This project(Strata CUDA-to-SYCL) owner is not have strata required resource. My test scope is compile passed by icpx.

How to build (SYCL / Intel GPU)

This fork ports the CUDA device layer to oneAPI SYCL for Intel GPUs. The ported kernels live in src/kernels/sycl/; the CMake build selects the backend with a flag.

Requirements

• Intel oneAPI Base Toolkit + DPC++/C++ Compiler (icpx) — tested with 2026.1
• CMake 3.24+ (the Ninja generator is recommended)
• An Intel GPU is only needed to run; building does not need one
• Roughly 10 GB of free disk for the build

1. Open a oneAPI environment (puts icpx and the SYCL runtime on PATH)

• Windows: run setvars.bat from the oneAPI install, or open the "Intel oneAPI command prompt for Intel 64 Visual Studio".
• Linux: source /opt/intel/oneapi/setvars.sh

2. Configure — SYCL backend on, CUDA off (the two are mutually exclusive)

cmake -S . -B build -G Ninja -DSTRATA_ENABLE_SYCL=ON -DSTRATA_ENABLE_CUDA=OFF

-G Ninja is optional (any CMake generator works). Add -DSTRATA_NATIVE_EXPERTS=OFF to skip fetching llama.cpp/ggml when the CPU i-quant experts are not needed.

3. Build

cmake --build build

On success the configure step prints

-- Strata: SYCL kernels built (Intel GPU); add each ported kernel to this library as it lands

and the build produces the ported kernel library:

• Windows: build/strata_kernels_sycl.lib
• Linux: build/libstrata_kernels_sycl.a

Every ported kernel in src/kernels/sycl/ is compiled with -fsycl by icpx.

Status

• Kernels: every CUDA kernel has a SYCL port under src/kernels/sycl/, and the whole kernel library strata_kernels_sycl compiles and links with icpx.
• Engine: the full engine now builds on the SYCL backend. A STRATA_ENABLE_SYCL=ON configure produces strata.exe (plus strata_core, strata_engine and strata_prefill on oneMKL) through a CUDA-to-SYCL compatibility shim (include/strata/core/sycl_compat/cuda_runtime.h); strata.exe starts and runs its host setup.
• Not verified end to end: a full token generation needs the model pack (~38 GB) and a CPU with AVX-512 VNNI/VBMI (the engine's expert kernel requires them) - neither was available here - so generation on the Intel GPU has not been exercised. The runtime numerical correctness of the ported kernels is likewise not yet validated against the CUDA oracle.
• Device backend (verified): a SYCL build's strata-device runs and enumerates the real GPU through oneAPI SYCL - device 0: Intel(R) Arc(TM) B570 Graphics, 160 compute units, 9.641 GiB, driver 100.

Credits

• Model: Qwen3.8-Flash-Next by the Qwen team; compressed versions by ISTA-DASLab; Swift 1.5 by UkisAI. Their licenses apply to the model files.
• Built with parts of llama.cpp / ggml (MIT). Ideas from Splash, ninfer and HyperQwen. More in the details.