llama.cpp Release b11048 Adds Metal Backend Support for qwen4exp Models
llama.cpp release b11048 adds Metal backend support for hidden component (HC) operator variants used by qwen4exp models. Specifically, it implements the gated hc_pre operator with a per-element sigmoid gate and the hc_post operator with identity mixing. This routine update expands hardware-accelerated LLM support on Apple Silicon devices for experimental Qwen model architectures. It ensures Mac and iOS users can run qwen4exp variants efficiently using native GPU acceleration. The patch addresses specific DSV4 HC operator variants, supporting cases where the combination tensor in hc_post is null (comb == nullptr). The release also provides pre-compiled binaries across macOS, iOS, Linux, Windows, and Android backends.
## BACKGROUND
llama.cpp is a high-performance open-source LLM inference framework written in C/C++ that enables local execution of large language models on consumer hardware. Its Metal backend is specifically optimized for Apple Silicon GPUs (M-series chips), allowing fast inference on macOS and iOS devices without requiring heavy external framework dependencies.