diff --git a/backends/vulkan/test/op_tests/quantize_test.cpp b/backends/vulkan/test/op_tests/quantize_test.cpp
index 6ff61dac19b..0ac08b65972 100644
--- a/backends/vulkan/test/op_tests/quantize_test.cpp
+++ b/backends/vulkan/test/op_tests/quantize_test.cpp
@@ -156,3 +156,387 @@ void check_quantize_args(
       " actual quant_max: ",
       quant_max);
 }
+/*
+ * Reference implementation of quantize_per_token
+ */
+at::Tensor quantize_per_token_reference_impl(
+    const at::Tensor& input,
+    const at::Tensor& scale,
+    const at::Tensor& 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);
+
+  // Calculate number of tokens
+  int num_tokens = 1;
+  for (int i = 0; i < input.dim() - 1; i++) {
+    num_tokens *= input.size(i);
+  }
+
+  // Verify that the number of tokens matches the size of scale and zero_point
+  // tensors
+  assert(num_tokens == scale.numel());
+  assert(num_tokens == zero_point.numel());
+
+  // Reshape input to [num_tokens, last_dim]
+  at::Tensor reshaped_input = input.reshape({num_tokens, input.size(-1)});
+  at::Tensor reshaped_out = out.reshape({num_tokens, input.size(-1)});
+
+  // Quantize each token separately
+  for (int token_idx = 0; token_idx < num_tokens; token_idx++) {
+    // Use float for scale since Vulkan doesn't support double
+    float token_scale = scale[token_idx].item<float>();
+    // Use int for zero_point since Vulkan doesn't support int64_t
+    int token_zero_point = zero_point[token_idx].item<int>();
+
+    float inv_scale = 1.0 / token_scale;
+
+    // Quantize the token
+    for (int i = 0; i < input.size(-1); i++) {
+      float value = reshaped_input[token_idx][i].item<float>();
+      int qvalue = token_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) {
+        reshaped_out[token_idx][i] = static_cast<uint8_t>(qvalue);
+      } else if (dtype == at::kChar) {
+        reshaped_out[token_idx][i] = static_cast<int8_t>(qvalue);
+      } else if (dtype == at::kShort) {
+        reshaped_out[token_idx][i] = static_cast<int16_t>(qvalue);
+      } else if (dtype == at::kInt) {
+        reshaped_out[token_idx][i] = static_cast<int32_t>(qvalue);
+      } else if (dtype == at::kLong) {
+        reshaped_out[token_idx][i] = static_cast<int64_t>(qvalue);
+      }
+    }
+  }
+
+  return out;
+}
+
+void test_vulkan_quantize_per_token_impl(
+    const std::vector<int>& input_sizes,
+    const std::vector<float>& scales,
+    const std::vector<int>& zero_points,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType in_dtype,
+    at::ScalarType dtype,
+    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_token(
+    const std::vector<int>& input_sizes,
+    const std::vector<float>& scales,
+    const std::vector<int>& zero_points,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType in_dtype = at::kFloat,
+    at::ScalarType dtype = at::kInt) {
+  // Test with buffer storage
+  test_vulkan_quantize_per_token_impl(
+      input_sizes,
+      scales,
+      zero_points,
+      quant_min,
+      quant_max,
+      in_dtype,
+      dtype,
+      vkcompute::utils::kBuffer,
+      vkcompute::utils::kBuffer);
+
+  // Test with texture storage
+  test_vulkan_quantize_per_token_impl(
+      input_sizes,
+      scales,
+      zero_points,
+      quant_min,
+      quant_max,
+      in_dtype,
+      dtype,
+      vkcompute::utils::kTexture3D,
+      vkcompute::utils::kTexture3D);
+}
+
+void test_reference_quantize_per_token(
+    const std::vector<int>& input_sizes,
+    const std::vector<float>& scales,
+    const std::vector<int>& zero_points,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType in_dtype = at::kFloat,
+    at::ScalarType dtype = at::kInt) {
+  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(in_dtype));
+
+  // Fill with a simple pattern: values from 0 to 1 in steps
+  float step = 1.0 / (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);
+
+  // Calculate number of tokens
+  int num_tokens = 1;
+  for (int i = 0; i < input.dim() - 1; i++) {
+    num_tokens *= input.size(i);
+  }
+
+  // Verify that the number of tokens matches the size of scales and zero_points
+  ASSERT_EQ(num_tokens, scales.size());
+  ASSERT_EQ(num_tokens, zero_points.size());
+
+  // Create scale and zero_point tensors
+  at::Tensor scale_tensor =
+      at::tensor(scales, at::device(at::kCPU).dtype(at::kDouble));
+  at::Tensor zero_point_tensor =
+      at::tensor(zero_points, at::device(at::kCPU).dtype(at::kLong));
+
+  // Get reference output
+  at::Tensor reference_out = quantize_per_token_reference_impl(
+      input, scale_tensor, zero_point_tensor, quant_min, quant_max, dtype);
+
+  // Get implementation output
+  at::Tensor impl_out = torch::executor::native::quantize_per_token_aten(
+      input, scale_tensor, zero_point_tensor, 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_out);
+  if (!output_correct) {
+    std::cout << "\n"
+              << "Failed with parameters: " << std::endl;
+    std::cout << "  scale(s):";
+    for (size_t i = 0; i < scales.size(); i++) {
+      std::cout << " " << scales[i] << " ";
+    }
+    std::cout << "" << std::endl;
+    std::cout << "  zero_point(s):";
+    for (size_t i = 0; i < zero_points.size(); i++) {
+      std::cout << " " << zero_points[i] << " ";
+    }
+    std::cout << "" << 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_out << std::endl;
+  }
+
+  ASSERT_TRUE(output_correct);
+}
+
+void test_vulkan_quantize_per_token_impl(
+    const std::vector<int>& input_sizes,
+    const std::vector<float>& scales,
+    const std::vector<int>& zero_points,
+    int64_t quant_min,
+    int64_t quant_max,
+    at::ScalarType in_dtype = at::kFloat,
+    at::ScalarType dtype = at::kInt,
+    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);
+  int num_tokens = 1;
+  for (int i = 0; i < input_sizes.size() - 1; i++) {
+    num_tokens *= input_sizes[i];
+  }
+
+  ASSERT_EQ(num_tokens, scales.size());
+  ASSERT_EQ(num_tokens, zero_points.size());
+
+  // Create input tensor with random values
+  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(in_dtype));
+  at::Tensor scale_tensor =
+      at::tensor(scales, at::device(at::kCPU).dtype(at::kDouble));
+  at::Tensor zero_point_tensor =
+      at::tensor(zero_points, at::device(at::kCPU).dtype(at::kLong));
+
+  // Get reference output to show what we would compare against
+  at::Tensor reference_out = torch::executor::native::quantize_per_token_aten(
+      input, scale_tensor, zero_point_tensor, quant_min, quant_max, dtype);
+
+  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);
+  IOValueRef r_scale = graph.add_input_tensor(
+      scale_tensor.sizes().vec(), vkapi::kFloat, in_storage);
+  IOValueRef r_zero_point = graph.add_input_tensor(
+      zero_point_tensor.sizes().vec(), vkapi::kInt, in_storage);
+
+  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_token.default")
+  (graph,
+   {
+       r_input.value,
+       r_scale.value,
+       r_zero_point.value,
+       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();
+
+  // Copy input data to GPU
+  graph.copy_into_staging(
+      r_input.staging, input.const_data_ptr(), input.numel());
+
+  // Convert scale tensor to float and copy to GPU
+  at::Tensor scale_float = scale_tensor.to(at::kFloat);
+  graph.copy_into_staging(
+      r_scale.staging, scale_float.const_data_ptr(), scale_float.numel());
+
+  // Convert zero_point tensor to int and copy to GPU
+  at::Tensor zero_point_int = zero_point_tensor.to(at::kInt);
+  graph.copy_into_staging(
+      r_zero_point.staging,
+      zero_point_int.const_data_ptr(),
+      zero_point_int.numel());
+
+  // Execute the graph
+  graph.execute();
+
+  // Copy output data back to CPU
+  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
+  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(s):";
+    for (size_t i = 0; i < scales.size(); i++) {
+      std::cout << " " << scales[i] << " ";
+    }
+    std::cout << "" << std::endl;
+    std::cout << "  zero_point(s):";
+    for (size_t i = 0; i < zero_points.size(); i++) {
+      std::cout << " " << zero_points[i] << " ";
+    }
+    std::cout << "" << std::endl;
+    std::cout << "  quant_min: " << quant_min << std::endl;
+    std::cout << "  quant_max: " << quant_max << std::endl;
+    std::cout << "  storage type: "
+              << (in_storage == vkcompute::utils::kBuffer ? "buffer"
+                                                          : "texture")
+              << 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_token_float_to_int8) {
+  std::vector<float> scales = {0.1, 0, 0.3, 0.1, 0.2, 0.3};
+  std::vector<int> zero_points = {1, 2, 3, 0, -1, -2};
+
+  test_reference_quantize_per_token(
+      {2, 3, 4}, // input sizes (2*3=6 tokens)
+      scales,
+      zero_points,
+      -128, // quant_min
+      127, // quant_max
+      at::kFloat,
+      at::kChar);
+}
+
+TEST(
+    VulkanQuantizePerTensorTest,
+    test_reference_quantize_per_token_float_to_int32) {
+  std::vector<float> scales = {0.1, 0, 0.3, 0.1, 0.2, 0.3};
+  std::vector<int> zero_points = {1, 2, 3, 0, -1, -2};
+
+  test_reference_quantize_per_token(
+      {2, 3, 4}, // input sizes (2*3=6 tokens)
+      scales,
+      zero_points,
+      std::numeric_limits<int32_t>::min(), // quant_min
+      std::numeric_limits<int32_t>::max(), // quant_max
+      at::kFloat,
+      at::kInt);
+}
+
+TEST(
+    VulkanQuantizePerTensorTest,
+    test_reference_quantize_per_token_half_to_int32) {
+  std::vector<float> scales = {0.1, 0, 0.3, 0.1, 0.2, 0.3};
+  std::vector<int> zero_points = {1, 2, 3, 0, -1, -2};
+
+  test_reference_quantize_per_token(
+      {2, 3, 4}, // input sizes (2*3=6 tokens)
+      scales,
+      zero_points,
+      std::numeric_limits<int32_t>::min(), // quant_min
+      std::numeric_limits<int32_t>::max(), // quant_max
+      at::kHalf,
+      at::kInt);
+}
+
+TEST(
+    VulkanQuantizePerTensorTest,
+    test_reference_quantize_per_token_half_to_uint8) {
+  std::vector<float> scales = {0.1, 0, 0.3, 0.1, 0.2, 0.3};
+  std::vector<int> zero_points = {1, 2, 3, 0, -1, -2};
+
+  test_reference_quantize_per_token(
+      {2, 3, 4}, // input sizes (2*3=6 tokens)
+      scales,
+      zero_points,
+      0, // quant_min
+      255, // quant_max
+      at::kHalf,
+      at::kByte);
+}