Skip to content

Fix to #652: SM100 Grouped GEMM: read the TMEM accumulator before releasing the overlapping stage #654

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
Anerudhan merged 1 commit into from
Aug 20, 2026
Merged

Fix to #652: SM100 Grouped GEMM: read the TMEM accumulator before releasing the overlapping stage #654

Show file tree
Hide file tree
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
6 changes: 3 additions & 3 deletions python/cudnn/gemm/cutedsl/grouped/dsrelu/moe_blockscaled_grouped_gemm_dsrelu_quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2150,16 +2150,16 @@ def kernel(
if reverse_subtile:
real_subtile_idx = self.cta_tile_shape_mnk[1] // self.epi_tile_n_required - 1 - subtile_idx

tTR_tAcc_mn = tTR_tAcc[(None, None, None, real_subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tAcc_mn, tTR_rAcc)

if cutlass.const_expr(self.overlapping_accum):
if subtile_idx == self.iter_acc_early_release_in_epilogue:
cute.arch.fence_view_async_tmem_load()
with cute.arch.elect_one():
acc_pipeline.consumer_release(acc_consumer_state)
acc_consumer_state.advance()

tTR_tAcc_mn = tTR_tAcc[(None, None, None, real_subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tAcc_mn, tTR_rAcc)

# Apply alpha scaling
if cutlass.const_expr(self.vectorized_f32):
for i in cutlass.range_constexpr(0, cute.size(tTR_rAcc), 2):
Expand Down
6 changes: 3 additions & 3 deletions python/cudnn/gemm/cutedsl/grouped/quant/grouped_gemm_quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1854,6 +1854,9 @@ def kernel(
if reverse_subtile:
real_subtile_idx = self.cta_tile_shape_mnk[1] // self.epi_tile_n_required - 1 - subtile_idx

tTR_tAcc_mn = tTR_tAcc[(None, None, None, real_subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tAcc_mn, tTR_rAcc)

# C1 fix: fence + early release for overlapping_accum
if cutlass.const_expr(self.overlapping_accum):
if subtile_idx == self.iter_acc_early_release_in_epilogue:
Expand All @@ -1862,9 +1865,6 @@ def kernel(
acc_pipeline.consumer_release(acc_consumer_state)
acc_consumer_state.advance()

tTR_tAcc_mn = tTR_tAcc[(None, None, None, real_subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tAcc_mn, tTR_rAcc)

if cutlass.const_expr(self.enable_bias):
# m7 fix: use real_subtile_idx directly (matches contiguous)
sBias_sub = sBias_subtiles[(None, real_subtile_idx)]
Expand Down
6 changes: 3 additions & 3 deletions python/cudnn/gemm/cutedsl/grouped/srelu/moe_blockscaled_grouped_gemm_srelu_quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1905,16 +1905,16 @@ def kernel(
if reverse_subtile:
real_subtile_idx = self.cta_tile_shape_mnk[1] // self.epi_tile_n_required - 1 - subtile_idx

tTR_tAcc_mn = tTR_tAcc[(None, None, None, real_subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tAcc_mn, tTR_rAcc)

if cutlass.const_expr(self.overlapping_accum):
if subtile_idx == self.iter_acc_early_release_in_epilogue:
cute.arch.fence_view_async_tmem_load()
with cute.arch.elect_one():
acc_pipeline.consumer_release(acc_consumer_state)
acc_consumer_state.advance()

tTR_tAcc_mn = tTR_tAcc[(None, None, None, real_subtile_idx)]
cute.copy(tiled_copy_t2r, tTR_tAcc_mn, tTR_rAcc)

if cutlass.const_expr(self.enable_bias):
# m7 fix: use real_subtile_idx directly (matches contiguous)
sBias_sub = sBias_subtiles[(None, real_subtile_idx)]
Expand Down