Offloading models in LLVM have traditionally registered device code through custom descriptor structures, and the SYCL implementation being upstreamed from Intel’s LLVM fork followed the same pattern. This talk describes a recent effort to drop that layer for SYCL in LLVM, by extending and using LLVM’s Offload Binary format. The talk will cover four parts of the pipeline: the multi-image Offload Binary V2 Format; the clang-sycl-linker device linker and LLVM module splitting; greatly simplified SYCL device images registration/wrapping code; and the runtime path that consumes a single OffloadBinary blob with no descriptor struct at all. The talk will be useful for offloading toolchain developers, who will learn how the multi-image format works, how a device linker can target it, and how the SYCL device-code pipeline was simplified by building directly on LLVM's offloading infrastructure.