diff --git a/onnxruntime/core/providers/cpu/tensor/upsample.h b/onnxruntime/core/providers/cpu/tensor/upsample.h
index 690008456b..2f65b22947 100644
--- a/onnxruntime/core/providers/cpu/tensor/upsample.h
+++ b/onnxruntime/core/providers/cpu/tensor/upsample.h
@@ -17,19 +17,32 @@ enum UpsampleMode {
 
 class UpsampleBase {
  protected:
-  UpsampleBase(OpKernelInfo info) {
+  UpsampleBase(OpKernelInfo info): scales_cached_(false) {
     std::string mode;
     ORT_ENFORCE(info.GetAttr<std::string>("mode", &mode).IsOK());
     mode_ = StringToUpsampleMode(mode);
 
-    if (info.GetInputCount() == 1) {
+    auto input_count = info.GetInputCount();
+    if (input_count == 1) {
       ORT_ENFORCE(info.GetAttrs<float>("scales", scales_).IsOK());
       ScalesValidation(scales_, mode_);
     }
+
+    // opset 9
+    if (input_count > 1) {
+      const Tensor* scale;
+      bool get_scale = info.TryGetConstantInput(1, &scale);
+
+      if (get_scale) {
+        ParseScalesData(scale, scales_);
+        scales_cached_ = true;
+      }
+    }
   }
 
   UpsampleMode mode_;
   std::vector<float> scales_;
+  bool scales_cached_;
 
   UpsampleMode StringToUpsampleMode(const std::string& mode) {
     if (strcmp(mode.c_str(), UpsampleModeNN) == 0) {
@@ -53,28 +66,7 @@ class UpsampleBase {
                           "Upsample: linear mode upsample only support bilinear, the first 2 scales should be 1.");
     }
   }
-};
 
-template <typename T>
-class Upsample : public UpsampleBase, public OpKernel {
- public:
-  Upsample(OpKernelInfo info) : UpsampleBase(info), OpKernel(info), scales_cached_(false) {
-    if (info.GetInputCount() > 1) {
-      const Tensor* scale;
-      bool get_scale = info.TryGetConstantInput(1, &scale);
-
-      if (get_scale) {
-        ParseScalesData(scale, scales_);
-        scales_cached_ = true;
-      }
-    }
-  }
-
-  Status Compute(OpKernelContext* context) const override;
-
-  Status BaseCompute(OpKernelContext* context, const std::vector<float>& scales) const;
-
-private:
   void ParseScalesData(const Tensor* scale, std::vector<float>& scales) const {
     const float* scale_data = scale->template Data<float>();
     int64_t scales_size = scale->Shape().Size();
@@ -82,9 +74,17 @@ private:
     memcpy(scales.data(), scale_data, scales_size * sizeof(float));
     ScalesValidation(scales, mode_);
   }
+};
 
-private:
-  bool scales_cached_;
+template <typename T>
+class Upsample : public UpsampleBase, public OpKernel {
+ public:
+  Upsample(OpKernelInfo info) : UpsampleBase(info), OpKernel(info) {
+  }
+
+  Status Compute(OpKernelContext* context) const override;
+
+  Status BaseCompute(OpKernelContext* context, const std::vector<float>& scales) const;
 };
 
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/cuda/cuda_execution_provider.cc b/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
index c5c5f29a51..cdf8f06b5b 100644
--- a/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
+++ b/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
@@ -498,10 +498,10 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, DynamicSlice);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int64_t, DynamicSlice);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, Compress);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, float, Upsample);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, double, Upsample);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, MLFloat16, Upsample);
-class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, int32_t, Upsample);
+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);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, int32_t, Upsample);
 
 static void RegisterCudaKernels(std::function<void(KernelCreateInfo&&)> fn) {
   fn(BuildKernel<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MemcpyFromHost)>());
@@ -761,10 +761,10 @@ static void RegisterCudaKernels(std::function<void(KernelCreateInfo&&)> fn) {
   fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int32_t, DynamicSlice)>());
   fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, int64_t, DynamicSlice)>());
   fn(BuildKernel<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, Compress)>());
-  fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, float, Upsample)>());
-  fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, double, Upsample)>());
-  fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, MLFloat16, Upsample)>());
-  fn(BuildKernel<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, int32_t, Upsample)>());
+  fn(BuildKernel<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, float, Upsample)>());
+  fn(BuildKernel<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, double, Upsample)>());
+  fn(BuildKernel<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, MLFloat16, Upsample)>());
+  fn(BuildKernel<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, int32_t, Upsample)>());
 }
 
 }  // namespace cuda
diff --git a/onnxruntime/core/providers/cuda/tensor/upsample.cc b/onnxruntime/core/providers/cuda/tensor/upsample.cc
index 11dcec7a9f..8347077852 100644
--- a/onnxruntime/core/providers/cuda/tensor/upsample.cc
+++ b/onnxruntime/core/providers/cuda/tensor/upsample.cc
@@ -11,13 +11,15 @@ namespace onnxruntime {
 namespace cuda {
 
 #define REGISTER_KERNEL_TYPED(T)                                  \
-  ONNX_OPERATOR_TYPED_KERNEL_EX(                                  \
+  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(                        \
       Upsample,                                                   \
       kOnnxDomain,                                                \
       7,                                                          \
+      9,                                                          \
       T,                                                          \
       kCudaExecutionProvider,                                     \
       KernelDefBuilder()                                          \
+          .InputMemoryType<OrtMemTypeCPUInput>(1)                 \
           .TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
       Upsample<T>);
 
@@ -27,7 +29,7 @@ REGISTER_KERNEL_TYPED(MLFloat16)
 REGISTER_KERNEL_TYPED(int32_t)
 
 template <typename T>
-Status Upsample<T>::ComputeInternal(OpKernelContext* context) const {
+Status Upsample<T>::BaseCompute(OpKernelContext* context, const std::vector<float>& scales) const {
   const Tensor* X = context->Input<Tensor>(0);
   ORT_ENFORCE(nullptr != X);
   const std::vector<int64_t>& X_dims = X->Shape().GetDims();
@@ -35,12 +37,12 @@ Status Upsample<T>::ComputeInternal(OpKernelContext* context) const {
   if (rank == 0)
     return Status(ONNXRUNTIME, INVALID_ARGUMENT, "Upsample: input tensor cannot be scalar.");
 
-  if (rank != scales_.size())
+  if (rank != scales.size())
     return Status(ONNXRUNTIME, INVALID_ARGUMENT, "Upsample: input tensor's dimension does not match the scales.");
 
   std::vector<int64_t> Y_dims;
   for (std::size_t i = 0; i < rank; i++) {
-    Y_dims.push_back(static_cast<int64_t>(scales_[i] * X_dims[i]));
+    Y_dims.push_back(static_cast<int64_t>(scales[i] * X_dims[i]));
   }
   Tensor* Y = context->Output(0, Y_dims);
   typedef typename ToCudaType<T>::MappedType CudaT;
@@ -55,14 +57,13 @@ Status Upsample<T>::ComputeInternal(OpKernelContext* context) const {
   CudaAsyncBuffer<fast_divmod> output_div_pitches(this, device_id, rank);
   gsl::span<fast_divmod> div_strides_span = output_div_pitches.CpuSpan();
 
-  
   CudaAsyncBuffer<fast_divmod> scales_div(this, device_id, rank);
   gsl::span<fast_divmod> scales_div_span = scales_div.CpuSpan();
 
   for (int i = 0; i < rank; ++i) {
     input_stride_span[i] = input_pitches[i];
     div_strides_span[i] = fast_divmod(gsl::narrow_cast<int>(output_pitches[i]));
-    scales_div_span[i] = fast_divmod(gsl::narrow_cast<int>(ceil(scales_[i])));
+    scales_div_span[i] = fast_divmod(gsl::narrow_cast<int>(ceil(scales[i])));
   }
   input_strides.CopyToGpu();
   output_div_pitches.CopyToGpu();
@@ -88,5 +89,21 @@ Status Upsample<T>::ComputeInternal(OpKernelContext* context) const {
   return Status::OK();
 }
 
+template <typename T>
+Status Upsample<T>::ComputeInternal(OpKernelContext* context) const {
+  // Opset 7
+  if (OpKernel::Node().InputDefs().size() == 1 || scales_cached_) {
+    return BaseCompute(context, scales_);
+  }
+
+  // Opset 9
+  const Tensor* scales = context->Input<Tensor>(1);
+  ORT_ENFORCE(scales != nullptr);
+  int64_t scales_size = scales->Shape().Size();
+  std::vector<float> scales_arrary(scales_size);
+  ParseScalesData(scales, scales_arrary);
+  return BaseCompute(context, scales_arrary);
+}
+
 }  // namespace cuda
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/cuda/tensor/upsample.h b/onnxruntime/core/providers/cuda/tensor/upsample.h
index 1d6823a29d..1c098bbaee 100644
--- a/onnxruntime/core/providers/cuda/tensor/upsample.h
+++ b/onnxruntime/core/providers/cuda/tensor/upsample.h
@@ -17,6 +17,7 @@ class Upsample : public UpsampleBase, public CudaKernel {
   }
 
   Status ComputeInternal(OpKernelContext* context) const override;
+  Status BaseCompute(OpKernelContext* context, const std::vector<float>& scales) const;
 };
 
 }  // namespace cuda
diff --git a/onnxruntime/test/onnx/main.cc b/onnxruntime/test/onnx/main.cc
index 4d9bb61d94..dff32db9e7 100644
--- a/onnxruntime/test/onnx/main.cc
+++ b/onnxruntime/test/onnx/main.cc
@@ -304,7 +304,6 @@ int real_main(int argc, char* argv[]) {
       {"operator_rnn_single_layer", "disable reason"},
       {"prelu_broadcast", "disable reason"},
       {"prelu_example", "disable reason"},
-      {"upsample_nearest", "opset 9 not supported yet"},
       {"sinh_example", "opset 9 not supported yet"},
       {"cosh_example", "opset 9 not supported yet"},
       {"asinh_example", "opset 9 not supported yet"},