AnalysisDevelopersOctober 25, 2026

llama.cpp PR fuses FFN gate/up matmuls in MMQ CUDA path

Read original source →github.com

Pull request #29948 by anujj fuses FFN gate/up matmuls and GLU in MMQ with interleaved gate/up rows. On a DGX Spark (GB10) with llama-bench at 16K prompt, Qwen3.8-27B Q4_K_M rose from 840.1 to 899.9 tok/s (+7.1%).

1 source

More stories today

Open the live feed