llama.cpp Release b11173 Fixes Apple Metal Graph Capture Bugs
llama.cpp released patch build b11173, bringing targeted fixes for Apple Metal graph capture and empty computational graph handling. The update handles zero-node graphs gracefully, cleans up capture counter logic, and improves capture error reporting. This patch improves stability and developer debugging when running local LLM inference on macOS and iOS devices via Apple Metal. It prevents unnecessary compute passes and provides clearer guidance when GPU capture fails. Specifically, the patch adjusts `ggml_metal_op_init` to emit GPU debug-group markers exclusively during captured compute steps and adds hints for `METAL_CAPTURE_ENABLED=1` in capture error messages.
## BACKGROUND
llama.cpp is an open-source inference engine designed for efficient Large Language Model execution across diverse hardware backends. Apple Metal is Apple's low-level graphics and parallel compute framework, enabling llama.cpp to leverage Apple Silicon GPUs for hardware acceleration.