Repository navigation

[Triton/Gluon] [gfx942] Support fp8 KV caches and fp8 dots in sparse_mla_fwd on gfx942 - #6199

Open

jin-amd wants to merge 4 commits into

jin-amd:gfx942-sparse-mla-fp8

Commits on Oct 8, 2026
- [Triton/Gluon] [gfx942] Read the fp8_scalar cache as e4m3fnuz on gfx942
  
  jin-amd
  and
  cursoragent
  committed
  e6d746e
- [Triton/Gluon] [gfx942] Run fp8 dots on gfx942
  
  jin-amd
  and
  cursoragent
  committed
  a83bacc
- [Triton/Gluon] [gfx942] Name the fp8 switches in sparse MLA traces, bench fp8 dots on gfx942
  
  jin-amd
  and
  cursoragent
  committed
  2782fad
- [Triton/Gluon] [gfx942] Refuse non-e4m3 byte caches, test the P_SCALE softmax tail
  
  jin-amd
  and
  cursoragent
  committed
  013f408