You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'm sharing this here because my environment (Intel XPU / SYCL) is not covered by JamePeng's prebuilt wheels — and this fork, with its vision handlers, has proven essential for my ComfyUI workflows.
Throughout the previous release iterations, I've maintained good communication with JamePeng, and the issues we discussed were ultimately resolved. This also convinced me that, for real-world ComfyUI usage, JamePeng's fork of llama-cpp-python (with vision handlers) is truly essential.
My environment is Intel XPU / SYCL, which is not covered by the prebuilt wheels in this repository. So I'm placing the related content here — you can read and use it directly, or hand it to an AI to generate execution steps / a skill from it.
Vision handlers are the differentiator: This fork ships Qwen3VLChatHandler, Qwen25VLChatHandler, and other MTMD vision handlers — the foundation for Qwen-series image/video captioning in ComfyUI. Other llama-cpp-python sources don't have them.
No local patches needed since 0.3.45: The SYCL flash-attention multi-turn garble fix (llama.cpp PR #25741 / #25880) is now natively merged upstream.
DLL handling matters: The wheel (Plan A, ~36 MB) excludes the oneAPI runtime DLLs (dnnl.dll, mkl_*.dll, tbb12.dll); they're provided by the installed oneAPI. Load the oneAPI environment (setvars.bat) in your ComfyUI startup script — no sycl-preloader plugin needed since 0.3.42.
Runtime alignment: Prebuilt wheels (0.3.43+) are compiled with oneAPI 2026.1. If you also use PyTorch XPU for image generation, keep PyTorch XPU ≥ 2.13 for ABI alignment.
Full Guides (attached)
The complete step-by-step build guides are attached to this post:
sycl-build-guide-en.md — English build guide (setup → compile → verify → ComfyUI integration → known issues)
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Throughout the previous release iterations, I've maintained good communication with JamePeng, and the issues we discussed were ultimately resolved. This also convinced me that, for real-world ComfyUI usage, JamePeng's fork of llama-cpp-python (with vision handlers) is truly essential.
My environment is Intel XPU / SYCL, which is not covered by the prebuilt wheels in this repository. So I'm placing the related content here — you can read and use it directly, or hand it to an AI to generate execution steps / a skill from it.
Environment at a Glance
Key Points
Qwen3VLChatHandler,Qwen25VLChatHandler, and other MTMD vision handlers — the foundation for Qwen-series image/video captioning in ComfyUI. Other llama-cpp-python sources don't have them.dnnl.dll,mkl_*.dll,tbb12.dll); they're provided by the installed oneAPI. Load the oneAPI environment (setvars.bat) in your ComfyUI startup script — nosycl-preloaderplugin needed since 0.3.42.Full Guides (attached)
The complete step-by-step build guides are attached to this post:
--
All reactions