Repository navigation

CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 - #24546

Merged

pwilkin merged 7 commits into

ggml-org:masterfrom

ravel7524:mmqx-rdna3-routed-moe-tiling

Sep 7, 2026

Merged

CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3#24546
pwilkin merged 7 commits into
ggml-org:masterfrom
ravel7524:mmqx-rdna3-routed-moe-tiling

Commits on Aug 5, 2026
Commits on Aug 6, 2026
- refactor: replace moe_ncols_min_cc with use_typical_moe_ncols in mmq configuration files
  ravel7524
  committed
  2adc35b
Commits on Aug 24, 2026
- HIP: mmq: enable typical moe ncols on RDNA4
  IMbackK
  committed
  115a4bd
Commits on Aug 30, 2026
- Merge branch 'master' of https://github.com/ggml-org/llama.cpp into mmqx-rdna3-routed-moe-tiling
  ravel7524
  committed
  7d8bc65