CYBER WATCH / EQUINOXE INTELLIGENCE

← All articles

CUDA: fuse NVFP4 gate/up matmuls and GLU in MMQ (#31296)

llama.cpp Releases · 2026-11-05 08:00:35

This update enhances CUDA support for NVFP4 by fusing gate/up and GLU operations in MMQ, honoring requested precision, sharing scale checks, skipping unused FP4 kernels, testing fused MMQ W4A8 without scales and with batches, cleaning up naming, and checking gate_scale before loading the fused gate scale.

Read original source ↗