llama.cpp b9503 fixes Gemma 4 audio projector embedding size bug
Google's Gemma 4 now runs properly on local hardware with this patch.
Deep Dive
Llama.cpp's b9503 release tackles a Gemma 4 audio fix: the multimodal model's audio projector embedding size is now handled correctly, with projection_dim removed from clip_n_mmproj_embd. Co-authored by Xuan Nguyen.
Key Points
- Fixes Gemma 4 audio projector embedding size mismatch (projection_dim) in clip_n_mmproj_embd.
- Addresses issue #24091 reported on llama.cpp GitHub, co-authored by Xuan Son Nguyen.
- Available immediately for all supported platforms: macOS, Linux, Windows, Android, and CUDA/Vulkan backends.
Why It Matters
Local multimodal AI inference gets a critical stability fix for Gemma 4 audio processing.