Repository navigation

rdna35 prefill optimizations other than mmq - #129

Draft

liangliangchang wants to merge 14 commits into

gfx11from

lichang.rdna35-prefill-quant-rope

Draft

rdna35 prefill optimizations other than mmq#129
liangliangchang wants to merge 14 commits into
gfx11from
lichang.rdna35-prefill-quant-rope

Commits on Oct 1, 2026
Commits on Oct 5, 2026
- hip: enable the causal attention fast path for 16x4 tiles
  liangliangchang
  committed
  f840324
- hip: extend RDNA3.5 quantizer interleaving to aligned row strides
  liangliangchang
  committed
  6c4d3b4
Commits on Oct 7, 2026
- hip: consolidate RDNA3.5 attention and prefill fusions
  liangliangchang
  committed
  5aa5aee
Commits on Oct 13, 2026
Commits on Oct 26, 2026
- hip: fuse RDNA3.5 vision bias and gated FFN conversions
  liangliangchang
  committed
  71d2d68
Commits on Oct 27, 2026
- hip: extend RDNA3.5 producer fusions to BF16
  liangliangchang
  committed
  667af94
Commits on Oct 28, 2026
- hip: reduce RDNA3.5 vision tensor traffic
  liangliangchang
  committed
  d79fcf2
Commits on Oct 29, 2026
- hip: extend shared vision fusions for Gemma
  liangliangchang
  committed
  5d2dac4
- hip: tune RDNA3.5 vision attention, copies and normalization
  liangliangchang
  committed
  257a478