llama.cpp SYCL backend removed Nvidia and AMD GPU support in 2026.02
[major] framework
Fields
id
gotchas/xpu-sycl-removed-nvidia-amd-supportClass
framework
Severity
major
Concepts
sycloneapillama.cpp
Symptom
As of the 2026.02 release, the llama.cpp SYCL backend no longer supports Nvidia or AMD GPUs. Previously, the SYCL backend could target any GPU via the oneAPI plugin for Nvidia and AMD, but the download/installation channels for that plugin are no longer available.
Root cause
The oneAPI plugin for Nvidia and AMD GPUs became unavailable because the download and installation channels stopped working. Users can no longer build the software for Nvidia or AMD GPUs using the SYCL backend.
Workaround
For Nvidia GPUs, use the CUDA backend. For AMD GPUs, use the HIP backend. The SYCL backend is now Intel-GPU-only. This change was made because the oneAPI plugin for non-Intel GPUs was not maintained and the installation channels were unreliable.
notes
This is a significant change for anyone who used the SYCL backend as a universal GPU backend. The SYCL backend is now Intel-GPU-only. For multi-vendor GPU support, use the Vulkan backend (which supports all GPUs but is not the fast path) or vendor-specific backends (CUDA for Nvidia, HIP for AMD).
Notes
This is a significant change for anyone who used the SYCL backend as a universal GPU backend. The SYCL backend is now Intel-GPU-only. For multi-vendor GPU support, use the Vulkan backend (which supports all GPUs but is not the fast path) or vendor-specific backends (CUDA for Nvidia, HIP for AMD).