ggerganov/llama.cpp Released b11684 with Metal Fusion Test Fixes
Llama.cpp release b11684 introduces test fixes for vanished rows in the Metal fusion baseline and updates dummy models. It ensures that missing fusion matches are correctly reported as regressions rather than being silently ignored. Maintaining robust test infrastructure ensures that optimizations on Apple Silicon hardware remain reliable and performance regressions are caught early. This prevents silent failures in operator fusions that accelerate local model inference. The update re-records the Metal fusion baseline to 277 rows, fixes softmax gating for the Mellum dummy model, and adds the muse-spark model to the test baseline. These changes target the test suite to prevent issues where model architecture updates cause fusions to stop matching without failing the CI job.
## BACKGROUND
llama.cpp is an open-source software library designed to perform efficient local inference on large language models using the GGML tensor library. The Metal framework is Apple's graphics and compute API, which llama.cpp leverages to accelerate AI model inference on Apple Silicon devices like MacBooks.