Standalone ggml ships CUDA graphs off. Capture/replay batches every
kernel launch of a graph into a single submission. Override with
-DGGML_CUDA_GRAPHS=OFF or at runtime with GGML_CUDA_DISABLE_GRAPHS=1
- Add sycl to the backend link loop so ggml-sycl links when built statically
- Add -fsycl link option for consumer executables (ggml-sycl links the
SYCL runtime PRIVATE, so consumers need it on their own link line)
- Gate _FORTIFY_SOURCE=2 with NOT GGML_SYCL: the fortified __memcpy_chk
symbol is unresolvable in SPIR-V device code and aborts kernel compilation
- Add buildsycl.sh convenience script (icx/icpx compilers)
Tested on Intel Arc A310 (DG2) with oneAPI 2026.0 + Level Zero: full TTS
generation produces correct audio output on the GPU.
Pascal
smashingtags