 18) generally work. For
utilisation while running, prefer xpu-smi dump.
dump -d <id> -m <metrics> -i <interval> — live CSV streamUseful metric IDs for LLM serving:
0 GPU utilisation (%)5 GPU memory used (MiB)18 GPU power (W)xpu-smi dump -d 0 -m 0,5,18 -i 1 > xpu.csv &
# ... run your model ...
kill %1
Privilege note: metric 0 reads MEI telemetry, restricted on
consumer parts. Without sudo, that column shows N/A. Metrics
5 and 18 work unprivileged.
topology -m — multi-GPU connectivityMatrix of Xe Link / PCIe switch / hostbridge between XPU pairs. Only useful on multi-XPU systems.
When diag --precheck flags a critical error:
| Category | Cause + fix |
|---|---|
| Level Zero Init Error | Driver/userspace mismatch. Confirm xe (Battlemage) or i915 (older) modules loaded: lsmod | grep -E 'i915|xe'. Reload or reboot. |
| GuC / HuC Not Running | Missing firmware blob. Check dmesg | grep -i 'GuC|HuC'; install linux-firmware. |
| IOMMU Catastrophic | Kernel cmdline. On consumer boards: intel_iommu=on iommu=pt. |
| PCIe Error | Reseat card / check slot; re-run xpu-smi diag -d <id> --singletest 5. |
| DRM Error | Stuck context from a crashed desktop session. Logout/login (or reboot) clears it. |
| i915 Not Loaded on Battlemage | Battlemage uses xe, not i915. Confirm modinfo xe; precheck error is misleading on this generation. |
| Symptom | Fix |
|---|---|
xpu-smi: command not found | Install Intel level-zero packages (Ubuntu / RHEL / Arch all package xpu-smi). Binary lands at /usr/bin/xpu-smi. |
discovery empty but card present | (1) Kernel didn't bind: lspci -k -s <bdf> should show Kernel driver in use:. (2) Bound to vfio: lsmod | grep vfio. (3) Inside container without /dev/dri: add --device /dev/dri --privileged. |
diag fails inside container | Container needs --privileged for diag ioctls beyond the standard render-node interface. |
| Two XPUs present, one visible | printenv ZE_AFFINITY_MASK; unset for full inventory, re-export for workloads. |
| Variable | Purpose |
|---|---|
ZE_AFFINITY_MASK | Which XPU(s) a process sees (0, 0,1, ...). Invalid IDs crash the L0 loader (unlike CUDA's silent-hide). |
ZE_FLAT_DEVICE_HIERARCHY | FLAT exposes tiles as separate root devices; COMPOSITE (default) groups under one. Battlemage is single-tile, doesn't matter. |
ZE_ENABLE_VALIDATION_LAYER=1 | L0 loader prints API misuse — useful when something silently returns wrong device count. |
xpu-smi source: https://github.com/intel/xpumanager