Skip to content

Enable sm120 support for fused attn if cuDNN is 9.18.1+ #2693

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

Draft
KshitijLakhani wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

Enable sm120 support for fused attn if cuDNN is 9.18.1+ #2693

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
998b3b8
Enable sm120 support for fused attn if cuDNN is 9.18.1+
KshitijLakhani Feb 20, 2026
dc282ea
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Feb 20, 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
5 changes: 3 additions & 2 deletions transformer_engine/pytorch/attention/dot_product_attention/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -610,6 +610,7 @@ def get_attention_backend(
qkv_layout,
)
use_fused_attention = False
# TODO: KL check if this condition is now supported or not ?
if (
device_compute_capability == (12, 0)
and (head_dim_qk > 128 or head_dim_qk % 8 != 0)
Expand Down Expand Up @@ -690,11 +691,11 @@ def _is_fa3_supported(num_heads, num_gqa_groups, head_dim_qk, head_dim_v, qkv_dt
"padding between sequences, i.e. [a, a, PAD, b, b, b, PAD, c, PAD]"
)
use_flash_attention = False
if device_compute_capability == (12, 0):
if device_compute_capability == (12, 0) and cudnn_version < (9, 18, 1):
if use_fused_attention:
logger.debug(
"Disabling FusedAttention as qkv_format = thd is"
" not supported for compute capability = sm120"
" not supported for compute capability = sm120 and cuDNN version < 9.18.1"
)
use_fused_attention = False

Expand Down