ad ut and e2e test for w8a16 quantization

Signed-off-by: yyt <yangyit139@gmail.com>

Author: TmacAaron

tests/e2e/singlecard/test_quantization.py

@@ -33,3 +33,18 @@ def test_quant_W8A8():
             quantization="ascend",
     ) as vllm_model:
         vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
+def test_quant_W8A16():
+    max_tokens = 5
+    example_prompts = [
+        "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
+    ]
+    with VllmRunner(
+            snapshot_download("vllm-ascend/Qwen3-0.6B-W8A16"),
+            max_model_len=8192,
+            enforce_eager=False,
+            gpu_memory_utilization=0.7,
+            quantization="ascend",
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)

tests/ut/quantization/test_w8a16.py

@@ -0,0 +1,88 @@
+import unittest
+from unittest.mock import MagicMock, patch
+
+import torch
+
+from tests.ut.base import TestBase
+from vllm_ascend.quantization.w8a16 import AscendW8A16LinearMethod
+
+
+class TestAscendW8A16LinearMethod(TestBase):
+
+    def setUp(self):
+        self.method = AscendW8A16LinearMethod()
+
+    def test_get_weight(self):
+        weight = self.method.get_weight(10, 20)
+        self.assertEqual(weight['weight'].dtype, torch.int8)
+        self.assertEqual(weight['weight'].shape, (20, 10))
+
+    @patch("torch_npu.npu_weight_quant_batchmatmul")
+    def test_apply_with_x_is_int8(self, mock_npu_weight_quant_batchmatmul):
+        layer = MagicMock()
+        layer.weight.data = torch.randn(128, 256)
+        layer.weight_scale.data = torch.randn(128, 1)
+        layer.weight_offset.data = torch.randn(128, 1)
+
+        x = torch.randn(32, 128)
+        bias = torch.randn(256)
+
+        expected_y_output = torch.randn(32, 256)
+        mock_npu_weight_quant_batchmatmul.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, bias)
+        expected_y_output += bias
+        self.assertTrue(torch.equal(output, expected_y_output))
+
+    @patch("vllm_ascend.quantization.w8a16.is_310p", return_value=True)
+    @patch("torch_npu.npu_weight_quant_batchmatmul")
+    def test_apply_with_x_is_310p(self, mock_npu_weight_quant_batchmatmul, mock_is_310p):
+        layer = MagicMock()
+        layer.weight.data = torch.randn(128, 256)
+        layer.weight_scale.data = torch.randn(128, 1)
+        layer.weight_offset.data = torch.randn(128, 1)
+
+        x = torch.randn(32, 128)
+        bias = torch.randn(256)
+
+        expected_y_output = torch.randn(32, 256)
+        mock_npu_weight_quant_batchmatmul.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, bias)
+        expected_y_output += bias
+        self.assertTrue(torch.equal(output, expected_y_output))
+
+    @patch("vllm_ascend.quantization.w8a16.is_enable_nz")
+    @patch('torch_npu.npu_format_cast')
+    def test_process_weights_after_loading_not_nz(self, mock_npu_format_cast,
+                                                  mock_is_nz):
+        layer = MagicMock()
+        layer.weight.data = torch.randn(128, 256)
+        layer.weight_scale.data = torch.randn(128, 1)
+        layer.weight_offset.data = torch.randn(128, 1)
+
+        mock_is_nz.return_value = 0
+        mock_npu_format_cast.return_value = MagicMock
+        self.method.process_weights_after_loading(layer)
+
+        self.assertEqual(layer.weight_scale.data.shape, (128, ))
+        self.assertEqual(layer.weight_offset.data.shape, (128, ))
+        mock_npu_format_cast.assert_not_called()
+
+    @patch("vllm_ascend.quantization.w8a16.is_enable_nz")
+    @patch('torch_npu.npu_format_cast')
+    def test_process_weights_after_loading_nz(self, mock_npu_format_cast,
+                                              mock_is_nz):
+        layer = MagicMock()
+
+        layer.weight.data = torch.randn(128, 256)
+        layer.weight_scale.data = torch.randn(128, 1)
+        layer.weight_offset.data = torch.randn(128, 1)
+
+        mock_is_nz.return_value = 1
+        mock_npu_format_cast.return_value = MagicMock
+        self.method.process_weights_after_loading(layer)
+
+        self.assertEqual(layer.weight_scale.data.shape, (128, ))
+        self.assertEqual(layer.weight_offset.data.shape, (128, ))
+        mock_npu_format_cast.assert_called_once()