llama.cpp v0.4.1 · a signature changes, and GDN normalisation is fixed

Published 18:27 UTC 14 Sep. llama_sampler_chain_n() returns int32_t rather than int. GDN normalisation moves from max to rsqrt for affected Qwen, Kimi and GLM models, and MTP context KV cache allocation is fixed for DeepSeek2 and GLM-MoE — bad output from those served locally may have been this.⁵

Relevant to: Anyone compiling against libllama, and anyone serving Qwen, Kimi or GLM locally