llamacpp:sycl — first-class Intel GPU backend

[TravisJ33304]

Summary

Lemonade documents Intel GPUs as Vulkan, but /system-info never detects intel_gpu, so llama.cpp never auto-selects a GPU on Arc. We want a first-class llamacpp:sycl variant (same shape as CUDA/ROCm/Vulkan).

Why SYCL, not only Vulkan

On Intel Arc Pro B70, llama.cpp SYCL is the validated high-performance path.
Vulkan on Arc has quality issues (garbled output unless GGML_VK_DISABLE_F16).
Vulkan can remain the no-oneAPI fallback.

Scope

• Detect intel_gpu (Linux: PCI 0x8086, xe/i915; prefer discrete over iGPU).
• sycl support row + llamacpp.sycl_bin BYO.
• Linux: do not download a Vulkan zip as SYCL. Builtin install only if a real
  ggml-org llama-*-sycl* asset exists; otherwise require sycl_bin.
• Launch: --device SYCL0, oneAPI/Level Zero env.
• Extend get_global_vram_usage_pct() for Intel so auto_evict works.
• Tests: C++ helpers + hardware server_llm.py --backend sycl (no GitHub-hosted SYCL CI).

Hardware

We can test on Intel Arc Pro B70 (Linux, xe driver).

Maintenance

1. CI can be used for automated tests that assert Intel GPU detection fixtures, parsers, and other code logic.
2. Follow the same convention as CUDA for inference testing by having manual inference tests run on local hardware.
3. Use contributors with Intel hardware such as me (hexzcracker on Discord) as a contact for testing when bumping backend version pins and reproducing bugs.
4. Don't add an sycl hardware matrix row to CI until an Intel self hosted runner exists.
5. Linux SYCL stays BYO (sycl_bin). That avoids lemonade-sdk having to publish/pin SYCL binaries the way they do CUDA/ROCm.

[jeremyfowers]

@TravisJ33304 I converted your issue to an RFC (see new policy landing soon: #3521 )

Your description above could use a note about maintainership: how do we ensure sycl continues to work after the contribution merges?

@kenvandine please review this RFC. Reference implementation available here: #3509

[TravisJ33304]

Thanks for the response and consideration @jeremyfowers. I added a section to the RFC covering a proposed maintenance plan.

CC @kenvandine