Optimizations for q40, q80, q5k, fix Cosmos VLM prefill - #92

Merged

jimw567 merged 8 commits into

gfx11from

lichang.mmq-opt-cont

Sep 17, 2026

Commits on Aug 25, 2026

HIP: batch RDNA3.5 Q4_0 J128 WMMA and prefetch X tiles
liangliangchang
committed
4e564f3

Commits on Aug 27, 2026

HIP: batch RDNA3.5 Q8_0 J128 WMMA and prefetch X tiles

liangliangchang
and
cursoragent
committed
4760c1a

Commits on Aug 28, 2026

HIP: batch RDNA3.5 Q5_K J128 WMMA like Q4_K

liangliangchang
and
cursoragent
committed
7050612
HIP: pipeline RDNA3.5 Q5_K tile loads

liangliangchang
and
cursoragent
committed
8a6872b

Commits on Sep 1, 2026

HIP: limit RDNA3.5 Q8_0 X-tile prefetch to J=128

liangliangchang
and
cursoragent
committed
ad9d427
HIP: keep RDNA3.5 block quants at ntx=1 for J=64 and J=96

liangliangchang
and
cursoragent
committed
2a6e364
HIP: drop the RDNA3.5 q8_0 J=64 -> J=96 widening

liangliangchang
and
cursoragent
committed
c0a6c7c

Commits on Sep 3, 2026

HIP: fold the RDNA3.5 batched WMMA into the generic q8_0 vec_dot

liangliangchang
and
cursoragent
committed
2ab083b