llama.cpp

e62fa13c - model : make Gemma 4 shared-KV tail attn_k tensors optional on load (#21739)