Skip to content

cuda: GDN decode conv-chain fusion and alpha/beta mat-vec pair on by default (GGML_CUDA_GDN_FUSION=0 disables), bit-identical, +7.9% 9B decode on H200 - #398

Merged
bri-prism merged 2 commits into
prismfrom
cuda/gdn-fusion-default
Oct 25, 2026
Merged

cuda: GDN decode conv-chain fusion and alpha/beta mat-vec pair on by default (GGML_CUDA_GDN_FUSION=0 disables), bit-identical, +7.9% 9B decode on H200#398
bri-prism merged 2 commits into
prismfrom
cuda/gdn-fusion-default

Commits

  1. Commits on Oct 25, 2026