Skip to content

[Cute,Fwd] improved block sparsity #2100

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
tridao merged 9 commits into from
Jan 5, 2026
+461 −827
Merged

[Cute,Fwd] improved block sparsity #2100

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
b23f4a4
improved block sparsity computation
reubenconducts Dec 21, 2025
33f79f5
refactor blocksparsity computation for tvm-ffi
reubenconducts Dec 23, 2025
9b3ae2d
refactor mask mod definitions and tests
reubenconducts Dec 23, 2025
7f81eae
refactor of block sparsity and mask mod application; eventually allow…
reubenconducts Dec 26, 2025
212d6ef
remove fastdivmods from compute block sparsity
reubenconducts Dec 30, 2025
0074730
remove unnecessary imports
reubenconducts Dec 30, 2025
2651f75
revert to 1-phase block sparsity computation
reubenconducts Dec 30, 2025
e79adcc
update bwd kernels to use new AttentionMaskCls api
reubenconducts Jan 5, 2026
c0677da
fix linter error
reubenconducts Jan 5, 2026
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
3 changes: 3 additions & 0 deletions flash_attn/cute/block_sparse_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -274,6 +274,7 @@ def consume_block_sparse_loads(
batch_idx,
head_idx,
m_block,
seqlen,
kv_consumer_state,
mma_pv_fn,
mma_one_n_block,
Expand Down Expand Up @@ -380,6 +381,7 @@ def consume_block_sparse_loads(
kv_consumer_state = process_first_half_block(
n_block=mask_n_block,
kv_consumer_state=kv_consumer_state,
seqlen=seqlen,
mask_fn=partial(
mask_fn,
mask_mod=mask_mod,
Expand All @@ -405,6 +407,7 @@ def consume_block_sparse_loads(
kv_consumer_state = process_first_half_block(
n_block=full_n_block,
kv_consumer_state=kv_consumer_state,
seqlen=seqlen,
mask_fn=partial(mask_fn, mask_mod=None, mask_seqlen=True),
score_mod_fn=score_mod_fn,
is_first_block=True,
Expand Down
Loading
Loading