Skip to content

[CuTeDSL] fix mha support for thor #3323

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

Open
HuPengsheet wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[CuTeDSL] fix mha support for thor #3323

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
15 changes: 8 additions & 7 deletions examples/python/CuTeDSL/cute/blackwell/kernel/attention/fmha/fmha.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -284,6 +284,11 @@ def __init__(
self.num_regs_other = 32
self.buffer_align_bytes = 1024
self.arch = BaseDSL._get_dsl().get_arch_enum()
# SM101 (e.g. Thor) shares the SM100 TMEM load path; other arches use LdRed.
self.uses_sm100_softmax_tmem_load = (
(self.arch >= Arch.sm_100 and self.arch <= Arch.sm_100f)
or (self.arch >= Arch.sm_101 and self.arch <= Arch.sm_101f)
)

if self.arch >= Arch.sm_103:
assert self.enable_ex2_emulation == False, (
Expand Down Expand Up @@ -2260,9 +2265,7 @@ def softmax_step(
old_row_max = row_max
skip_softmax = cutlass.Boolean(False)
if whether_apply_mask:
if cutlass.const_expr(
self.arch >= Arch.sm_100 and self.arch <= Arch.sm_100f
):
if cutlass.const_expr(self.uses_sm100_softmax_tmem_load):
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
else:
tTMEM_LOADrMax = cute.make_rmem_tensor(
Expand Down Expand Up @@ -2308,9 +2311,7 @@ def softmax_step(
inplace_producer.commit()
inplace_producer.advance()
else:
if cutlass.const_expr(
self.arch >= Arch.sm_100 and self.arch <= Arch.sm_100f
):
if cutlass.const_expr(self.uses_sm100_softmax_tmem_load):
cute.copy(
tiled_tmem_load,
tTMEM_LOADtS[None, 0, None, None],
Expand Down Expand Up @@ -2731,7 +2732,7 @@ def softmax(
)
tmem_p_offset = self.tmem_p0_offset if stage == 0 else self.tmem_p1_offset
tStS_P = cute.make_tensor(tStS.iterator + tmem_p_offset, tStS_P_layout)
if cutlass.const_expr(self.arch >= Arch.sm_100 and self.arch <= Arch.sm_100f):
if cutlass.const_expr(self.uses_sm100_softmax_tmem_load):
tmem_load_atom = cute.make_copy_atom(
tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)),
self.qk_acc_dtype,
Expand Down