diff --git a/backends/vulkan/test/op_tests/quantize_test.cpp b/backends/vulkan/test/op_tests/quantize_test.cpp
index 6e9962e88f..fe591ac282 100644
--- a/backends/vulkan/test/op_tests/quantize_test.cpp
+++ b/backends/vulkan/test/op_tests/quantize_test.cpp
@@ -275,6 +275,56 @@ void check_quantize_args(
       " actual quant_max: ",
       quant_max);
 }
+
+//
+// Reference Implementation
+//
+
+/*
+ * Reference implementation of quantize_per_tensor
+ */
+at::Tensor quantize_per_tensor_reference_impl(
+    const at::Tensor& input,
+    double scale,
+    int64_t zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType dtype) {
+  // Create output tensor with the target dtype
+  at::Tensor out = at::empty_like(input, dtype);
+
+  // Quantize the input tensor
+  float inv_scale = 1.0 / scale;
+
+  // Iterate through the tensor and quantize each element
+  at::Tensor float_input = input.to(at::kFloat);
+  at::Tensor float_values = float_input.flatten();
+
+  auto out_flat = out.flatten();
+
+  for (int i = 0; i < float_values.numel(); i++) {
+    float value = float_values[i].item<float>();
+    int64_t qvalue = zero_point + std::nearbyint(inv_scale * value);
+
+    qvalue = std::max<int64_t>(qvalue, quant_min);
+    qvalue = std::min<int64_t>(qvalue, quant_max);
+
+    if (dtype == at::kByte) {
+      out_flat[i] = static_cast<uint8_t>(qvalue);
+    } else if (dtype == at::kChar) {
+      out_flat[i] = static_cast<int8_t>(qvalue);
+    } else if (dtype == at::kShort) {
+      out_flat[i] = static_cast<int16_t>(qvalue);
+    } else if (dtype == at::kInt) {
+      out_flat[i] = static_cast<int32_t>(qvalue);
+    } else if (dtype == at::kLong) {
+      out_flat[i] = static_cast<int64_t>(qvalue);
+    }
+  }
+
+  return out.reshape(input.sizes());
+}
+
 /*
  * Reference implementation of quantize_per_token
  */
@@ -337,6 +387,17 @@ at::Tensor quantize_per_token_reference_impl(
   return out;
 }
 
+// Forward declaration of implementation functions
+void test_vulkan_quantize_per_tensor_impl(
+    const std::vector<int>& input_sizes,
+    float scale,
+    int zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType dtype,
+    const vkcompute::utils::StorageType in_storage,
+    const vkcompute::utils::StorageType out_storage);
+
 void test_vulkan_quantize_per_token_impl(
     const std::vector<int>& input_sizes,
     const std::vector<float>& scales,
@@ -347,6 +408,37 @@ void test_vulkan_quantize_per_token_impl(
     const vkcompute::utils::StorageType in_storage,
     const vkcompute::utils::StorageType out_storage);
 
+// Wrapper function to test both buffer and texture storage types
+void test_vulkan_quantize_per_tensor(
+    const std::vector<int>& input_sizes,
+    float scale,
+    int zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType dtype) {
+  // Test with buffer storage
+  test_vulkan_quantize_per_tensor_impl(
+      input_sizes,
+      scale,
+      zero_point,
+      quant_min,
+      quant_max,
+      dtype,
+      vkcompute::utils::kBuffer,
+      vkcompute::utils::kBuffer);
+
+  // Test with texture storage
+  test_vulkan_quantize_per_tensor_impl(
+      input_sizes,
+      scale,
+      zero_point,
+      quant_min,
+      quant_max,
+      dtype,
+      vkcompute::utils::kTexture3D,
+      vkcompute::utils::kTexture3D);
+}
+
 // Wrapper function to test both buffer and texture storage types
 void test_vulkan_quantize_per_token(
     const std::vector<int>& input_sizes,
@@ -378,6 +470,166 @@ void test_vulkan_quantize_per_token(
       vkcompute::utils::kTexture3D);
 }
 
+void test_reference_quantize_per_tensor(
+    const std::vector<int>& input_sizes,
+    float scale,
+    int zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType dtype) {
+  check_quantize_args(quant_min, quant_max, dtype);
+  std::vector<int64_t> input_sizes_int64(
+      input_sizes.begin(), input_sizes.end());
+  at::Tensor input =
+      at::zeros(input_sizes_int64, at::device(at::kCPU).dtype(at::kFloat));
+
+  // Fill with a simple pattern: values from 0 to 1 in steps
+  float step = 1.0f / (input.numel() - 1);
+  auto flat_input = input.flatten();
+  for (int i = 0; i < flat_input.numel(); i++) {
+    flat_input[i] = i * step;
+  }
+
+  // Reshape back to original dimensions
+  input = flat_input.reshape(input_sizes_int64);
+
+  // Get reference output
+  at::Tensor reference_out = quantize_per_tensor_reference_impl(
+      input, scale, zero_point, quant_min, quant_max, dtype);
+
+  // Get implementation output
+  at::Tensor impl_out = torch::executor::native::quantize_per_tensor_aten(
+      input, scale, zero_point, quant_min, quant_max, dtype);
+
+  // Convert to int for consistent display regardless of underlying type
+  at::Tensor reference_int = reference_out.to(at::kInt);
+  at::Tensor impl_int = impl_out.to(at::kInt);
+
+  const bool output_correct = at::equal(reference_int, impl_int);
+  if (!output_correct) {
+    at::Tensor diffs = at::abs(reference_int - impl_int);
+
+    std::cout << "\n"
+              << "Failed with parameters: " << std::endl;
+    std::cout << "  scale: " << scale << std::endl;
+    std::cout << "  zero_point: " << zero_point << std::endl;
+    std::cout << "  quant_min: " << quant_min << std::endl;
+    std::cout << "  quant_max: " << quant_max << std::endl;
+
+    std::cout << "input:" << std::endl;
+    std::cout << input << std::endl;
+    std::cout << "reference:" << std::endl;
+    std::cout << reference_int << std::endl;
+    std::cout << "my_reference:" << std::endl;
+    std::cout << impl_int << std::endl;
+  }
+
+  ASSERT_TRUE(output_correct);
+}
+
+void test_vulkan_quantize_per_tensor_impl(
+    const std::vector<int>& input_sizes,
+    float scale,
+    int zero_point,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType dtype,
+    const vkcompute::utils::StorageType in_storage =
+        vkcompute::utils::kTexture3D,
+    const vkcompute::utils::StorageType out_storage =
+        vkcompute::utils::kTexture3D) {
+  check_quantize_args(quant_min, quant_max, dtype);
+  std::vector<int64_t> input_sizes_int64(
+      input_sizes.begin(), input_sizes.end());
+  at::Tensor input =
+      at::rand(input_sizes_int64, at::device(at::kCPU).dtype(at::kFloat));
+
+  // Get reference output
+  at::Tensor reference_out = torch::executor::native::quantize_per_tensor_aten(
+      input, scale, zero_point, quant_min, quant_max, dtype);
+
+  // Build Vulkan quantize_per_tensor graph
+  using namespace vkcompute;
+
+  GraphConfig config;
+  config.set_storage_type_override(in_storage);
+  ComputeGraph graph(config);
+
+  IOValueRef r_input = graph.add_input_tensor(
+      input.sizes().vec(), from_at_scalartype(input.scalar_type()), in_storage);
+
+  const ValueRef r_scale = graph.add_scalar<double>(scale);
+  const ValueRef r_zero_point = graph.add_scalar<int64_t>(zero_point);
+  const ValueRef r_quant_min = graph.add_scalar<int64_t>(quant_min);
+  const ValueRef r_quant_max = graph.add_scalar<int64_t>(quant_max);
+
+  const ValueRef r_out = graph.add_tensor(
+      input.sizes().vec(), from_at_scalartype(dtype), out_storage);
+
+  VK_GET_OP_FN("quantize_per_tensor.default")
+  (graph,
+   {
+       r_input.value,
+       r_scale,
+       r_zero_point,
+       r_quant_min,
+       r_quant_max,
+       r_out,
+   });
+
+  ValueRef staging_out = graph.set_output_tensor(r_out);
+
+  graph.prepare();
+  graph.encode_prepack();
+  graph.prepack();
+  graph.encode_execute();
+
+  // Run Vulkan quantize_per_tensor
+  graph.copy_into_staging(
+      r_input.staging, input.const_data_ptr(), input.numel());
+
+  graph.execute();
+
+  at::Tensor vk_out = at::empty_like(reference_out).contiguous();
+  graph.copy_from_staging(
+      staging_out, vk_out.mutable_data_ptr(), vk_out.numel());
+
+  // Compare outputs
+  // For quantized types, we need to compare the actual integer values
+  at::Tensor reference_int = reference_out.to(at::kInt);
+  at::Tensor vk_int = vk_out.to(at::kInt);
+
+  const bool output_correct = at::equal(reference_int, vk_int);
+  if (!output_correct) {
+    at::Tensor diffs = at::abs(reference_int - vk_int);
+
+    std::cout << "\n"
+              << "Failed with parameters: " << std::endl;
+    std::cout << "  scale: " << scale << std::endl;
+    std::cout << "  zero_point: " << zero_point << std::endl;
+    std::cout << "  quant_min: " << quant_min << std::endl;
+    std::cout << "  quant_max: " << quant_max << std::endl;
+
+    std::cout << "input:" << std::endl;
+    std::cout << input << std::endl;
+    std::cout << "reference:" << std::endl;
+    std::cout << reference_int << std::endl;
+    std::cout << "vulkan:" << std::endl;
+    std::cout << vk_int << std::endl;
+  }
+
+  ASSERT_TRUE(output_correct);
+}
+
+TEST(VulkanQuantizePerTensorTest, test_reference_quantize_per_tensor_int8) {
+  test_reference_quantize_per_tensor(
+      {2, 3, 4}, // input sizes
+      0.1, // scale
+      0, // zero_point
+      -128, // quant_min
+      127, // quant_max
+      at::kChar);
+}
 void test_reference_quantize_per_token(
     const std::vector<int>& input_sizes,
     const std::vector<float>& scales,