Skip to content

Fill CUDA EP opset gap for Shape operator (23→25) #27734

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 4 commits into from
Mar 24, 2026
+21 −6
Merged

Fill CUDA EP opset gap for Shape operator (23→25) #27734

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
4404985
Update Shape CUDA operator to fill opset gap (23→25)
Copilot Mar 18, 2026
38e0bc0
Merge remote-tracking branch 'origin/main' into copilot/update-shape-…
Copilot Mar 18, 2026
e69559d
Update docs/OperatorKernels.md for Shape CUDA opset 25 registration
Copilot Mar 18, 2026
6f84b5a
Merge remote-tracking branch 'origin/main' into copilot/update-shape-…
tianleiwu Mar 24, 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
3 changes: 2 additions & 1 deletion docs/OperatorKernels.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -941,7 +941,8 @@ Do not modify directly.*
|SequenceErase|*in* input_sequence:**S**<br> *in* position:**I**<br> *out* output_sequence:**S**|11+|**I** = tensor(int32), tensor(int64)<br/> **S** = seq(tensor(bfloat16)), seq(tensor(bool)), seq(tensor(double)), seq(tensor(float)), seq(tensor(float16)), seq(tensor(int16)), seq(tensor(int32)), seq(tensor(int64)), seq(tensor(int8)), seq(tensor(uint16)), seq(tensor(uint32)), seq(tensor(uint64)), seq(tensor(uint8))|
|SequenceInsert|*in* input_sequence:**S**<br> *in* tensor:**T**<br> *in* position:**I**<br> *out* output_sequence:**S**|11+|**I** = tensor(int32), tensor(int64)<br/> **S** = seq(tensor(bfloat16)), seq(tensor(bool)), seq(tensor(double)), seq(tensor(float)), seq(tensor(float16)), seq(tensor(int16)), seq(tensor(int32)), seq(tensor(int64)), seq(tensor(int8)), seq(tensor(uint16)), seq(tensor(uint32)), seq(tensor(uint64)), seq(tensor(uint8))|
|SequenceLength|*in* input_sequence:**S**<br> *out* length:**I**|11+|**I** = tensor(int64)<br/> **S** = seq(tensor(bfloat16)), seq(tensor(bool)), seq(tensor(double)), seq(tensor(float)), seq(tensor(float16)), seq(tensor(int16)), seq(tensor(int32)), seq(tensor(int64)), seq(tensor(int8)), seq(tensor(uint16)), seq(tensor(uint32)), seq(tensor(uint64)), seq(tensor(uint8))|
|Shape|*in* data:**T**<br> *out* shape:**T1**|23+|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T1** = tensor(int64)|
|Shape|*in* data:**T**<br> *out* shape:**T1**|25+|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T1** = tensor(int64)|
|||[23, 24]|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T1** = tensor(int64)|
|||[21, 22]|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T1** = tensor(int64)|
|||[19, 20]|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T1** = tensor(int64)|
|||[15, 18]|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T1** = tensor(int64)|
Expand Down
10 changes: 6 additions & 4 deletions onnxruntime/core/providers/cuda/cuda_execution_provider.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1578,7 +1578,6 @@ class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kO
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Float8E4M3FN, Cast);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Float8E5M2, Cast);
#endif
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Shape);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Reshape);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Transpose);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Squeeze);
Expand All @@ -1592,6 +1591,7 @@ class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDom
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, If);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Loop);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Scan);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Shape);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Size);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Flatten);

Expand Down Expand Up @@ -1665,7 +1665,6 @@ class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kO
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 23, double, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 23, MLFloat16, Pad);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 23, bool, Pad);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Shape);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Reshape);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Transpose);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 23, Squeeze);
Expand All @@ -1675,6 +1674,7 @@ class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDom
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, If);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Loop);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Scan);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Shape);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Size);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Flatten);

Expand Down Expand Up @@ -1706,6 +1706,7 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, I
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, If);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Loop);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Scan);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Shape);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Size);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Flatten);

Expand Down Expand Up @@ -2736,7 +2737,6 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Float8E4M3FN, Cast)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Float8E5M2, Cast)>,
#endif
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Shape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Reshape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Transpose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Squeeze)>,
Expand All @@ -2746,6 +2746,7 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, If)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Loop)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Scan)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Shape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Size)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 21, 22, Flatten)>,

Expand Down Expand Up @@ -2822,7 +2823,6 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
#if !defined(DISABLE_FLOAT4_TYPES)
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Float4E2M1x2, Cast)>,
#endif
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Shape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Reshape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, Transpose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 23, Squeeze)>,
Expand All @@ -2832,6 +2832,7 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, If)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Loop)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Scan)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Shape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Size)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 23, 24, Flatten)>,

Expand Down Expand Up @@ -2859,6 +2860,7 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, If)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Loop)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Scan)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Shape)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Size)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Flatten)>,
#endif
Expand Down
14 changes: 13 additions & 1 deletion onnxruntime/core/providers/cuda/tensor/shape_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,10 +68,22 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
.TypeConstraint("T1", DataTypeImpl::GetTensorType<int64_t>()),
Shape);

ONNX_OPERATOR_VERSIONED_KERNEL_EX(
Shape,
kOnnxDomain,
23, 24,
kCudaExecutionProvider,
(*KernelDefBuilder::Create())
// properly force CPU/GPU synch inside the kernel
.OutputMemoryType(OrtMemTypeCPUInput, 0)
.TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypesIRv9())
.TypeConstraint("T1", DataTypeImpl::GetTensorType<int64_t>()),
Shape);

ONNX_OPERATOR_KERNEL_EX(
Shape,
kOnnxDomain,
23,
25,
kCudaExecutionProvider,
(*KernelDefBuilder::Create())
// properly force CPU/GPU synch inside the kernel
Expand Down
Loading