[ET-VK][testing] Q/DQ/CQP op comprehensive delegate dynamic quantization testing

Pull Request resolved: #12210

# Context

We need to ensure that most of the operators that were created work in tandem with each other for dynamic quantization.

# Changes

This creates two test cases to test the per_token and per_tensor pipeline to ensure that the whole full quantization workflow works as intended.
ghstack-source-id: 295489091
@exported-using-ghexport

Differential Revision: [D77746139](https://our.internmc.facebook.com/intern/diff/D77746139/)

Author: morelos

backends/vulkan/test/test_vulkan_delegate.py

@@ -46,6 +46,9 @@ def lower_module(
     if dynamic_shapes is not None:
         compile_options["require_dynamic_shapes"] = True
 
+    # Enable downcast_64_bit by default to handle float64/int64 tensors
+    compile_options["downcast_64_bit"] = True
+
     edge_compile_config = EdgeCompileConfig(
         _skip_dim_order=False,  # TODO(T182928844): Delegate dim order op to backend.
     )
@@ -1964,3 +1967,99 @@ def forward(self, x):
                     GroupNormModule(num_groups, num_channels),
                     sample_inputs,
                 )
+
+    def test_vulkan_backend_full_quantization_workflow(self):
+        class FullQuantizationWorkflowModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                # Step 1: Choose quantization parameters per tensor
+                scale, zero_point = (
+                    torch.ops.quantized_decomposed.choose_qparams.tensor(
+                        x,
+                        quant_min=-2147483648,  # int32 min
+                        quant_max=2147483647,  # int32 max
+                        eps=1e-5,
+                        dtype=torch.int32,
+                    )
+                )
+
+                # Step 2: Quantize using the calculated parameters
+                quantized = torch.ops.quantized_decomposed.quantize_per_tensor.tensor(
+                    x,
+                    scale,
+                    zero_point,
+                    quant_min=-2147483648,  # int32 min
+                    quant_max=2147483647,  # int32 max
+                    dtype=torch.int32,
+                )
+
+                # Step 3: Dequantize back to float
+                dequantized = (
+                    torch.ops.quantized_decomposed.dequantize_per_tensor.tensor(
+                        quantized,
+                        scale,
+                        zero_point,
+                        quant_min=-2147483648,  # int32 min
+                        quant_max=2147483647,  # int32 max
+                        dtype=torch.int32,
+                    )
+                )
+
+                return dequantized
+
+        full_workflow_module = FullQuantizationWorkflowModule()
+        sample_inputs = (torch.rand(size=(2, 3, 4), dtype=torch.float32),)
+
+        # Use higher tolerance since quantization introduces some error
+        self.lower_module_and_test_output(
+            full_workflow_module, sample_inputs, atol=5e-3, rtol=5e-3
+        )
+
+    def test_vulkan_backend_full_per_token_quantization_workflow(self):
+        class FullPerTokenQuantizationWorkflowModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                # Step 1: Choose quantization parameters per token
+                scale, zero_point = (
+                    torch.ops.quantized_decomposed.choose_qparams_per_token_asymmetric.default(
+                        x,
+                        dtype=torch.int32,
+                    )
+                )
+
+                # Step 2: Quantize using the calculated parameters per token
+                quantized = torch.ops.quantized_decomposed.quantize_per_token.default(
+                    x,
+                    scale,
+                    zero_point,
+                    quant_min=-2147483648,  # int32 min
+                    quant_max=2147483647,  # int32 max
+                    dtype=torch.int32,
+                )
+
+                # Step 3: Dequantize back to float per token
+                dequantized = (
+                    torch.ops.quantized_decomposed.dequantize_per_token.default(
+                        quantized,
+                        scale,
+                        zero_point,
+                        quant_min=-2147483648,  # int32 min
+                        quant_max=2147483647,  # int32 max
+                        dtype=torch.int32,
+                        output_dtype=torch.float32,
+                    )
+                )
+
+                return dequantized
+
+        full_per_token_workflow_module = FullPerTokenQuantizationWorkflowModule()
+        sample_inputs = (torch.rand(size=(6, 4), dtype=torch.float32),)
+
+        # Use higher tolerance since quantization introduces some error
+        self.lower_module_and_test_output(
+            full_per_token_workflow_module, sample_inputs, atol=5e-3, rtol=5e-3
+        )