Skip to content

Fix GPT-2 no-past attention fusion for transformers >= 4.27 #27449

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
tianleiwu merged 3 commits into from
Mar 5, 2026
Merged

Fix GPT-2 no-past attention fusion for transformers >= 4.27 #27449

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
2030a97
Fix GPT-2 no-past attention fusion for transformers >= 4.27
Rishi-Dave Feb 25, 2026
4b881b5
Use verify_fusion in no-past attention test per review feedback
Rishi-Dave Feb 25, 2026
604aa89
Fix line length lint in no-past attention test
Rishi-Dave Feb 25, 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
25 changes: 11 additions & 14 deletions onnxruntime/python/tools/transformers/fusion_gpt_attention_no_past.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,9 +61,6 @@ def create_attention_node(self, gemm, gemm_qkv, input, output):
self.node_name_to_graph_name[add_node.name] = self.this_graph_name

def fuse(self, normalize_node, input_name_to_nodes, output_name_to_node):
# (TODO) hasesh/tlwu: Investigate what fixes the following logic needs in order
# to fuse the Attention sub-graph. With some changes to other fusions, this stopped
# working.
return_indice = []

is_normalize_node_skiplayernorm = normalize_node.op_type == "SkipLayerNormalization"
Expand Down Expand Up @@ -187,20 +184,20 @@ def fuse(self, normalize_node, input_name_to_nodes, output_name_to_node):
qk_nodes = self.model.match_parent_path(matmul_qkv, ["Softmax", "Where", "Div", "MatMul"], [0, 0, 1, 0])
if qk_nodes is not None:
(softmax_qk, where_qk, div_qk, matmul_qk) = qk_nodes
mask_nodes = self.model.match_parent_path(
_, mask_nodes, _ = self.model.match_parent_paths(
where_qk,
[
"Cast",
"Slice",
"Slice",
"Unsqueeze",
"Sub",
"Squeeze",
"Slice",
"Shape",
"Div",
(
["Cast", "Slice", "Slice", "Unsqueeze", "Sub", "Squeeze", "Slice", "Shape", "Div"],
[0, 0, 0, 1, 0, 0, 0, 0, 0],
),
# For transformers >= 4.27, causal mask uses torch.bool instead of torch.uint8.
(
["Slice", "Slice", "Unsqueeze", "Sub", "Squeeze", "Slice", "Shape", "Div"],
[0, 0, 1, 0, 0, 0, 0, 0],
),
],
[0, 0, 0, 1, 0, 0, 0, 0, 0],
output_name_to_node,
)
if mask_nodes is None:
logger.debug("fuse_attention: failed to match mask path")
Expand Down
Loading
Loading