Skip to content

[DRAFT] Extend GQA fusion for Gemma3 #2639

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
gramalingam merged 2 commits into from
Oct 16, 2025
Merged

[DRAFT] Extend GQA fusion for Gemma3 #2639

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
00ffc66
Extend GQA fusion for Gemma3
gramalingam Oct 16, 2025
7c5f947
Merge branch 'main' into rama/gqafix
justinchuby Oct 16, 2025
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
39 changes: 39 additions & 0 deletions onnxscript/rewriter/ort_fusions/gqa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -166,11 +166,23 @@ def pattern(
# Transpose from (B, S, H, D/H) to (B, H, S, D/H)
query_BHSDh = op.Transpose(query_BSHDh, perm=[0, 2, 1, 3])

# Gemma variant uses normalization of query/key before rotary embedding:
query_BHSDh_normalized = op.SimplifiedLayerNormalization(
query_BHSDh, pattern.ANY_VALUE, axis=-1, _outputs=["query_BHSDh_normalized"]
)
query_BHSDh = pattern.OrValue([query_BHSDh, query_BHSDh_normalized])

# Reshape key from (B, S, Dkv) to (B, S, Hkv, D/H)
key_BSHkvDh = op.Reshape(key_BSDkv, pattern.ANY_VALUE, _outputs=["key_BSHkvDh"])
# Transpose from (B, S, Hkv, D/H) to (B, Hkv, S, D/H)
key_BHkvSDh = op.Transpose(key_BSHkvDh, perm=[0, 2, 1, 3])

# Gemma variant uses normalization of query/key before rotary embedding:
key_BHkvSDh_normalized = op.SimplifiedLayerNormalization(
key_BHkvSDh, pattern.ANY_VALUE, axis=-1, _outputs=["key_BHkvSDh_normalized"]
)
key_BHkvSDh = pattern.OrValue([key_BHkvSDh, key_BHkvSDh_normalized])

# Reshape value from (B, S, Dkv) to (B, S, Hkv, D/H)
value_BSHkvDh = op.Reshape(value_BSDkv, pattern.ANY_VALUE, _outputs=["value_BSHkvDh"])
# Transpose from (B, S, Hkv, D/H) to (B, Hkv, S, D/H)
Expand Down Expand Up @@ -316,6 +328,10 @@ def rewrite(
cos,
sin,
mask,
query_BSHDh,
key_BSHkvDh,
query_BHSDh_normalized=None,
key_BHkvSDh_normalized=None,
**_,
):
# Note that the following optimization is specific to current ORT GenAI attention-mask
Expand All @@ -335,6 +351,29 @@ def rewrite(
seqlens_k = op.Cast(seqlens_k_int64, to=ir.DataType.INT32)
max_seq_length = op.ReduceMax(seqlens_k, zero_int64_1d, keepdims=0)
total_seq_length_int32 = op.Add(max_seq_length, one_int32_0d)

if query_BHSDh_normalized is not None:
# We apply normalization without the transpose, which is fused into GQA
norm_node = query_BHSDh_normalized.producer()
norm_attrs = norm_node.attributes
norm_scale = norm_node.inputs[1]
query_BSHDh_normalized = op.SimplifiedLayerNormalization(
query_BSHDh, norm_scale, **norm_attrs
)
reshape_BSHDh_to_BSD = op.Constant(value_ints=[0, 0, -1])
query_BSD = op.Reshape(query_BSHDh_normalized, reshape_BSHDh_to_BSD)

if key_BHkvSDh_normalized is not None:
# We apply normalization without the transpose, which is fused into GQA
norm_node = key_BHkvSDh_normalized.producer()
norm_attrs = norm_node.attributes
norm_scale = norm_node.inputs[1]
key_BSHkvDh_normalized = op.SimplifiedLayerNormalization(
key_BSHkvDh, norm_scale, **norm_attrs
)
reshape_BSHkvDh_to_BSDkv = op.Constant(value_ints=[0, 0, -1])
key_BSDkv = op.Reshape(key_BSHkvDh_normalized, reshape_BSHkvDh_to_BSDkv)

return op.GroupQueryAttention(
query_BSD,
key_BSDkv,
Expand Down
Loading