At this point, given the >4GB allocation bug was fixed, if possible, I'd like to see the SYCL backend integrated into koboldcpp. While Intel GPUs are supported by Vulkan, SYCL is much faster at prompt processing (4-7x in some cases, though Vulkan is getting faster.) Even if it's not included by default with koboldcpp (much like hipBLAS) it should be supported as a backend.
At this point, given the >4GB allocation bug was fixed, if possible, I'd like to see the SYCL backend integrated into koboldcpp. While Intel GPUs are supported by Vulkan, SYCL is much faster at prompt processing (4-7x in some cases, though Vulkan is getting faster.) Even if it's not included by default with koboldcpp (much like hipBLAS) it should be supported as a backend.