diff --git a/onnxruntime/core/providers/cpu/nn/flatten.h b/onnxruntime/core/providers/cpu/nn/flatten.h
index 5892c8ce59..04abbbeb7a 100644
--- a/onnxruntime/core/providers/cpu/nn/flatten.h
+++ b/onnxruntime/core/providers/cpu/nn/flatten.h
@@ -23,13 +23,12 @@ class Flatten final : public OpKernel {
 
     const TensorShape& X_shape = X->Shape();
     auto axis = axis_;
-    
+
     // Valid axis range is [-rank, rank] instead of [-rank, rank-1], add additional check to only handle neg axis case.
-    if (axis < 0)
-    {
+    if (axis < 0) {
       axis = HandleNegativeAxis(axis, X_shape.NumDimensions());  // handle negative and enforce axis is valid
     }
-    
+
     ORT_ENFORCE(gsl::narrow_cast<int64_t>(X_shape.NumDimensions()) >= axis, "The rank of input tensor must be >= axis");
 
     Tensor* Y = context->Output(0, TensorShape({X_shape.SizeToDimension(axis), X_shape.SizeFromDimension(axis)}));
diff --git a/onnxruntime/core/providers/cpu/tensor/compress.cc b/onnxruntime/core/providers/cpu/tensor/compress.cc
index bf5502fda4..3007feef92 100644
--- a/onnxruntime/core/providers/cpu/tensor/compress.cc
+++ b/onnxruntime/core/providers/cpu/tensor/compress.cc
@@ -11,16 +11,18 @@ ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Compress,
     9,
     10,
-    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::AllTensorTypes())
-                      .TypeConstraint("T1", DataTypeImpl::GetTensorType<bool>()),
+    KernelDefBuilder()
+        .TypeConstraint("T", DataTypeImpl::AllTensorTypes())
+        .TypeConstraint("T1", DataTypeImpl::GetTensorType<bool>()),
     Compress);
 
 // Opset 11 starts to support Neg Axis.
 ONNX_CPU_OPERATOR_KERNEL(
     Compress,
     11,
-    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::AllTensorTypes())
-                      .TypeConstraint("T1", DataTypeImpl::GetTensorType<bool>()),
+    KernelDefBuilder()
+        .TypeConstraint("T", DataTypeImpl::AllTensorTypes())
+        .TypeConstraint("T1", DataTypeImpl::GetTensorType<bool>()),
     Compress);
 
 Status Compress::Compute(OpKernelContext* ctx) const {
@@ -30,7 +32,6 @@ Status Compress::Compute(OpKernelContext* ctx) const {
   int64_t axis = axis_;
   if (has_axis_) {
     axis = HandleNegativeAxis(axis, rank);  // handle negative and enforce axis is valid
-    ORT_ENFORCE(axis < static_cast<int64_t>(rank), "axis greater than input data dimension!");
   }
 
   const auto* condition = ctx->Input<Tensor>(1);
@@ -95,7 +96,7 @@ Status Compress::Compute(OpKernelContext* ctx) const {
         if (is_string_type) {
           for (int idxItem = 0; idxItem < axes_right_stride; ++idxItem) {
             reinterpret_cast<std::string*>(output_data)[output_index + idxItem] =
-              reinterpret_cast<const std::string*>(input_data)[i * axes_included_right_stride + j * axes_right_stride + idxItem];
+                reinterpret_cast<const std::string*>(input_data)[i * axes_included_right_stride + j * axes_right_stride + idxItem];
           }
           output_index += axes_right_stride;
         } else {
@@ -118,7 +119,7 @@ Status Compress::Compute(OpKernelContext* ctx) const {
     }
   }
 
-return Status::OK();
+  return Status::OK();
 }
 
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/cuda/cuda_execution_provider.cc b/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
index 19e0c5e148..07dc8eab41 100644
--- a/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
+++ b/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
@@ -210,13 +210,13 @@ namespace cuda {
 // opset 1 to 9
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MemcpyFromHost);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MemcpyToHost);
-class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 4, Concat);
-class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Unsqueeze);
+class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 4, 10, Concat);
+class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, Unsqueeze);
 class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 8, Flatten);
-class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Squeeze);
+class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, Squeeze);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Identity);
 class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, Dropout);
-class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Gather);
+class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, Gather);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 8, float, Gemm);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 8, double, Gemm);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 8, MLFloat16, Gemm);
@@ -259,9 +259,9 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, Softplus);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, Softplus);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, Softplus);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, Softmax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, Softmax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, Softmax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, Softmax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, Softmax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, Softmax);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, float, Pow);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, double, Pow);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, MLFloat16, Pow);
@@ -410,49 +410,49 @@ class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kO
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, GlobalMaxPool);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, GlobalMaxPool);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, GlobalMaxPool);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ArgMax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ArgMax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ArgMax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ArgMin);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ArgMin);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ArgMin);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceL1);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceL1);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceL1);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceL1);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceL2);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceL2);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceL2);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceL2);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceMax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceMax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceMax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceMax);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceMean);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceMean);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceMean);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceMean);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceMin);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceMin);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceMin);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceMin);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceProd);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceProd);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceProd);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceProd);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceSum);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceSum);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceSum);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceSum);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceLogSum);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceLogSum);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceLogSum);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceSumSquare);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceSumSquare);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceSumSquare);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceLogSumExp);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceLogSumExp);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceLogSumExp);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ArgMax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ArgMax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ArgMax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ArgMin);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ArgMin);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ArgMin);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceL1);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceL1);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceL1);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceL1);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceL2);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceL2);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceL2);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceL2);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceMax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceMax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceMax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceMax);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceMean);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceMean);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceMean);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceMean);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceMin);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceMin);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceMin);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceMin);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceProd);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceProd);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceProd);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceProd);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceSum);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceSum);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceSum);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceSum);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceLogSum);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceLogSum);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceLogSum);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceSumSquare);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceSumSquare);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceSumSquare);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceLogSumExp);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceLogSumExp);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceLogSumExp);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 8, float, Cast);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 8, double, Cast);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 8, MLFloat16, Cast);
@@ -502,8 +502,8 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 9, int32_t, Slice);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 9, int64_t, Slice);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 9, float, Slice);
-class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, Compress);
-class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, Flatten);
+class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, 10, Compress);
+class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, 10, Flatten);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, float, Upsample);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, double, Upsample);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, MLFloat16, Upsample);
@@ -562,36 +562,91 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, ReverseSequence);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, RoiAlign);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, RoiAlign);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, int32_t, Slice);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, int64_t, Slice);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, Slice);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, int32_t, Slice);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, int64_t, Slice);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, float, Slice);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, ThresholdedRelu);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, ThresholdedRelu);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, MLFloat16, ThresholdedRelu);
 class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, TopK);
 
 // opset 11
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ArgMax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ArgMax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ArgMax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ArgMin);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ArgMin);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ArgMin);
+class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Compress);
+class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Concat);
+class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Flatten);
+class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Gather);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, GatherElements);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Gemm);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, Gemm);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, Gemm);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, NonMaxSuppression);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Range);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceL1);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceL1);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceL1);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceL1);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceL2);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceL2);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceL2);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceL2);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceLogSum);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceLogSum);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceLogSum);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceLogSumExp);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceLogSumExp);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceLogSumExp);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceMax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceMax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceMax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceMax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceMean);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceMean);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceMean);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceMean);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceMin);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceMin);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceMin);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceMin);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceProd);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceProd);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceProd);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceProd);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceSum);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceSum);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceSum);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceSum);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceSumSquare);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceSumSquare);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceSumSquare);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, ScatterElements);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, Slice);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int64_t, Slice);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Slice);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Softmax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, Softmax);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, Softmax);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Split);
+class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Squeeze);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, TopK);
+class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Unsqueeze);
 
 static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
   static const BuildKernelCreateInfoFn function_table[] = {
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MemcpyFromHost)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MemcpyToHost)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 4, Concat)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Unsqueeze)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 4, 10, Concat)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, Unsqueeze)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 8, Flatten)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Squeeze)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, Squeeze)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Identity)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, Dropout)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, Gather)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, Gather)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 8, float, Gemm)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 8, double, Gemm)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 8, MLFloat16, Gemm)>,
@@ -635,9 +690,9 @@ static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, Softplus)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, Softplus)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, Softplus)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, Softmax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, Softmax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, Softmax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, Softmax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, Softmax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, Softmax)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, float, Pow)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, double, Pow)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, MLFloat16, Pow)>,
@@ -785,49 +840,49 @@ static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, GlobalMaxPool)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, GlobalMaxPool)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, GlobalMaxPool)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ArgMax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ArgMax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ArgMax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ArgMin)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ArgMin)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ArgMin)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceL1)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceL1)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceL1)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceL1)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceL2)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceL2)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceL2)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceL2)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceMax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceMax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceMax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceMax)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceMean)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceMean)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceMean)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceMean)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceMin)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceMin)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceMin)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceMin)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceProd)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceProd)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceProd)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceProd)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceSum)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceSum)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceSum)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, ReduceSum)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceLogSum)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceLogSum)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceLogSum)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceSumSquare)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceSumSquare)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceSumSquare)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ReduceLogSumExp)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ReduceLogSumExp)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ReduceLogSumExp)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ArgMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ArgMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ArgMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ArgMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ArgMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ArgMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, int32_t, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceLogSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceLogSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceLogSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceSumSquare)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceSumSquare)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceSumSquare)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ReduceLogSumExp)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ReduceLogSumExp)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ReduceLogSumExp)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 8, float, Cast)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 8, double, Cast)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 8, MLFloat16, Cast)>,
@@ -877,8 +932,8 @@ static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 9, int32_t, Slice)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 9, int64_t, Slice)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 9, float, Slice)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, Compress)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, Flatten)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, 10, Compress)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, 10, Flatten)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, float, Upsample)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, double, Upsample)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, MLFloat16, Upsample)>,
@@ -937,24 +992,79 @@ static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, ReverseSequence)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, RoiAlign)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, RoiAlign)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, int32_t, Slice)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, int64_t, Slice)>,
-      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, Slice)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, int32_t, Slice)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, int64_t, Slice)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, float, Slice)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, ThresholdedRelu)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, ThresholdedRelu)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, MLFloat16, ThresholdedRelu)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, TopK)>,
 
       // opset 11
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ArgMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ArgMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ArgMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ArgMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ArgMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ArgMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Compress)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Concat)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Flatten)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Gather)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, GatherElements)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, Gemm)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Gemm)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, Gemm)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, NonMaxSuppression)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Range)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceL1)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceL2)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceLogSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceLogSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceLogSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceLogSumExp)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceLogSumExp)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceLogSumExp)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceMax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceMean)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceMin)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceProd)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, ReduceSum)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ReduceSumSquare)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ReduceSumSquare)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ReduceSumSquare)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, ScatterElements)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int32_t, Slice)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, int64_t, Slice)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Slice)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Softmax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, Softmax)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, Softmax)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Split)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Squeeze)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, TopK)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Unsqueeze)>,
   };
 
   for (auto& function_table_entry : function_table) {
@@ -1139,15 +1249,18 @@ CUDAExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph,
             // or the input declared as CPU memory and is from CPU
             // in that case we should still keep the node on CUDA
             bool initializer_input = graph.GetInitializedTensor(def.Name(), initializer);
-            if ((!initializer_input && !defs_outside_cuda.count(&def)) ||
-                (defs_outside_cuda.count(&def) && cuda_kernel_def->kernel_def->IsInputOnCpu(index)))
+            bool input_is_on_cpu = defs_outside_cuda.count(&def) > 0;
+            if ((!initializer_input && !input_is_on_cpu) ||
+                (input_is_on_cpu && cuda_kernel_def->kernel_def->IsInputOnCpu(index)))
               should_force_outside = false;
+
             if (!initializer_input) {
               all_input_are_initializer = false;
             }
             return Status::OK();
           });
-      // If all the inputs are initialier, we shouldn't force it to CPU
+
+      // If all the inputs are initializers, we shouldn't force it to CPU
       if (should_force_outside && !all_input_are_initializer) {
         force_outside = true;
       }
diff --git a/onnxruntime/core/providers/cuda/math/softmax.cc b/onnxruntime/core/providers/cuda/math/softmax.cc
index ba7d437c2e..6f0b648739 100644
--- a/onnxruntime/core/providers/cuda/math/softmax.cc
+++ b/onnxruntime/core/providers/cuda/math/softmax.cc
@@ -8,11 +8,20 @@
 namespace onnxruntime {
 namespace cuda {
 
+// opset 11 added explicit support for negative axis. implementation already supported it.
 #define REGISTER_KERNEL_TYPED(T)                                                \
+  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(                                      \
+      Softmax,                                                                  \
+      kOnnxDomain,                                                              \
+      1, 10,                                                                    \
+      T,                                                                        \
+      kCudaExecutionProvider,                                                   \
+      KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
+      Softmax<T>);                                                              \
   ONNX_OPERATOR_TYPED_KERNEL_EX(                                                \
       Softmax,                                                                  \
       kOnnxDomain,                                                              \
-      1,                                                                        \
+      11,                                                                       \
       T,                                                                        \
       kCudaExecutionProvider,                                                   \
       KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
@@ -25,7 +34,7 @@ Status Softmax<T>::ComputeInternal(OpKernelContext* ctx) const {
   const TensorShape& input_shape{X.Shape()};
 
   Tensor* Y = ctx->Output(0, input_shape);
-  // special case when there is a dim value of 0 in the shape.  
+  // special case when there is a dim value of 0 in the shape.
   if (input_shape.Size() == 0)
     return Status::OK();
 
diff --git a/onnxruntime/core/providers/cuda/reduction/reduction_ops.cc b/onnxruntime/core/providers/cuda/reduction/reduction_ops.cc
index 07b72fd3ac..f461fcd223 100644
--- a/onnxruntime/core/providers/cuda/reduction/reduction_ops.cc
+++ b/onnxruntime/core/providers/cuda/reduction/reduction_ops.cc
@@ -13,11 +13,20 @@ using namespace onnxruntime::common;
 namespace onnxruntime {
 namespace cuda {
 
+// opset 11 explicitly added support for negative axis. implementation already allowed it.
 #define REGISTER_KERNEL_TYPED(name, T)                                          \
+  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(                                      \
+      name,                                                                     \
+      kOnnxDomain,                                                              \
+      1, 10,                                                                    \
+      T,                                                                        \
+      kCudaExecutionProvider,                                                   \
+      KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
+      name<T>);                                                                 \
   ONNX_OPERATOR_TYPED_KERNEL_EX(                                                \
       name,                                                                     \
       kOnnxDomain,                                                              \
-      1,                                                                        \
+      11,                                                                       \
       T,                                                                        \
       kCudaExecutionProvider,                                                   \
       KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
diff --git a/onnxruntime/core/providers/cuda/tensor/compress.cc b/onnxruntime/core/providers/cuda/tensor/compress.cc
index 9e35e3bc38..56445f8333 100644
--- a/onnxruntime/core/providers/cuda/tensor/compress.cc
+++ b/onnxruntime/core/providers/cuda/tensor/compress.cc
@@ -3,18 +3,31 @@
 
 #include "compress.h"
 #include "core/providers/cpu/tensor/utils.h"
+#include "core/providers/common.h"
 #include "compress_impl.h"
 
 namespace onnxruntime {
 namespace cuda {
 
+ONNX_OPERATOR_VERSIONED_KERNEL_EX(
+    Compress,
+    kOnnxDomain,
+    9, 10,
+    kCudaExecutionProvider,
+    KernelDefBuilder()
+        .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
+        .TypeConstraint("T1", DataTypeImpl::GetTensorType<bool>()),
+    Compress);
+
+// explicit negative axis support
 ONNX_OPERATOR_KERNEL_EX(
     Compress,
     kOnnxDomain,
-    9,
+    11,
     kCudaExecutionProvider,
-    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
-                      .TypeConstraint("T1", DataTypeImpl::GetTensorType<bool>()),
+    KernelDefBuilder()
+        .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
+        .TypeConstraint("T1", DataTypeImpl::GetTensorType<bool>()),
     Compress);
 
 Status Compress::ComputeInternal(OpKernelContext* ctx) const {
@@ -22,8 +35,9 @@ Status Compress::ComputeInternal(OpKernelContext* ctx) const {
   ORT_ENFORCE(input_tensor);
   size_t rank = input_tensor->Shape().NumDimensions();
   auto& input_dimensions = input_tensor->Shape().GetDims();
+  int64_t axis = 0;
   if (has_axis_) {
-    ORT_ENFORCE(axis_ < static_cast<int64_t>(rank), "axis greater than input data dimension!");
+    axis = HandleNegativeAxis(axis_, rank);
   }
 
   const Tensor* condition = ctx->Input<Tensor>(1);
@@ -33,19 +47,19 @@ Status Compress::ComputeInternal(OpKernelContext* ctx) const {
 
   // if has axis, we need to compress on dimension[axis], otherwise compress on the flattened input data
   int64_t input_size = input_tensor->Shape().Size();
-  int64_t compress_input_length = has_axis_ ? input_dimensions[axis_] : input_size;
+  int64_t compress_input_length = has_axis_ ? input_dimensions[axis] : input_size;
   int64_t valid_condition_length = compress_input_length < condition_length ? compress_input_length : condition_length;
 
   auto condition_cumulative_sum_buffer = GetScratchBuffer<int32_t>(valid_condition_length);
   auto condition_cumulative_sum = condition_cumulative_sum_buffer.get();
   PrefixSumImpl(reinterpret_cast<const int8_t*>(condition_data), condition_cumulative_sum, valid_condition_length);
-  
+
   int32_t positive_condition_count = 0;
   CUDA_RETURN_IF_ERROR(cudaMemcpy(&positive_condition_count, condition_cumulative_sum + valid_condition_length - 1, sizeof(int32_t), cudaMemcpyDeviceToHost));
 
   std::vector<int64_t> output_dims(input_dimensions);
   if (has_axis_) {
-    output_dims[axis_] = positive_condition_count;
+    output_dims[axis] = positive_condition_count;
   } else {
     output_dims.resize(1);
     output_dims[0] = positive_condition_count;
@@ -61,21 +75,22 @@ Status Compress::ComputeInternal(OpKernelContext* ctx) const {
 
   int64_t axis_right_stride = 1;
   if (has_axis_) {
-    for (auto i = static_cast<size_t>(axis_ + 1); i < rank; ++i) {
+    for (auto i = static_cast<size_t>(axis + 1); i < rank; ++i) {
       axis_right_stride *= input_dimensions[i];
     }
   }
 
   ORT_RETURN_IF_ERROR(CompressImpl(element_bytes,
-                                           gsl::narrow_cast<int32_t>(valid_condition_length),
-                                           gsl::narrow_cast<int32_t>(axis_right_stride),
-                                           has_axis_ ? gsl::narrow_cast<int32_t>(input_dimensions[axis_]) : gsl::narrow_cast<int32_t>(input_size),
-                                           gsl::narrow_cast<int32_t>(positive_condition_count),
-                                           condition_cumulative_sum,
-                                           condition_data,
-                                           input_tensor->DataRaw(),
-                                           output_tensor->MutableDataRaw(),
-                                           input_size));
+                                   gsl::narrow_cast<int32_t>(valid_condition_length),
+                                   gsl::narrow_cast<int32_t>(axis_right_stride),
+                                   has_axis_ ? gsl::narrow_cast<int32_t>(input_dimensions[axis])
+                                             : gsl::narrow_cast<int32_t>(input_size),
+                                   gsl::narrow_cast<int32_t>(positive_condition_count),
+                                   condition_cumulative_sum,
+                                   condition_data,
+                                   input_tensor->DataRaw(),
+                                   output_tensor->MutableDataRaw(),
+                                   input_size));
 
   return Status::OK();
 }
diff --git a/onnxruntime/core/providers/cuda/tensor/concat.cc b/onnxruntime/core/providers/cuda/tensor/concat.cc
index 66dfd27d72..4a10297957 100644
--- a/onnxruntime/core/providers/cuda/tensor/concat.cc
+++ b/onnxruntime/core/providers/cuda/tensor/concat.cc
@@ -6,14 +6,22 @@
 
 namespace onnxruntime {
 namespace cuda {
-ONNX_OPERATOR_KERNEL_EX(
-    Concat,
-    kOnnxDomain,
-    4,
-    kCudaExecutionProvider,
-    KernelDefBuilder()
-        .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
-    Concat);
+ONNX_OPERATOR_VERSIONED_KERNEL_EX(Concat,
+                                  kOnnxDomain,
+                                  4, 10,
+                                  kCudaExecutionProvider,
+                                  KernelDefBuilder()
+                                      .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
+                                  Concat);
+
+// opset 11 explicitly support negative axis
+ONNX_OPERATOR_KERNEL_EX(Concat,
+                        kOnnxDomain,
+                        11,
+                        kCudaExecutionProvider,
+                        KernelDefBuilder()
+                            .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
+                        Concat);
 
 Status Concat::ComputeInternal(OpKernelContext* ctx) const {
   auto input_count = Node().InputArgCount().front();
diff --git a/onnxruntime/core/providers/cuda/tensor/flatten.cc b/onnxruntime/core/providers/cuda/tensor/flatten.cc
index 9b09ba2f27..0db4187018 100644
--- a/onnxruntime/core/providers/cuda/tensor/flatten.cc
+++ b/onnxruntime/core/providers/cuda/tensor/flatten.cc
@@ -2,6 +2,7 @@
 // Licensed under the MIT License.
 
 #include "flatten.h"
+#include "core/providers/common.h"
 
 namespace onnxruntime {
 namespace cuda {
@@ -16,10 +17,21 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
         .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
     Flatten);
 
+ONNX_OPERATOR_VERSIONED_KERNEL_EX(
+    Flatten,
+    kOnnxDomain,
+    9, 10,
+    kCudaExecutionProvider,
+    KernelDefBuilder()
+        .Alias(0, 0)
+        .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
+    Flatten);
+
+// explicitly support negative axis
 ONNX_OPERATOR_KERNEL_EX(
     Flatten,
     kOnnxDomain,
-    9,
+    11,
     kCudaExecutionProvider,
     KernelDefBuilder()
         .Alias(0, 0)
@@ -29,14 +41,22 @@ ONNX_OPERATOR_KERNEL_EX(
 Status Flatten::ComputeInternal(OpKernelContext* ctx) const {
   const Tensor* X = ctx->Input<Tensor>(0);
   const TensorShape& X_shape = X->Shape();
-  ORT_ENFORCE(gsl::narrow_cast<int64_t>(X_shape.NumDimensions()) >= axis_, "The rank of input tensor must be >= axis");
 
-  Tensor* Y = ctx->Output(0, TensorShape({X_shape.SizeToDimension(axis_), X_shape.SizeFromDimension(axis_)}));
+  auto axis = axis_;
+  // Valid axis range is [-rank, rank] instead of [-rank, rank-1], add additional check to only handle neg axis case.
+  if (axis < 0) {
+    axis = HandleNegativeAxis(axis, X_shape.NumDimensions());  // handle negative and enforce axis is valid
+  }
+
+  ORT_ENFORCE(gsl::narrow_cast<int64_t>(X_shape.NumDimensions()) >= axis, "The rank of input tensor must be >= axis");
+
+  Tensor* Y = ctx->Output(0, TensorShape({X_shape.SizeToDimension(axis), X_shape.SizeFromDimension(axis)}));
   //If source and target pointers are not equal (non-inplace operation), we need to copy the data.
   const void* source = X->DataRaw();
   void* target = Y->MutableDataRaw();
   if (target != source) {
-    CUDA_RETURN_IF_ERROR(cudaMemcpyAsync(target, source, X_shape.Size() * X->DataType()->Size(), cudaMemcpyDeviceToDevice));
+    CUDA_RETURN_IF_ERROR(cudaMemcpyAsync(target, source, X_shape.Size() * X->DataType()->Size(),
+                                         cudaMemcpyDeviceToDevice));
   }
 
   return Status::OK();
diff --git a/onnxruntime/core/providers/cuda/tensor/gather.cc b/onnxruntime/core/providers/cuda/tensor/gather.cc
index da14721783..b4434ae6e4 100644
--- a/onnxruntime/core/providers/cuda/tensor/gather.cc
+++ b/onnxruntime/core/providers/cuda/tensor/gather.cc
@@ -8,10 +8,23 @@
 
 namespace onnxruntime {
 namespace cuda {
+ONNX_OPERATOR_VERSIONED_KERNEL_EX(
+    Gather,
+    kOnnxDomain,
+    1, 10,
+    kCudaExecutionProvider,
+    KernelDefBuilder()
+        .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
+        .TypeConstraint("Tind", std::vector<MLDataType>{
+                                    DataTypeImpl::GetTensorType<int32_t>(),
+                                    DataTypeImpl::GetTensorType<int64_t>()}),
+    Gather);
+
+// explicit negative axis support
 ONNX_OPERATOR_KERNEL_EX(
     Gather,
     kOnnxDomain,
-    1,
+    11,
     kCudaExecutionProvider,
     KernelDefBuilder()
         .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes())
diff --git a/onnxruntime/core/providers/cuda/tensor/slice.cc b/onnxruntime/core/providers/cuda/tensor/slice.cc
index 88cf111ed7..29b3a87b1e 100644
--- a/onnxruntime/core/providers/cuda/tensor/slice.cc
+++ b/onnxruntime/core/providers/cuda/tensor/slice.cc
@@ -7,34 +7,62 @@
 
 namespace onnxruntime {
 namespace cuda {
-#define REGISTER_VERSIONED_TYPED_SLICE(TIND)                                                                                                       \
-  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(                                                                                                         \
-      Slice,                                                                                                                                       \
-      kOnnxDomain,                                                                                                                                 \
-      1, 9,                                                                                                                                        \
-      TIND,                                                                                                                                        \
-      kCudaExecutionProvider,                                                                                                                      \
-      KernelDefBuilder().TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()).TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIND>()), \
+#define REGISTER_VERSIONED_TYPED_SLICE(TIND)                            \
+  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(                              \
+      Slice,                                                            \
+      kOnnxDomain,                                                      \
+      1, 9,                                                             \
+      TIND,                                                             \
+      kCudaExecutionProvider,                                           \
+      KernelDefBuilder()                                                \
+          .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()) \
+          .TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIND>()), \
       Slice<TIND, false>);
 
 REGISTER_VERSIONED_TYPED_SLICE(int32_t)
 REGISTER_VERSIONED_TYPED_SLICE(int64_t)
 REGISTER_VERSIONED_TYPED_SLICE(float)
 
-#define REGISTER_V10_TYPED_SLICE(TIND)                                                                                                                                                                                                                                                                         \
-  ONNX_OPERATOR_TYPED_KERNEL_EX(                                                                                                                                                                                                                                                                               \
-      Slice,                                                                                                                                                                                                                                                                                                   \
-      kOnnxDomain,                                                                                                                                                                                                                                                                                             \
-      10,                                                                                                                                                                                                                                                                                                      \
-      TIND,                                                                                                                                                                                                                                                                                                    \
-      kCudaExecutionProvider,                                                                                                                                                                                                                                                                                  \
-      KernelDefBuilder().InputMemoryType<OrtMemTypeCPUInput>(1).InputMemoryType<OrtMemTypeCPUInput>(2).InputMemoryType<OrtMemTypeCPUInput>(3).InputMemoryType<OrtMemTypeCPUInput>(4).TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()).TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIND>()), \
+#define REGISTER_V10_TYPED_SLICE(TIND)                                  \
+  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(                              \
+      Slice,                                                            \
+      kOnnxDomain,                                                      \
+      10, 10,                                                           \
+      TIND,                                                             \
+      kCudaExecutionProvider,                                           \
+      KernelDefBuilder()                                                \
+          .InputMemoryType<OrtMemTypeCPUInput>(1)                       \
+          .InputMemoryType<OrtMemTypeCPUInput>(2)                       \
+          .InputMemoryType<OrtMemTypeCPUInput>(3)                       \
+          .InputMemoryType<OrtMemTypeCPUInput>(4)                       \
+          .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()) \
+          .TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIND>()), \
       Slice<TIND, true>);
 
 REGISTER_V10_TYPED_SLICE(int32_t)
 REGISTER_V10_TYPED_SLICE(int64_t)
 REGISTER_V10_TYPED_SLICE(float)
 
+#define REGISTER_V11_TYPED_SLICE(TIND)                                  \
+  ONNX_OPERATOR_TYPED_KERNEL_EX(                                        \
+      Slice,                                                            \
+      kOnnxDomain,                                                      \
+      11,                                                               \
+      TIND,                                                             \
+      kCudaExecutionProvider,                                           \
+      KernelDefBuilder()                                                \
+          .InputMemoryType<OrtMemTypeCPUInput>(1)                       \
+          .InputMemoryType<OrtMemTypeCPUInput>(2)                       \
+          .InputMemoryType<OrtMemTypeCPUInput>(3)                       \
+          .InputMemoryType<OrtMemTypeCPUInput>(4)                       \
+          .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()) \
+          .TypeConstraint("Tind", DataTypeImpl::GetTensorType<TIND>()), \
+      Slice<TIND, true>);
+
+REGISTER_V11_TYPED_SLICE(int32_t)
+REGISTER_V11_TYPED_SLICE(int64_t)
+REGISTER_V11_TYPED_SLICE(float)
+
 template <typename Tind, bool dynamic>
 Status Slice<Tind, dynamic>::ComputeInternal(OpKernelContext* ctx) const {
   const auto* input_tensor = ctx->Input<Tensor>(0);
diff --git a/onnxruntime/core/providers/cuda/tensor/squeeze.cc b/onnxruntime/core/providers/cuda/tensor/squeeze.cc
index 1f7c658def..9592588782 100644
--- a/onnxruntime/core/providers/cuda/tensor/squeeze.cc
+++ b/onnxruntime/core/providers/cuda/tensor/squeeze.cc
@@ -6,10 +6,21 @@
 namespace onnxruntime {
 namespace cuda {
 
+ONNX_OPERATOR_VERSIONED_KERNEL_EX(
+    Squeeze,
+    kOnnxDomain,
+    1, 10,
+    kCudaExecutionProvider,
+    KernelDefBuilder()
+        .Alias(0, 0)
+        .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
+    Squeeze);
+
+// explicit support for negative axis.
 ONNX_OPERATOR_KERNEL_EX(
     Squeeze,
     kOnnxDomain,
-    1,
+    11,
     kCudaExecutionProvider,
     KernelDefBuilder()
         .Alias(0, 0)
diff --git a/onnxruntime/core/providers/cuda/tensor/unsqueeze.cc b/onnxruntime/core/providers/cuda/tensor/unsqueeze.cc
index 3da89a7dfe..d1c303b8b3 100644
--- a/onnxruntime/core/providers/cuda/tensor/unsqueeze.cc
+++ b/onnxruntime/core/providers/cuda/tensor/unsqueeze.cc
@@ -6,10 +6,21 @@
 namespace onnxruntime {
 namespace cuda {
 
+ONNX_OPERATOR_VERSIONED_KERNEL_EX(
+    Unsqueeze,
+    kOnnxDomain,
+    1, 10,
+    kCudaExecutionProvider,
+    KernelDefBuilder()
+        .Alias(0, 0)
+        .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()),
+    Unsqueeze);
+
+// explicitly support negative axis
 ONNX_OPERATOR_KERNEL_EX(
     Unsqueeze,
     kOnnxDomain,
-    1,
+    11,
     kCudaExecutionProvider,
     KernelDefBuilder()
         .Alias(0, 0)