Skip to content

Update optimizer opset version checks for latest ONNX opset 26 #28966

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
yuslepukhin merged 8 commits into from
Jun 12, 2026
Merged

Update optimizer opset version checks for latest ONNX opset 26 #28966

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
6a12338
Update optimizer opset version checks for latest ONNX opset 25
yuslepukhin Jun 9, 2026
62264db
Update optimizer opset version checks for ONNX opset 26 and add curre…
yuslepukhin Jun 10, 2026
1e044d7
Address review comments: fix .at() crash risk, remove redundant check
yuslepukhin Jun 11, 2026
8ca6320
Add Shape opset 15+ start/end attribute guards to prevent incorrect f…
yuslepukhin Jun 11, 2026
033ac63
Fix Shape end-attr guard to accept INT64_MAX as full-shape default
yuslepukhin Jun 11, 2026
992c85e
Add comments explaining partial-shape guard logic
yuslepukhin Jun 11, 2026
7941263
Add scope comment to LoadModelAtCurrentOpset helper
yuslepukhin Jun 11, 2026
06ef3db
Extract IsFullShapeNode helper into graph_utils
yuslepukhin Jun 11, 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
9 changes: 9 additions & 0 deletions onnxruntime/core/graph/graph_utils.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@
#include "core/common/logging/logging.h"

#include <algorithm>
#include <limits>

Check warning on line 11 in onnxruntime/core/graph/graph_utils.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Found C++ system header after other header. Should be: graph_utils.h, c system, c++ system, other.  [build/include_order] [4]

Raw Output:
onnxruntime/core/graph/graph_utils.cc:11:  Found C++ system header after other header. Should be: graph_utils.h, c system, c++ system, other.  [build/include_order] [4]
#include <queue>
#include <string>
#include <vector>
Expand Down Expand Up @@ -411,6 +412,14 @@
return iter == attrs.end() ? nullptr : &iter->second;
}

bool IsFullShapeNode(const Node& node) {
const auto* start_attr = GetNodeAttribute(node, "start");
const auto* end_attr = GetNodeAttribute(node, "end");
// end=INT64_MAX is the runtime default meaning "all dimensions" (full shape).
return (!start_attr || start_attr->i() == 0) &&
(!end_attr || end_attr->i() == std::numeric_limits<int64_t>::max());
}

static NodeArg& GetOrCreateNodeArg(Graph& graph, const ONNX_NAMESPACE::TensorProto& new_initializer) {
ONNX_NAMESPACE::TypeProto new_type;
auto* typeproto_tensor = new_type.mutable_tensor_type();
Expand Down
5 changes: 5 additions & 0 deletions onnxruntime/core/graph/graph_utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@
#include "core/graph/onnx_protobuf.h"
#include "core/graph/graph.h"

#include <limits>

Check warning on line 11 in onnxruntime/core/graph/graph_utils.h

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Found C++ system header after other header. Should be: graph_utils.h, c system, c++ system, other.  [build/include_order] [4]

Raw Output:
onnxruntime/core/graph/graph_utils.h:11:  Found C++ system header after other header. Should be: graph_utils.h, c system, c++ system, other.  [build/include_order] [4]
#include <string>
#include <vector>

Expand All @@ -31,6 +32,10 @@
/** Returns the attribute of a Node with a given name. */
const ONNX_NAMESPACE::AttributeProto* GetNodeAttribute(const Node& node, const std::string& attr_name);

/** Checks whether a Shape node returns the full tensor shape (all dimensions).
* Returns false if start/end attributes restrict the output to a subset of dimensions. */
bool IsFullShapeNode(const Node& node);

/** Add a new initializer to 'graph'.
Checks that new_initializer does not already exist in 'graph' before adding it.
@returns The NodeArg for the new initializer.
Expand Down
58 changes: 29 additions & 29 deletions onnxruntime/core/optimizer/attention_fusion.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -349,22 +349,22 @@ static bool TryFuseMobileClipMHA(Node& qkv_matmul,

const Node* sequence_transpose = graph_utils::GetInputNode(qkv_matmul, 0);
if (sequence_transpose == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*sequence_transpose, "Transpose", {1, 13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*sequence_transpose, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain) ||
!HasExpectedPerm(*sequence_transpose, {0, 2, 1}) ||
!optimizer_utils::CheckOutputEdges(graph, *sequence_transpose, 1)) {
return false;
}

const Node* input_reshape = graph_utils::GetInputNode(*sequence_transpose, 0);
if (input_reshape == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*input_reshape, "Reshape", {5, 13, 14}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*input_reshape, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain) ||
!optimizer_utils::CheckOutputEdges(graph, *input_reshape, 1)) {
return fail("missing input Reshape before sequence transpose");
}

Node* qkv_reshape = GetOnlyChildByOutputIndex(graph, qkv_matmul, 0, "Reshape");
if (qkv_reshape == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*qkv_reshape, "Reshape", {5, 13, 14}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*qkv_reshape, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain) ||
!optimizer_utils::CheckOutputEdges(graph, *qkv_reshape, 1)) {
return fail("qkv Reshape after MatMul not matched");
}
Expand All @@ -379,9 +379,9 @@ static bool TryFuseMobileClipMHA(Node& qkv_matmul,
Node* k_squeeze = GetOnlyChildByOutputIndex(graph, *split, 1, "Squeeze");
Node* v_transpose = GetOnlyChildByOutputIndex(graph, *split, 2, "Transpose");
if (q_transpose == nullptr || k_squeeze == nullptr || v_transpose == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*q_transpose, "Transpose", {1, 13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*k_squeeze, "Squeeze", {13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*v_transpose, "Transpose", {1, 13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*q_transpose, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*k_squeeze, "Squeeze", {13, 21, 23, 24, 25}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*v_transpose, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain) ||
!HasExpectedPerm(*q_transpose, {2, 0, 3, 1, 4}) ||
!HasExpectedPerm(*v_transpose, {2, 0, 3, 1, 4}) ||
!HasExpectedAxesInput(graph, *k_squeeze, {2})) {
Expand All @@ -391,8 +391,8 @@ static bool TryFuseMobileClipMHA(Node& qkv_matmul,
Node* q_squeeze = GetOnlyChildByOutputIndex(graph, *q_transpose, 0, "Squeeze");
Node* v_squeeze = GetOnlyChildByOutputIndex(graph, *v_transpose, 0, "Squeeze");
if (q_squeeze == nullptr || v_squeeze == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*q_squeeze, "Squeeze", {13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*v_squeeze, "Squeeze", {13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*q_squeeze, "Squeeze", {13, 21, 23, 24, 25}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*v_squeeze, "Squeeze", {13, 21, 23, 24, 25}, kOnnxDomain) ||
!HasExpectedAxesInput(graph, *q_squeeze, {0}) ||
!HasExpectedAxesInput(graph, *v_squeeze, {0})) {
return fail("q/v squeeze pattern not matched");
Expand All @@ -402,7 +402,7 @@ static bool TryFuseMobileClipMHA(Node& qkv_matmul,
Node* k_transpose = GetOnlyChildByOutputIndex(graph, *k_squeeze, 0, "Transpose");
if (q_scale_mul == nullptr || k_transpose == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*q_scale_mul, "Mul", {7, 13, 14}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*k_transpose, "Transpose", {1, 13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*k_transpose, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain) ||
!HasExpectedPerm(*k_transpose, {0, 2, 3, 1})) {
return fail("q scale Mul or k Transpose(0,2,3,1) not matched");
}
Expand Down Expand Up @@ -460,15 +460,15 @@ static bool TryFuseMobileClipMHA(Node& qkv_matmul,

Node* transpose_3 = GetOnlyChildByOutputIndex(graph, *qkv_matmul_1, 0, "Transpose");
if (transpose_3 == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*transpose_3, "Transpose", {1, 13}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*transpose_3, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain) ||
!HasExpectedPerm(*transpose_3, {0, 2, 1, 3}) ||
!optimizer_utils::CheckOutputEdges(graph, *transpose_3, 1)) {
return fail("output Transpose(0,2,1,3) not matched");
}

Node* reshape_2 = GetOnlyChildByOutputIndex(graph, *transpose_3, 0, "Reshape");
if (reshape_2 == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*reshape_2, "Reshape", {5, 13, 14}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*reshape_2, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain) ||
!optimizer_utils::CheckOutputEdges(graph, *reshape_2, 1)) {
return fail("output Reshape not matched");
}
Expand Down Expand Up @@ -497,7 +497,7 @@ static bool TryFuseMobileClipMHA(Node& qkv_matmul,
if (proj_gemm == nullptr) {
proj_gemm_input_reshape = GetOnlyChildByOutputIndex(graph, *reshape_2, 0, "Reshape");
if (proj_gemm_input_reshape == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*proj_gemm_input_reshape, "Reshape", {5, 13, 14}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*proj_gemm_input_reshape, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain) ||
!optimizer_utils::CheckOutputEdges(graph, *proj_gemm_input_reshape, 1)) {
Comment thread
yuslepukhin marked this conversation as resolved.
return fail("projection MatMul/Gemm not matched");
}
Expand All @@ -511,7 +511,7 @@ static bool TryFuseMobileClipMHA(Node& qkv_matmul,

proj_gemm_output_reshape = GetOnlyChildByOutputIndex(graph, *proj_gemm, 0, "Reshape");
if (proj_gemm_output_reshape == nullptr ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*proj_gemm_output_reshape, "Reshape", {5, 13, 14}, kOnnxDomain) ||
!graph_utils::IsSupportedOptypeVersionAndDomain(*proj_gemm_output_reshape, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain) ||
!optimizer_utils::CheckOutputEdges(graph, *proj_gemm_output_reshape, 1)) {
return fail("normalized projection Gemm output Reshape not matched");
}
Expand Down Expand Up @@ -920,11 +920,11 @@ static bool FuseSubGraphQKImpl(Node& layer_norm,
}

std::vector<graph_utils::EdgeEndToMatch> q_path{
{0, 0, "Transpose", {1, 13}, kOnnxDomain},
{0, 0, "Reshape", {5, 13}, kOnnxDomain},
{0, 0, "Add", {7, 13}, kOnnxDomain},
{0, 0, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "Add", {7, 13, 14}, kOnnxDomain},
{0, 0, "MatMul", {1, 9, 13}, kOnnxDomain},
{0, 0, "LayerNormalization", {1}, kOnnxDomain}};
{0, 0, "LayerNormalization", {1, 17}, kOnnxDomain}};
if (!graph_utils::FindPath(edges[edges.size() - 1]->GetNode(), true, q_path, edges, logger)) {
DEBUG_LOG("Failed to find path for q");
return false;
Expand Down Expand Up @@ -953,9 +953,9 @@ static bool FuseSubGraphQKImpl(Node& layer_norm,
}

std::vector<graph_utils::EdgeEndToMatch> k_path{
{0, 1, "Transpose", {1, 13}, kOnnxDomain},
{0, 0, "Reshape", {5, 13}, kOnnxDomain},
{0, 0, "Add", {7, 13}, kOnnxDomain},
{0, 1, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "Add", {7, 13, 14}, kOnnxDomain},
{0, 0, "MatMul", {1, 9, 13}, kOnnxDomain},
{0, 0, "LayerNormalization", {1, 17}, kOnnxDomain}};

Expand Down Expand Up @@ -1070,8 +1070,8 @@ static bool FuseSubGraphQK(Node& layer_norm,
const logging::Logger& logger) {
// path to q
std::vector<graph_utils::EdgeEndToMatch> q_varience_path{
{0, 0, "Div", {7, 13}, kOnnxDomain},
{0, 0, "MatMul", {1, 9}, kOnnxDomain}};
{0, 0, "Div", {7, 13, 14}, kOnnxDomain},
{0, 0, "MatMul", {1, 9, 13}, kOnnxDomain}};
std::vector<const Node::EdgeEnd*> edges;
if (!graph_utils::FindPath(*(mask_nodes.add), true, q_varience_path, edges, logger)) {
DEBUG_LOG("Failed to find path for q");
Expand Down Expand Up @@ -1163,7 +1163,7 @@ static bool FuseSubGraphQKDistilBert(Node& layer_norm,
// path to q
std::vector<graph_utils::EdgeEndToMatch> q_varience_path{
{0, 2, "MatMul", {1, 9, 13}, kOnnxDomain},
{0, 0, "Div", {7, 13}, kOnnxDomain}};
{0, 0, "Div", {7, 13, 14}, kOnnxDomain}};
std::vector<const Node::EdgeEnd*> edges;
if (!graph_utils::FindPath(*(mask_nodes.where), true, q_varience_path, edges, logger)) {
DEBUG_LOG("Failed to find path for q");
Expand Down Expand Up @@ -1265,14 +1265,14 @@ bool AttentionFusion::FuseSubGraph(Node& layer_norm,
std::map<std::string, NodeArg*>& mask_int32_map,
const logging::Logger& logger) {
std::vector<graph_utils::EdgeEndToMatch> parent_path{
{0, 0, "Add", {7, 13}, kOnnxDomain},
{0, 0, "Add", {7, 13, 14}, kOnnxDomain},
{0, 0, "MatMul", {1, 9, 13}, kOnnxDomain},
{0, 0, "Reshape", {5, 13}, kOnnxDomain},
{0, 0, "Transpose", {1, 13}, kOnnxDomain},
{0, 0, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "MatMul", {1, 9, 13}, kOnnxDomain},
{0, 1, "Transpose", {1, 13}, kOnnxDomain},
{0, 0, "Reshape", {5, 13}, kOnnxDomain},
{0, 0, "Add", {7, 13}, kOnnxDomain},
{0, 1, "Transpose", {1, 13, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "Reshape", {5, 13, 14, 19, 21, 23, 24, 25}, kOnnxDomain},
{0, 0, "Add", {7, 13, 14}, kOnnxDomain},
{0, 0, "MatMul", {1, 9, 13}, kOnnxDomain},
{0, 0, "LayerNormalization", {1, 17}, kOnnxDomain}};

Expand Down
Loading
Loading