gfx11 MoE experiments#908

Draft

mgehre-amd wants to merge 23 commits into

gfx11from

matthias.moe-perf-opt-experiments

Commits on Apr 25, 2026

Commits on Apr 27, 2026

Add experimental MoE megakernel for single-token AWQ-int4 decode
mgehre-amd
committed
Add experimental HIP MoE megakernel: fused GEMM2 + topk-weighted reduce
mgehre-amd
committed
moe_megakernel: keep slot partials in LDS, drop grid-wide barrier
mgehre-amd
committed
moe_megakernel: bump WvPrGrp from 4 to 8 (256-thread workgroups)
mgehre-amd
committed
moe_megakernel: bump WvPrGrp from 8 to 12 (384-thread workgroups)
mgehre-amd
committed
moe_megakernel: bump WvPrGrp from 12 to 16 (512-thread workgroups)
mgehre-amd
committed
moe_megakernel: widen A_CHUNK from 16 to 32 (4x DOT2 ops per inner unroll)
mgehre-amd
committed

Commits on Apr 28, 2026