Skip to content

Enable head dimension 128 test cases for backward and forward equivalence #137

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
merged 2 commits into from
Aug 29, 2025
Merged

Enable head dimension 128 test cases for backward and forward equivalence #137

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a347663
Enables head dimension 128 test cases in backward equivalence
LoserCheems Aug 29, 2025
1df70eb
Enables head dimension 128 test cases in CUDA forward equivalence
LoserCheems Aug 29, 2025
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
27 changes: 13 additions & 14 deletions benchmarks/backward_equivalence.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -598,23 +598,22 @@ def test_cuda_backward_equivalence(accuracy_threshold=0.95):
(1, 2, 1, 4096, 4096, 96, True),
(1, 2, 1, 4096, 4096, 96, False),

# Not support head_dim => 128 in sm89 yet
# Because fwd uses splitkv branch, this branch does not support head_dim>=128 for now
# Head dim 128
# (1, 2, 1, 128, 128, 128, True),
# (1, 2, 1, 128, 128, 128, True),
# (1, 2, 1, 256, 256, 128, True),
# (1, 2, 1, 256, 256, 128, False),
# (1, 2, 1, 512, 512, 128, True),
# (1, 2, 1, 512, 512, 128, False),
# (1, 2, 1, 1024, 1024, 128, True),
# (1, 2, 1, 1024, 1024, 128, False),
# (1, 2, 1, 2048, 2048, 128, True),
# (1, 2, 1, 2048, 2048, 128, False),
# (1, 2, 1, 4096, 4096, 128, True),
# (1, 2, 1, 4096, 4096, 128, False),
(1, 2, 1, 128, 128, 128, True),
(1, 2, 1, 128, 128, 128, False),
Comment on lines +602 to +603
Copy link

Copilot AI Aug 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Line 603 should have False as the last parameter, but the original commented line 603 had True. This creates an inconsistency where there are two test cases with True and none with False for the (1, 2, 1, 128, 128, 128) configuration.

Copilot uses AI. Check for mistakes.
(1, 2, 1, 256, 256, 128, True),
(1, 2, 1, 256, 256, 128, False),
(1, 2, 1, 512, 512, 128, True),
(1, 2, 1, 512, 512, 128, False),
(1, 2, 1, 1024, 1024, 128, True),
(1, 2, 1, 1024, 1024, 128, False),
(1, 2, 1, 2048, 2048, 128, True),
(1, 2, 1, 2048, 2048, 128, False),
(1, 2, 1, 4096, 4096, 128, True),
(1, 2, 1, 4096, 4096, 128, False),

# Head dim 256
# Because fwd uses splitkv branch, this branch does not support head_dim=256 for now
# For head_dim=256, besides the reason of splitkv branch, bwd itself does not support it, not enough shared memory
# (1, 2, 1, 128, 128, 256, True),
# (1, 2, 1, 128, 128, 256, False),
Expand Down
28 changes: 14 additions & 14 deletions benchmarks/forward_equivalence.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -561,22 +561,22 @@ def test_cuda_forward_equivalence(accuracy_threshold=0.95):
(1, 2, 1, 4096, 4096, 96, True),
(1, 2, 1, 4096, 4096, 96, False),

# Not support head_dim >= 128 in sm89 yet
# Because fwd uses splitkv branch by default, and shared memory is not enough for sm89
# Head dim 128
# (1, 2, 1, 128, 128, 128, True),
# (1, 2, 1, 128, 128, 128, False),
# (1, 2, 1, 256, 256, 128, True),
# (1, 2, 1, 256, 256, 128, False),
# (1, 2, 1, 512, 512, 128, True),
# (1, 2, 1, 512, 512, 128, False),
# (1, 2, 1, 1024, 1024, 128, True),
# (1, 2, 1, 1024, 1024, 128, False),
# (1, 2, 1, 2048, 2048, 128, True),
# (1, 2, 1, 2048, 2048, 128, False),
# (1, 2, 1, 4096, 4096, 128, True),
# (1, 2, 1, 4096, 4096, 128, False),
(1, 2, 1, 128, 128, 128, True),
(1, 2, 1, 128, 128, 128, False),
(1, 2, 1, 256, 256, 128, True),
(1, 2, 1, 256, 256, 128, False),
(1, 2, 1, 512, 512, 128, True),
(1, 2, 1, 512, 512, 128, False),
(1, 2, 1, 1024, 1024, 128, True),
(1, 2, 1, 1024, 1024, 128, False),
(1, 2, 1, 2048, 2048, 128, True),
(1, 2, 1, 2048, 2048, 128, False),
(1, 2, 1, 4096, 4096, 128, True),
(1, 2, 1, 4096, 4096, 128, False),

# Not support head_dim = 256 in sm89 yet
# Because fwd uses splitkv branch by default, and shared memory is not enough for sm89
# Head dim 256
# (1, 2, 1, 128, 128, 256, True),
# (1, 2, 1, 128, 128, 256, False),
Expand Down