Repository navigation

[Kernel][FA] Support paged FP8 Flash attention with asymmetric K/V head dim on gfx950 - #1066

Merged

coderfeli merged 59 commits into

ROCm:mainfrom

sammysun0711:upstream/fa-paged-fp8-asymmetric-v

Sep 14, 2026

Merged

[Kernel][FA] Support paged FP8 Flash attention with asymmetric K/V head dim on gfx950#1066
coderfeli merged 59 commits into
ROCm:mainfrom
sammysun0711:upstream/fa-paged-fp8-asymmetric-v

Commits on Sep 13, 2026
- feat(fmha): support paged FP8 D192 with asymmetric values
  sammysun0711
  committed
  45f1bb9
- test(fa): guard large offset test by available VRAM
  sammysun0711
  committed
  ef39d59
- fix(fmha): order paged copies on the launch stream
  sammysun0711
  committed
  490bb16
- fix(fmha): guard paged FP8 flattened dimensions
  sammysun0711
  committed
  ff77071
- refactor(fmha): clean up paged FP8 address helpers
  sammysun0711
  committed
  e40644e
- fix(fmha): preserve paged BF16 seqlen compatibility
  sammysun0711
  committed
  a3c0d34
- fix(fmha): validate paged output device
  sammysun0711
  committed
  2175c2f
- perf(fmha): reduce paged FP8 V192 register pressure
  sammysun0711
  committed
  4b04045
- feat(fmha): enable paged FP8 D128 attention
  sammysun0711
  committed
  5b8fe46
- perf(fmha): add paired-page D128 fast path
  sammysun0711
  committed
  2fc285e
- perf(fmha): relax paged D128 P*V accumulator anchors
  sammysun0711
  committed
  f56ab1f
- perf(fmha): vectorize paged D128 V fragment loads
  sammysun0711
  committed
  aa3126c
- perf(fmha): use permlane16 for paged D128 V staging
  sammysun0711
  committed
  3db264f
- perf(fmha): keep paged D128 permlane results in native order
  sammysun0711
  committed
  c554b1d
- perf(fmha): overlap paged D128 V prefetch
  sammysun0711
  committed
  b48cb40
- perf(fmha): defer paged D128 logit scaling
  sammysun0711
  committed
  fe33810
- perf(fmha): overlap paged D128 V loads with QK
  sammysun0711
  committed
  88ff0b9
- perf(fmha): defer paged D192 logit scaling
  sammysun0711
  committed
  9232c17
- perf(fmha): use direct FP8 P*V for paged V128
  sammysun0711
  committed
  7f4f709
- perf(fmha): cache compiled fp8 page attention
  sammysun0711
  committed
  7c02fdb
- perf(fmha): schedule paged BF16 V192 fragments on demand
  sammysun0711
  committed
  fc7f9b8
- perf(fmha): use segmented FP8 P*V for paged V192
  sammysun0711
  committed
  a3a4c4f
- perf(fmha): pipeline paged FP8 V prefetch
  sammysun0711
  committed
  9d07ca5
- perf(fmha): extend paged FP8 BN128 to varlen batches
  sammysun0711
  committed
  688884b
- perf(fmha): balance paged FP8 D192 batches
  sammysun0711
  committed
  c388f3b
- refactor(fmha): use typed paged FP8 helpers
  sammysun0711
  committed
  e82ca92
- refactor(fmha): remove unreachable paged FP8 split-K code
  sammysun0711
  committed
  fc99957
- refactor(fmha): remove unreachable paged FP8 non-causal paths
  sammysun0711
  committed
  484646f
- docs(fmha): trim paged FP8 schedule narration
  sammysun0711
  committed
  e92b34b
- refactor(fmha): attach compile hints to paged launcher
  sammysun0711
  committed
  fcb549e
- refactor(fmha): use architecture-dispatched paged FP8 waits
  sammysun0711
  committed
  4c0912b
- refactor(fmha): simplify static paged FP8 contracts
  sammysun0711
  committed
  dded5be
- fix(fmha): rebase and fix dense and paged FP8 trait integration
  sammysun0711
  committed
  50d98a1
- perf(fmha): extend paged BN128 to D192/V128
  sammysun0711
  committed
  c9e72fa
- perf(fmha): use an eight-slot paged D128 prefetch ring
  sammysun0711
  committed
  c15f035
- perf(fmha): add compact paired-page FP8 V192 pipeline
  sammysun0711
  committed
  f537e4e
- refactor(fmha): specialize paged FP8 layout contracts
  sammysun0711
  committed
  cfbdfb5
- refactor(fmha): use explicit-width paged FP8 arithmetic
  sammysun0711
  committed
  3791131
- test(fmha): share the paged FP8 reference calculation
  sammysun0711
  committed
  fa3cee0
- refactor(fmha): use public paged FP8 memory views
  sammysun0711
  committed
  752c52e
- refactor(fmha): share paged launch and compile preparation
  sammysun0711
  committed
  7148042
- docs(fmha): clean up paged FP8 implementation comments
  sammysun0711
  committed
  44302e9
- refactor(fmha): keep segmented LDS offsets in tile units
  sammysun0711
  committed
  4a72f6e
- perf(fmha): share a balanced maximum across paged FP8 tile pairs
  sammysun0711
  committed
  e6e7f86
- perf(fmha): split the D192 V128 unmasked attention prefix
  sammysun0711
  committed
  cf79b38
- perf(fmha): select scalar centering for native paged FP8
  sammysun0711
  committed
  9df3371
- perf(fmha): interleave paired D192/V128 batches in grid X
  sammysun0711
  committed
  9b291b2
- perf(fmha): prune bounded paged FP8 rescaling
  sammysun0711
  committed
  1c2a27a
- perf(fmha): interleave small paired D128 batches
  sammysun0711
  committed
  eb85926
- refactor(fmha): simplify paged FP8 pipeline helpers
  sammysun0711
  committed
  3cc57af
- refactor(fmha): centralize paged dispatch policy and test helpers
  sammysun0711
  committed
  9e8c5ef
- docs(fmha): clarify paged FP8 layout and scheduling invariants
  sammysun0711
  committed
  cacbae0
- refactor(fmha): adopt official LLVM memory primitives
  sammysun0711
  committed
  3bbbeef
- refactor(fmha): unify paged FP8 attention on official APIs
  sammysun0711
  committed
  497c76f
- perf(fmha): optimize native small-page FP8 attention
  sammysun0711
  committed
  9251049
- perf(fmha): streamline small-page FP8 data movement
  sammysun0711
  committed
  5196a22
- fix(fmha): harden paged FP8 contracts and use public FMA
  sammysun0711
  committed
  3579e32
Commits on Sep 14, 2026
- refactor(fmha): simplify paged FP8 scheduling and tests
  sammysun0711
  committed
  08ac7cb
- increase CI test timeout to 70 min
  sammysun0711
  committed
  8a662b0

Commits

Commits on Sep 13, 2026

Commits on Sep 14, 2026