Intel GPU AI Skills Brings 20 Agent Skills for Hugging Face Models on Arc GPUs [2026]

By Indie Kings | October 1, 2026

Updated October 1, 2026: Intel published v0.1.0 of gpu-ai-skills on October 1, 2026, a collection of 20 Agent Skills for setting up, running, benchmarking and profiling Hugging Face models on Intel Arc, Arc Pro and Battlemage GPUs. Per the official README, it covers PyTorch plus Transformers, vLLM-XPU, SGLang-XPU and llama.cpp-SYCL, plus CUDA to XPU migration, and installs into Claude Code, opencode, Codex, Cursor, Copilot CLI, Gemini CLI and four more agents.

Intel gpu-ai-skills GitHub repository social preview card

Image: gpu-ai-skills repository preview. Credit: GitHub.

What the pack contains

Per the official README, gpu-ai-skills is a collection of Agent Skills in the format of the Agent Skills specification at agentskills.io. Per the same source, the 20 skills break into six groups: setup, plan, run, benchmark, profile and migrate.

Per the GitHub API, the repository was created August 21, 2026, carries the Apache-2.0 license, and stands at 26 stars and 11 forks at time of check. Version 0.1.0 was published October 1, 2026. This site covered vLLM on Arc B70 hardware before, but no prior post mentions Agent Skills, SKILL.md, or this repository, so this is new ground rather than a re-run.

Group	Skills	Job
Setup	4	Host, driver, container and preflight checks
Plan	3	VRAM fit estimates and deployment plans
Run	5	Serve models via PyTorch, vLLM, SGLang or llama.cpp
Benchmark	3	Measure TTFT, decode rate and throughput
Profile	3	Find hot ops and kernel level costs
Migrate	2	Assess and execute CUDA to XPU ports

Setup and plan skills

Per the README setup table, four skills get the host ready. The xpu-system-setup skill installs xpu-smi, the Level Zero runtime and Docker, configures the render group and Intel GPU PPA, and fixes Battlemage specific prerequisites. The xpu-discover skill inventories Intel GPUs and checks driver health, which the README calls the CUDA analogue of nvidia-smi. The xpu-runtime-preflight skill checks host, device, Docker, proxy, storage and container readiness. The xpu-container-run skill launches a Docker container with GPU access.

Per the README plan table, three skills size work before launch. The model-can-it-fit skill estimates VRAM from a Hugging Face config.json, covering decoder-only LLMs, MoE and VLM, and refuses diffusion. The model-config-recommend skill, marked experimental, recommends a vLLM-XPU deployment config using roofline math against Arc B-series specs. The xpu-deploy-plan skill chains preflight, fit, config and a runtime skill into one PLAN.md with a launch command, smoke test and rollback.

Run, benchmark and profile skills

Per the README run table, five skills put models on the GPU. The xpu-model-type-detect skill detects a model actual type before loading so the agent picks the right AutoModel class. The torch-xpu-run skill runs safetensors models via upstream PyTorch plus Transformers plus torch.xpu. The vllm-xpu-run skill serves through the vLLM-XPU OpenAI compatible HTTP API. The sglang-xpu-run skill serves through the SGLang XPU backend with RadixAttention prefix caching. The llamacpp-xpu-run skill runs GGUF models via the llama.cpp SYCL backend with no Python or PyTorch required.

Per the README benchmark and profile tables, measurement splits by framework too. Each of PyTorch, vLLM and SGLang gets a bench skill (TTFT, decode rate, peak memory, throughput at concurrency) and a profile skill (torch.profiler plus Kineto traces, vLLM start and stop profile endpoints, and SYCL and Level Zero kernel timing via unitrace). The README notes that descriptions disambiguate by deployment shape, so an agent should ask which framework is meant when a request just says benchmark.

Framework	Run skill	Bench skill	Profile skill
PyTorch	torch-xpu-run	torch-xpu-bench	torch-xpu-profile
vLLM-XPU	vllm-xpu-run	vllm-xpu-bench	vllm-xpu-profile
SGLang-XPU	sglang-xpu-run	sglang-xpu-bench	xpu-profile-unitrace
llama.cpp SYCL	llamacpp-xpu-run	Not listed	Not listed

Migrate skills for CUDA code

Per the README migrate section, two skills handle CUDA code. The cuda-to-xpu-migration skill assesses a CUDA repo, notebook, Dockerfile or launch script, produces a migration report and routes to the right XPU skills, and plans and routes only. The xpu-port skill executes a single target port with a libcst based scan, mechanical rewrite and CPU FP64 versus target dtype correctness check.

Per the README resources section, the NVIDIA equivalents live at vllm-project/vllm-skills and Hub workflows at huggingface/skills. The migration pair is the reason this pack matters beyond Intel owners: it is the on ramp the other two packs do not provide.

How to install it

Per the README install section, the recommended path is a local clone that installs into every detected agent skills directory at once.

Recommended: per README, run git clone on intel/gpu-ai-skills, then bash scripts/install.sh. Flags include all for agents not yet used, agent plus a name for one agent, and uninstall to reverse.
Claude Code: per README, add the marketplace then install a skill by name.
Gemini CLI: per README, install from the shipped gemini-extension.json, and uninstall old copies first.
Other agents: per README, copy skill folders into the agent skills directory. The table names Claude Code, opencode, OpenAI Codex, Copilot CLI, Cursor, Qwen Code, Kimi Code, Hermes Agent, OpenClaw and a generic AGENTS.md path.
Battlemage hosts: per README, run the check_battlemage_prerequisites.sh diagnostic first, with a fix flag for remediation.
Contributors: per README, new skills go under the skills directory from the SKILL.md template, validated by scripts/check-skills.sh and tests/static.sh.

Limits to know before use

Per the README, the fit estimator refuses diffusion models, and the config recommender is experimental. Per the same source, upgrading an older Gemini install requires a full uninstall first. The test suite notes 71 passed with 6 skipped without HF_TOKEN, and model configs are fetched from the Hub at pinned revisions rather than committed, so a config that cannot be fetched skips rather than fails.

Per the GitHub API at time of check, the project is young: created late August 2026, first tagged release October 1, 2026, one open issue. Treat skill counts and directory paths as moving targets and recheck the README before scripting installs.

FAQ

What is Intel gpu-ai-skills?

Per the official README, it is a collection of 20 Agent Skills for setting up, running, benchmarking and profiling Hugging Face safetensors and GGUF models on Intel Arc, Arc Pro and Battlemage GPUs, following the Agent Skills specification.

Which frameworks does it cover?

Per the README, PyTorch plus Transformers with torch.xpu, vLLM-XPU, SGLang-XPU and llama.cpp with the SYCL backend, plus CUDA to XPU migration skills. Diffusion models are refused by the fit estimator.

How do I install it?

Per the README, clone intel/gpu-ai-skills and run bash scripts/install.sh, which installs into every detected agent skills directory. Claude Code uses a plugin marketplace command and Gemini CLI uses the shipped extension file.

Does it work with my coding agent?

Per the README, it works with Claude Code, opencode, OpenAI Codex, Cursor, GitHub Copilot CLI, Gemini CLI, Qwen Code, Kimi Code, Hermes Agent and OpenClaw, plus a generic AGENTS.md path for anything else.

Can it migrate CUDA code to Intel GPUs?

Per the README, yes in two steps. The cuda-to-xpu-migration skill assesses a repo and produces a report, and the xpu-port skill executes a single target port with correctness verification against CPU FP64.

Is it stable enough to depend on?

Per the GitHub API, v0.1.0 shipped October 1, 2026 on a repo created August 21, 2026, with 26 stars and one open issue at time of check. Treat it as early and recheck the README before scripting anything, since paths and skill counts can move.

Bottom Line

Intel shipped the missing manual for its own GPUs in agent readable form: 20 skills that take a bare Ubuntu box from driver checks to a profiled serving deployment, with the migration pair as the standout. The vLLM and SGLang coverage overlaps our earlier Arc B70 post, but that post tuned servers by hand and this pack automates the same loop, so the two complement rather than repeat. Clone it, run the Battlemage diagnostic first if that is your card, and pin to v0.1.0 until the skill list settles.