Repository navigation

cuda: gate (SwiGLU) fused PTQ1_0 mat-vec for 2-4 columns - #310

Open

sb32445 wants to merge 6 commits into

PrismML-Eng:prismfrom

sb32445:pr/ptq1-gate-up-fuse-mc

Commits on Oct 4, 2026
- cuda: gate (SwiGLU) fused PTQ1_0 mat-vec for 2-4 columns
  
  sb32445
  and
  claude
  committed
  c7866ac
- cuda: keep the gate/up SwiGLU fusion when the GLU output overlaps the q8 rows
  
  sb32445
  and
  claude
  committed
  15b61ce
- cuda: reset the q8 rows registry pointer after the graph evaluation
  
  sb32445
  and
  claude
  committed
  bdb0083
- cuda: remove the GGML_CUDA_PTQ1_FUSE_MC and GGML_CUDA_FWHT_GLU_POOL switches
  
  sb32445
  and
  claude
  committed
  1fde296
Commits on Oct 7, 2026
- cuda : require a contiguous bias for the fused PTQ1_0 mat-vec with 2 to 4 columns
  sb32445
  committed
  830c19b
- tests : add gate fused PTQ1_0 mat-vec cases with 2 to 4 columns
  sb32445
  committed
  791eb61