Skip to content

Add NaN handling in softmax pattern in SDPA fusion #2593

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 1 commit into from
Oct 1, 2025
Merged

Add NaN handling in softmax pattern in SDPA fusion #2593

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
3 changes: 3 additions & 0 deletions onnxscript/rewriter/ort_fusions/sdpa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,6 +88,9 @@ def pattern(
)

attn_weight = op.Softmax(attn_score, axis=-1)
is_nan = op.IsNaN(attn_weight)
adj_attn_weight = op.Where(is_nan, 0.0, attn_weight)
attn_weight = pattern.OrValue([adj_attn_weight, attn_weight])
attn_output = op.MatMul(attn_weight, value)
return attn_output

Expand Down
85 changes: 68 additions & 17 deletions onnxscript/rewriter/ort_fusions/sdpa_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,10 @@ def _unmasked_pre_div_sdpa_script(query, key, value):
scaled_key = op.Div(key_transposed, divisor)
attn_score = op.MatMul(scaled_query, scaled_key)
attn_weight = op.Softmax(attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -56,7 +59,10 @@ def _unmasked_pre_mul_sdpa_script(query, key, value):
scaled_key = op.Mul(key_transposed, multiplier)
attn_score = op.MatMul(scaled_query, scaled_key)
attn_weight = op.Softmax(attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -67,7 +73,10 @@ def _unmasked_post_div_sdpa_script(query, key, value):
attn_score = op.MatMul(query, key_transposed)
scaled_attn_score = op.Div(attn_score, divisor)
attn_weight = op.Softmax(scaled_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -78,7 +87,10 @@ def _unmasked_post_mul_sdpa_script(query, key, value):
attn_score = op.MatMul(query, key_transposed)
scaled_attn_score = op.Mul(attn_score, multiplier)
attn_weight = op.Softmax(scaled_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -90,7 +102,10 @@ def _custom_scale_pre_div_sdpa_script(query, key, value):
scaled_key = op.Div(key_transposed, divisor)
attn_score = op.MatMul(scaled_query, scaled_key)
attn_weight = op.Softmax(attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -102,7 +117,10 @@ def _custom_scale_pre_mul_sdpa_script(query, key, value):
scaled_key = op.Mul(key_transposed, multiplier)
attn_score = op.MatMul(scaled_query, scaled_key)
attn_weight = op.Softmax(attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -115,7 +133,10 @@ def _custom_multi_scale_pre_mul_sdpa_script(query, key, value):
scaled_key = op.Mul(key_transposed, multiplier_k)
attn_score = op.MatMul(scaled_query, scaled_key)
attn_weight = op.Softmax(attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -126,7 +147,10 @@ def _custom_scale_post_div_sdpa_script(query, key, value):
attn_score = op.MatMul(query, key_transposed)
scaled_attn_score = op.Div(attn_score, divisor)
attn_weight = op.Softmax(scaled_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -137,7 +161,10 @@ def _custom_scale_post_mul_sdpa_script(query, key, value):
attn_score = op.MatMul(query, key_transposed)
scaled_attn_score = op.Mul(attn_score, multiplier)
attn_weight = op.Softmax(scaled_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -150,7 +177,10 @@ def _masked_pre_div_sdpa_script(query, key, value, mask):
attn_score = op.MatMul(scaled_query, scaled_key)
masked_attn_score = op.Add(attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -163,7 +193,10 @@ def _masked_pre_mul_sdpa_script(query, key, value, mask):
attn_score = op.MatMul(scaled_query, scaled_key)
masked_attn_score = op.Add(attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -175,7 +208,10 @@ def _masked_post_div_sdpa_script(query, key, value, mask):
scaled_attn_score = op.Div(attn_score, divisor)
masked_attn_score = op.Add(scaled_attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -187,7 +223,10 @@ def _masked_post_mul_sdpa_script(query, key, value, mask):
scaled_attn_score = op.Mul(attn_score, multiplier)
masked_attn_score = op.Add(scaled_attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -200,7 +239,10 @@ def _masked_custom_scale_pre_div_sdpa_script(query, key, value, mask):
attn_score = op.MatMul(scaled_query, scaled_key)
masked_attn_score = op.Add(attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -213,7 +255,10 @@ def _masked_custom_scale_pre_mul_sdpa_script(query, key, value, mask):
attn_score = op.MatMul(scaled_query, scaled_key)
masked_attn_score = op.Add(attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -225,7 +270,10 @@ def _masked_custom_scale_post_div_sdpa_script(query, key, value, mask):
scaled_attn_score = op.Div(attn_score, divisor)
masked_attn_score = op.Add(scaled_attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand All @@ -237,7 +285,10 @@ def _masked_custom_scale_post_mul_sdpa_script(query, key, value, mask):
scaled_attn_score = op.Mul(attn_score, multiplier)
masked_attn_score = op.Add(scaled_attn_score, mask)
attn_weight = op.Softmax(masked_attn_score, axis=-1)
attn_output = op.MatMul(attn_weight, value)
is_nan = op.IsNaN(attn_weight)
zero = op.Constant(value_float=0.0)
adj_attn_weight = op.Where(is_nan, zero, attn_weight)
attn_output = op.MatMul(adj_attn_weight, value)
return attn_output


Expand Down
Loading