Skip to content

tune FP8 tile sizes with two-level accumulation #125

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
LucasWilkinson merged 3 commits into from
Mar 16, 2026
+32 −13
Merged

tune FP8 tile sizes with two-level accumulation #125

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
192c71a
tune FP8 tile sizes with two-level accumulation
jmkuebler Mar 6, 2026
e1efa3e
refactor: reorganize FP8 tile_size.h and add hd96 tuning
jmkuebler Mar 9, 2026
98fb960
Merge branch 'main' into tune-fp8-tile-sizes
jmkuebler Mar 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
45 changes: 32 additions & 13 deletions hopper/tile_size.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,24 +51,43 @@ constexpr std::tuple<int, int, bool, bool> tile_size_fwd_sm90(
return {128, is_local ? 64 : 80, true, true}; // 128 x 80 hits the limit of smem
}
} else {
if (headdim <= 64) {
if (use_one_mma_wg) {
// FP8 path
if (use_one_mma_wg) {
// Decode tiles — independent of two-level accumulation setting
if (headdim <= 96) {
return {64, 128, true, true};
} else {
return {192, 160, true, true};
}
} else if (headdim <= 96) {
return {192, 128, true, true};
} else if (headdim <= 128) {
if (use_one_mma_wg) {
return {64, 96, true, true};
} else{
return {128, paged_kv_non_TMA ? 160 : (v_colmajor || (softcap && is_local) ? 192 : 224), true, true};
}
} else if (headdim <= 192) {
return {128, (paged_kv_non_TMA || softcap) && is_local ? 128 : 160, true, true};
} else {
return {128, is_local ? 64 : 128, true, !paged_kv_non_TMA}; // PagedKV uses more registers so we disabled IntraWGOverlap
// Prefill tiles — two-level accumulation needs smaller tiles to reduce
// register pressure from the separate fp32 accumulator (tOrO_accum).
// Currently just optimized for causal case.
#ifndef FLASHATTENTION_DISABLE_FP8_TWO_LEVEL_ACCUMULATION
if (headdim <= 64) {
return {192, 128, true, true};
} else if (headdim <= 96) {
return {128, 128, true, true};
} else if (headdim <= 128) {
return {128, 192, true, true};
} else if (headdim <= 192) {
return {128, 96, true, true};
} else {
return {128, is_local ? 64 : 128, true, !paged_kv_non_TMA}; // TODO: FP8 prefill ~0.54x of BF16 with two-level accum at hd256
}
#else
if (headdim <= 64) {
return {192, 160, true, true};
} else if (headdim <= 96) {
return {192, 128, true, true};
} else if (headdim <= 128) {
return {128, paged_kv_non_TMA ? 160 : (v_colmajor || (softcap && is_local) ? 192 : 224), true, true};
} else if (headdim <= 192) {
return {128, (paged_kv_non_TMA || softcap) && is_local ? 128 : 160, true, true};
} else {
return {128, is_local ? 64 : 128, true, !paged_kv_non_TMA};
}
#endif
}
}
}
Expand Down