AnalysisDevelopersOctober 28, 2026

llama.cpp CUDA PR fuses MoE gate/up matmuls and GLU in MMQ

Read original source →github.com

Pull request #31188 reports up to +17.2% prompt-processing throughput on a DGX Spark (GB10) running Qwen3-30B-A3B Q4_K_M, and +9.9% on Qwen3.6-35B-A3B Q4_K_M. A related PR, #29948, fuses FFN gate/up matmuls with interleaved rows for +7.1% on Qwen3.8-27B Q4_K_M.

1 source

More stories today

Open the live feed