cuda-to-xpu-migration
Create a CUDA-to-XPU migration assessment for an existing AI repo. Identify CUDA-specific assumptions, route to the right XPU skills, produce a migration report. Use when the user has a CUDA repo, notebook, Dockerfile, launch script, HF / vLLM / SGLang workload, or Triton kernel and asks to migrate it to Intel Arc / Arc Pro / Battlemage / XPU — including "convert this to XPU" / "move it to XPU" and the bare "migrate this repo" request where scope is not yet set. A request that says "port" routes to xpu-port. Plans and routes only. Not for executing an already-scoped rewrite (use xpu-port), running migrated code, or measuring it.
intel · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add intel/gpu-ai-skills --skill cuda-to-xpu-migration技能说明
name: cuda-to-xpu-migration description: Create a CUDA-to-XPU migration assessment for an existing AI repo. Identify CUDA-specific assumptions, route to the right XPU skills, produce a migration report. Use when the user has a CUDA repo, notebook, Dockerfile, launch script, HF / vLLM / SGLang workload, or Triton kernel and asks to migrate it to Intel Arc / Arc Pro / Battlemage / XPU — including "convert this to XPU" / "move it to XPU" and the bare "migrate this repo" request where scope is not yet set. A request that says "port" routes to xpu-port. Plans and routes only. Not for executing an already-scoped rewrite (use xpu-port), running migrated code, or measuring it.
cuda-to-xpu-migration
Create a migration assessment for CUDA-oriented repos that want an Intel XPU path.
Scope
This skill is for assessment and planning only.
- Identify CUDA-specific assumptions and likely migration surfaces.
- Reuse existing skills for code translation, runtime setup, serving, profiling, and sizing.
- Produce a clean migration report with recommended edits, blockers, and next steps.
- Do not execute or validate migrated code.
- Do not duplicate detailed commands or framework runbooks that are already covered by other skills.
Always assess; route forward once
This skill assesses the workflow verbs — "assess", "migrate", "convert", "move" — with the same read-only pass: inventory, classify, report. A request that says "port" (or otherwise names the rewrite explicitly) belongs to xpu-port, not here; route it there. For the workflow verbs, do not bounce the request away before the report exists: a mis-routed execution request costs one cheap read-only assessment, while the reverse mis-route would rewrite a user's repo unasked.
The report's Next steps (with the "Non-namespace CUDA surfaces → Route" table) is the single onward-routing authority. It names the executor for each surface — xpu-port for portable Python, xpu-deploy-plan for API-first / serving repos, container and runtime skills per tier. Routing flows one way, assess → execute; nothing here routes backward.
Use with
- xpu-discover for an XPU environment / driver preflight check before recommending a migration target.
- xpu-runtime-preflight for a read-only go/no-go check of host/container XPU readiness (drivers, /dev/dri, permissions, runtime, and essentials) before any XPU workload.
- torch-xpu-run for PyTorch / Transformers CUDA-to-XPU code translation guidance.
- xpu-container-run for container runtime setup on Intel GPU.
- vllm-xpu-run for vLLM launch and runtime guidance.
- sglang-xpu-run for SGLang launch and runtime guidance.
- model-can-it-fit and model-config-recommend for fit and config decisions.
- xpu-deploy-plan for a coordinated end-to-end serving plan (preflight → fit → config → launch → smoke test) once the assessment routes an API-first / serving-stack repo to a local XPU endpoint.
Migration flow
-
Inventory
- If the request is a git URL (not a local path), shallow-clone it to a stated scratch location first (e.g.
git clone --depth 1 <url> /tmp/<repo>-assess) and assess that path. State the clone target before scanning — do not assess a URL from memory. - If the user's request does not include a repo path, search the local workspace for the relevant scripts or repo (e.g.
find ~/workspace -name "*.sh" -path "*vllm*"). State which path you are assessing before running the scan — do not silently assume a location. - Run
scripts/inventory.shfrom the target repo root to surface CUDA/NVIDIA references and common launch/runtime surfaces. - Group findings by area: framework code, runtime/container, serving stack, Triton kernels, and CUDA-native dependencies.
- Notebooks:
inventory.shgreps.ipynbfiles as text, so CUDA references in notebook cells are surfaced. Notebook code cells are not directly rewritable: for a notebook-heavy repo, flag that code cells need extraction to.py(or hand-porting) before the port executes, and note it under "Needs version-aware review".
- If the request is a git URL (not a local path), shallow-clone it to a stated scratch location first (e.g.
-
Classify
- Safe translation candidates: portable PyTorch / Transformers code.
- Runtime changes: container flags, environment variables, launch scripts.
- Version-sensitive changes: vLLM / SGLang internals and plugin code.
- Likely blockers: CUDA extensions, CuPy, NCCL, TensorRT, flash-attn, bitsandbytes, cuBLAS / CUTLASS / cuTENSOR / cuSPARSE / cuDNN consumers, custom
.cu/.cuhcode. - Cloud-API / provider references (not CUDA): many "NVIDIA" hits are HTTP-client calls, not local GPU code — NVIDIA NIM (
integrate.api.nvidia.com,langchain-nvidia-ai-endpoints,build.nvidia.com), or unrelated cloud services (ElevenLabs, OpenAI). These do not need a CUDA port. Classify them as: provider swap (retarget a local vLLM-XPU / SGLang-XPU OpenAI-compatible endpoint and change theapi_base/ model name — no Python rewrite) or unaffected (GPU-agnostic cloud dependency, e.g. a TTS API). Do not count these toward the CUDA migration surface — the pdf-to-podcast blueprint had ~170 grep hits at the time of assessment (upstream is unpinned, so the exact count drifts) and zero actual CUDA code.
-
Route
- For XPU environment / driver preflight, use xpu-discover.
- For PyTorch / Transformers code translation, use torch-xpu-run.
- For container changes, use xpu-container-run.
- For vLLM / SGLang serving paths, use vllm-xpu-run or sglang-xpu-run.
- For deployment and capacity planning, use model-can-it-fit, and model-config-recommend.
- For non-namespace CUDA constructs, route per the per-construct guidance in the "Non-namespace CUDA surfaces" section of
templates/migration_report.md:torch.compileandpin_memory→ xpu-port (execution) / torch-xpu-run (reference);ProfilerActivity.CUDA→ torch-xpu-profile;set_float32_matmul_precision/ TF32 toggles