TL;DR: The llama.cpp release b9503 fixes compatibility with Gemma 4 audio models by correctly handling the audio projector embedding size.
Summary: This release updates llama.cpp to handle the Gemma 4 audio projector embedding size, resolving a compatibility issue with audio projector models. The update removes projection_dim from clip_n_mmproj_embd to align with the new embedding size structure. Additionally, it contains build fixes such as skipping cvector-generator and export-lora when CPU backends are disabled.
Why it matters: Developers building local audio-capable agents or multimodal apps using Gemma 4 can now run these models directly on edge devices using llama.cpp. Keep an eye on llama.cpp releases for updated multimodal support and verify compatibility when compiling with disabled CPU backends.
Source: github.com