Repository navigation

perf(attention): fast fp8 kv decode and prefill on rdna3 (gfx11) - #899

Draft

zihaomu wants to merge 4 commits into

FlashML-org:mainfrom

zihaomu:perf/rdna3-fp8-kv-pr897-followup-20261112

Draft

perf(attention): fast fp8 kv decode and prefill on rdna3 (gfx11)#899
zihaomu wants to merge 4 commits into
FlashML-org:mainfrom
zihaomu:perf/rdna3-fp8-kv-pr897-followup-20261112

Commits on Nov 12, 2026
- perf(attention): add packed fp8 kv reads on gfx1201
  zihaomu
  committed
  29c98e7
- refactor(attention): encapsulate packed fp8 kv readers
  zihaomu
  committed
  252528e
- fix(attention): isolate packed fp8 kv reads to amd
  zihaomu
  committed
  01f3b47
- perf(attention): fast fp8 kv path for rdna3 (gfx11)
  zihaomu
  committed
  cf44001