RDNA3.5 (gfx11 / gfx1151) MMQ prefill optimizations#32

Open

liangliangchang wants to merge 14 commits into

liangliangchang:lichang.mmq_opt

Commits on Jul 10, 2026

Fix Q6_K B-tile prefetch using wrong K offset in vec_dot

liangliangchang
and
cursoragent
committed
Improve RDNA3.5 MMQ occupancy, LDS layout, and activation pipelining

authored and
Jim Wu
committed
Improve RDNA3.5 Q4_K MMQ dequant/vec_dot and add phase profiling

authored and
Jim Wu
committed
cuda: optimize RDNA3.5 tiny-M MMQ for gate Q8_0 prefill

authored and
Jim Wu
committed
Improve RDNA3.5 Q6_K MMQ prefill and narrow-N tile sizing

authored and
Jim Wu
committed
cuda: fix RDNA3.5 MMQ codegen regression after profiling cleanup

authored and
Jim Wu
committed
Fix Q6_K B-tile prefetch using wrong K offset in vec_dot

authored and
Jim Wu
committed

Commits on Jul 14, 2026