CYBER WATCH / EQUINOXE INTELLIGENCE
← All articles
b12300: llama – removal of non‑fused GDN and batch‑size dependent graph nodes (#32078)
llama.cpp Releases · 2026-11-12 00:01:09
This update removes build_delta_net_chunking and build_delta_net_autoregressive, always uses the fused gated delta net, makes graph_max_nodes independent of batch size, and drops the fused GDN probes and fused node tags. Assisted by dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF on Mac Studio M5 Ultra 256 GB.Read original source ↗