Skip to content

Fix d=256 SDPA #326

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
Jun 26, 2026
Merged

Fix d=256 SDPA #326

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
10 changes: 5 additions & 5 deletions include/cudnn_frontend/node/scaled_dot_product_flash_attention.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1722,6 +1722,11 @@ class CompositeSDPABackwardNode : public NodeCRTP<CompositeSDPABackwardNode> {
sub_nodes.emplace_back(node_);
}

// last_output = last_output - stats
last_output = pointwise(last_output,
attributes.inputs[input_names::Stats],
Pointwise_attributes().set_name("sub_s_m").set_mode(PointwiseMode_t::SUB));

// (optional) Apply padding mask
if (attributes.padding_mask) {
auto graph_ = std::make_shared<Graph>();
Expand All @@ -1734,11 +1739,6 @@ class CompositeSDPABackwardNode : public NodeCRTP<CompositeSDPABackwardNode> {
sub_nodes.emplace_back(node_);
}

// last_output = last_output - stats
last_output = pointwise(last_output,
attributes.inputs[input_names::Stats],
Pointwise_attributes().set_name("sub_s_m").set_mode(PointwiseMode_t::SUB));

// Explicitly put the padding value again after the stats have been loaded
if (attributes.padding_mask && detail::get_backend_version() >= 90000 &&
detail::get_backend_version() < 91000) {
Expand Down
2 changes: 1 addition & 1 deletion test/python/test_mhas_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -478,7 +478,7 @@ def test_sdpa_random_bwd_ragged_L0(env_info, test_no, request, cudnn_handle):
with RandomizationContext(
batches=RandomBatchSize(min=8, max=16),
s_q_s_kv = RandomSequenceLength(s_q_min=1, s_q_max=4096, s_kv_min=1, s_kv_max=4096, s_q_distribution={"s_q=1":0, "s_q=s_kv":5, "s_q=random":10}),
d_qk_d_v=RandomHiddenDimSize(d_qk_min=1, d_qk_max=192, d_v_min=1, d_v_max=128, head_dim_distribution={"d_qk=d_v":5, "d_qk=random":1}, with_high_probability=[(64,64), (128,128), (192,128)]),
d_qk_d_v=RandomHiddenDimSize(d_qk_min=1, d_qk_max=192, d_v_min=1, d_v_max=128, head_dim_distribution={"d_qk=d_v":5, "d_qk=random":1}, with_high_probability=[(64,64), (128,128), (192,128), (256,256)]),
head_count=RandomHeadGenerator(min=1, max=8, head_group_options=(1, 4, 1)),
data_type=RandomChoice({torch.float16 : 1, torch.bfloat16 : 2}),
with_sliding_mask=SlidingWindowMaskGenerator(causal=10, left_window_only=5, right_window_only=5, band_around_diag=10, no_mask=10),
Expand Down