Intel GPU AI Skills plugin
Agent skills for running, benchmarking and profiling Hugging Face models on Intel Arc GPUs, plus CUDA to XPU code migration.
- 51 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- The host setup skill installs driver packages and Docker and changes render group rights.
- Skills launch containers and model servers on your machine.
Install
- Works in
Manual install
/plugin marketplace add intel/gpu-ai-skillsThen `/plugin install <skill-name>@intel-gpu-ai-skills`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The host setup skill installs driver packages and Docker and changes render group rights.
- Skills launch containers and model servers on your machine.
This is third-party code. Review the repository files before installing.
What it does
A collection of Agent Skills for working with arbitrary Hugging Face models (safetensors and GGUF) on Intel Arc, Arc Pro and Battlemage GPUs. The skills cover PyTorch with Transformers, vLLM for XPU, SGLang for XPU and llama.cpp on SYCL. The set is split into stages. Host setup: `xpu-system-setup`, `xpu-discover`, `xpu-container-run`. Planning: `model-can-it-fit` estimates VRAM from `config.json`, `xpu-deploy-plan` assembles one launch plan. Running: `torch-xpu-run`, `vllm-xpu-run`, `sglang-xpu-run`, `llamacpp-xpu-run`. Benchmarking and profiling follow.
Who it is for. Engineers who bring up local models on Intel GPUs and want the agent to choose the image and flags and verify the launch.
Good fit when
- A clean Ubuntu host needs the driver, `xpu-smi` and Docker prepared for Intel Arc
- You want to know whether Qwen2.5-32B in int4 fits an Arc Pro B70 at 8K context with four requests
- Code written for CUDA is being ported to `torch.xpu`
- A GGUF model is served through `llama-server` without Python or PyTorch
- An OpenAI-compatible vLLM server is needed for an Intel GPU
Not a fit when
- You have an NVIDIA or AMD card: the skills target Intel XPU
- The model is a diffusion model, which `model-can-it-fit` refuses to estimate
- You need a skill for a hosted inference platform instead of your own hardware
Example request
Will Qwen2.5-32B in int4 fit on my Arc Pro B70 at 8K context and concurrency 4? If so, draft a vLLM deployment plan.Limitations
The skills target Intel Arc, Arc Pro and Battlemage cards with XPU support. Some recommendation skills, such as `model-config-recommend`, are marked experimental by the author. Memory estimates rely on formulas over `config.json` and do not replace a measurement on your model. The host setup skill installs driver packages and Docker and changes render group rights, which needs elevated access. The repository moves fast and vLLM and SGLang image and flag versions change quickly, so check them against current documentation.
How to disable. Delete the installed skill folders from your agent directory or disable the plugin through `/plugin`. The install script supports reversal with the `--uninstall` flag.
FAQ
How do I install Intel GPU AI Skills in Claude Code?
Run `/plugin marketplace add intel/gpu-ai-skills`, then `/plugin install <skill-name>@intel-gpu-ai-skills` for the skill you need.
Which agents are supported?
Claude Code, Cursor, Gemini CLI, OpenAI Codex, GitHub Copilot CLI, opencode, Qwen Code, Kimi Code, Hermes Agent and OpenClaw. `scripts/install.sh` places skills in every agent directory it finds.
How do `vllm-xpu-run` and `llamacpp-xpu-run` differ?
The first serves a model through the vLLM OpenAI-compatible HTTP API, the second runs GGUF through `llama-server` on SYCL without Python or PyTorch.
Can it estimate whether a model fits in memory?
Yes, `model-can-it-fit` estimates weights, KV cache and activations from `config.json` for decoder-only LLMs, MoE and VLMs. It refuses diffusion models.
Is there CUDA code migration?
Yes, the `cuda-to-xpu-migration` skill with an inventory script and a migration report template. `torch-xpu-run` also translates CUDA code to `torch.xpu`.
Discussion
No questions or reviews yet.
Related
PluginsA GPU profiler for AI agents: the agent runs vLLM or SGLang under graphsignal-run, reads the metrics and tunes flags for speed
13 NVIDIA skills for an agent on a Jetson device: diagnostics, memory, LLM serving, packages and video and inference benchmarks
Official AMD skills teach an agent to set up ROCm, serve LLMs on Instinct and EPYC, profile GPUs and quantize models.
The official GPUtw skill and MCP server: an agent rents a GPU in Taiwan, runs a command and shuts the instance down
Skill that powers AutoDL instances on and off from a plain-language request, keeps a budget and shuts down idle machines.
Inference Autopilot tunes SGLang launch parameters for your GPU and workload and returns a launch command backed by benchmarks.
README in short
The README lists skills by stage: setup, planning, running, benchmarking and profiling. It then covers installation for several agents plus testing and security notes.