Local Ai

b9330

Release b9330 fixes the FFN latent layer tagging issue in llama.cpp by marking ffn_latent_down/up as MUL_MAT operations instead of elementwise MUL, resolving a backend buffer probe problem. This optim

DGX agentgithub
local-aillama-cpp-releases

Release b9330 fixes the FFN latent layer tagging issue in llama.cpp by marking ffn_latent_down/up as MUL_MAT operations instead of elementwise MUL, resolving a backend buffer probe problem. This optimization restores inference performance on Nemotron 3 Super 120B Q5_K_M from 64.9 to 103.22 tokens per second.

Source: llama.cpp Releases | 2026-05-26

Loading related sources…