Skip to content

[release/2.7] Fix for flex attention tuning #2589

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
Sep 2, 2025
Merged

[release/2.7] Fix for flex attention tuning #2589

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1d39f46
Fix for flex attention tuning
jataylo Sep 2, 2025
00b714b
Update flex_attention.py
jataylo Sep 2, 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
14 changes: 0 additions & 14 deletions torch/_inductor/kernel/flex_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2476,20 +2476,6 @@ def flex_attention_backward(*args, **kwargs):
SPARSE_KV_BLOCK_SIZE = V.graph.sizevars.evaluate_static_shape(SPARSE_KV_BLOCK_SIZE)

choices: list[Any] = []
configs: list[tuple[int, int, int, int]] = []
configs.append(_get_default_config_bwd(query))
if config.max_autotune:
num_stages_list = [1, 3, 4, 5] if torch.version.hip is None else [1]
configs.extend(
[
(BLOCK1, BLOCK2, w, s)
for BLOCK1 in [32, 64]
for BLOCK2 in [32, 64, 128]
for w in ([4, 8] if BLOCK1 >= 128 or BLOCK2 >= 128 else [4])
for s in num_stages_list
if BLOCK2 % BLOCK1 == 0
]
)

dtype = query.get_dtype()
head_dim = V.graph.sizevars.evaluate_static_shape(query.get_size()[-1])
Expand Down