Skip to content

[None][perf] CuTe DSL act-fusion GEMMs: key the early accumulator release on the loop index #21842

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
jvpflum wants to merge 1 commit into
base: main
Choose a base branch
from
+8 −2
Open

[None][perf] CuTe DSL act-fusion GEMMs: key the early accumulator release on the loop index #21842

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
5 changes: 4 additions & 1 deletion tensorrt_llm/_torch/kernels/blackwell/dense_blockscaled_gemm_act_fusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1551,6 +1551,9 @@ def kernel(
real_subtile_idx = subtile_idx // 2
else:
real_subtile_idx = subtile_idx
# Position in processing order. The early accumulator release below counts
# processed subtiles, so it must not use the reversed index.
iter_subtile_idx = real_subtile_idx
if cutlass.const_expr(self.overlapping_accum):
if reverse_subtile:
real_subtile_idx = (
Expand Down Expand Up @@ -1584,7 +1587,7 @@ def kernel(
# Async arrive accumulator buffer empty earlier when overlapping_accum is enabled
#
if cutlass.const_expr(self.overlapping_accum):
if real_subtile_idx == self.iter_acc_early_release_in_epilogue:
if iter_subtile_idx == self.iter_acc_early_release_in_epilogue:
# Fence for TMEM load
cute.arch.fence_view_async_tmem_load()
with cute.arch.elect_one():
Expand Down
5 changes: 4 additions & 1 deletion ...llm/_torch/moe/kernels/blackwell/blockscaled_contiguous_gather_grouped_gemm_act_fusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2501,6 +2501,9 @@ def kernel(
real_subtile_idx = subtile_idx // 2
else:
real_subtile_idx = subtile_idx
# Position in processing order. The early accumulator release below counts
# processed subtiles, so it must not use the reversed index.
iter_subtile_idx = real_subtile_idx
if cutlass.const_expr(self.overlapping_accum):
if reverse_subtile:
real_subtile_idx = (
Expand Down Expand Up @@ -2528,7 +2531,7 @@ def kernel(
# Async arrive accumulator buffer empty earlier when overlapping_accum is enabled
#
if cutlass.const_expr(self.overlapping_accum):
if real_subtile_idx == self.iter_acc_early_release_in_epilogue:
if iter_subtile_idx == self.iter_acc_early_release_in_epilogue:
# Fence for TMEM load
cute.arch.fence_view_async_tmem_load()
acc_pipeline.consumer_release(acc_consumer_state)
Expand Down
Loading