Community
llama.cpp PR fuses Qwen and Llama FFN gate/up matmuls in CUDA for 2 to 11 percent faster prefill
The pull request interleaves gate and up rows in MMQ kernels. Benchmarks on DGX Spark (GB10) and RTX 5090 show prompt processing gains of roughly 2 to 11 percent across Qwen3.6-27B, Qwen3.8-27B and Llama-3.1-8B quantizations.