init v0.23.0

Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>

tests/ut/_310p/quantization/test_modelslim_config_310.py

@@ -0,0 +1,138 @@
+#
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from unittest.mock import MagicMock, patch
+
+from vllm.model_executor.layers.fused_moe.config import FusedMoEConfig, FusedMoEParallelConfig
+from vllm.model_executor.layers.linear import LinearBase
+
+from tests.ut.base import TestBase
+from vllm_ascend._310p.fused_moe.fused_moe import AscendUnquantizedFusedMoEMethod310
+from vllm_ascend._310p.quantization.modelslim_config import AscendModelSlimConfig310
+from vllm_ascend.ops.linear import AscendUnquantizedLinearMethod
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.23.0"):
+    from vllm.model_executor.layers.fused_moe import FusedMoE
+else:
+    from vllm.model_executor.layers.fused_moe import RoutedExperts
+
+
+class TestAscendModelSlimConfig310(TestBase):
+    def setUp(self):
+        self.sample_config = {
+            "weight": "INT8",
+            "layer1.weight": "INT8",
+            "layer2.weight": "FLOAT",
+            "fused_layer.weight": "FLOAT",
+            "fused_layer.shard1.weight": "FLOAT",
+            "fused_layer.shard2.weight": "FLOAT",
+            "shard1.weight": "FLOAT",
+            "shard2.weight": "FLOAT",
+        }
+        self.ascend_config = AscendModelSlimConfig310(self.sample_config)
+        self.ascend_config.packed_modules_mapping = None
+
+    def test_get_quant_method_for_linear_310(self):
+        mock_config = MagicMock()
+        mock_config.model_config.hf_config.model_type = None
+        linear_layer = MagicMock(spec=LinearBase)
+        # Test skipped layer
+        with (
+            patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
+            patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=True),
+        ):
+            method = self.ascend_config.get_quant_method(linear_layer, ".attn")
+            self.assertIsInstance(method, AscendUnquantizedLinearMethod)
+
+        # Test quantized layer
+        mock_scheme = MagicMock()
+        with (
+            patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=False),
+            patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
+            patch("vllm_ascend._310p.quantization.modelslim_config.create_scheme_for_layer", return_value=mock_scheme),
+            patch(
+                "vllm_ascend._310p.quantization.modelslim_config.AscendLinearMethod", return_value=MagicMock()
+            ) as mock_ascend_linear,
+        ):
+            method = self.ascend_config.get_quant_method(linear_layer, ".attn")
+            self.assertIs(method, mock_ascend_linear.return_value)
+            mock_ascend_linear.assert_called_once_with(mock_scheme)
+
+    def test_get_quant_method_maps_lm_head_prefix_310(self):
+        config = AscendModelSlimConfig310({"language_model.lm_head.weight": "INT8"})
+        linear_layer = MagicMock(spec=LinearBase)
+        mock_config = MagicMock()
+        mock_config.model_config.hf_config.model_type = "qwen3_5_moe"
+        mock_scheme = MagicMock()
+
+        with (
+            patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
+            patch(
+                "vllm_ascend._310p.quantization.modelslim_config.create_scheme_for_layer",
+                return_value=mock_scheme,
+            ) as mock_create_scheme,
+            patch("vllm_ascend._310p.quantization.modelslim_config.AscendLinearMethod", return_value=MagicMock()),
+        ):
+            config.get_quant_method(linear_layer, "lm_head")
+
+        mock_create_scheme.assert_called_once_with(
+            quant_description=config.quant_description,
+            prefix="language_model.lm_head",
+            layer_type="linear",
+            packed_modules_mapping=config.packed_modules_mapping,
+        )
+
+    def test_get_quant_method_for_fused_moe_310(self):
+        if vllm_version_is("0.23.0"):
+            fused_moe_cls = FusedMoE
+        else:
+            fused_moe_cls = RoutedExperts
+        fused_moe_layer = MagicMock(spec=fused_moe_cls)
+        fused_moe_layer.moe = MagicMock(spec=FusedMoEConfig)
+        fused_moe_layer.moe_config = MagicMock(spec=FusedMoEConfig)
+        fused_moe_layer.moe_config.moe_backend = "auto"
+        fused_moe_layer.moe_config.moe_parallel_config = MagicMock(spec=FusedMoEParallelConfig)
+        fused_moe_layer.moe_config.moe_parallel_config.use_ep = True
+        fused_moe_layer.moe_config.moe_parallel_config.dp_size = 1
+        mock_config = MagicMock()
+        mock_config.model_config.hf_config.model_type = None
+        mock_config.compilation_config.custom_ops = ["all"]
+        mock_scheme = MagicMock()
+        # Test skipped layer
+        with (
+            patch("vllm.config.vllm.get_current_vllm_config", return_value=mock_config),
+            patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
+            patch("vllm_ascend.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
+            patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=True),
+        ):
+            method = self.ascend_config.get_quant_method(fused_moe_layer, ".moe")
+            self.assertIsInstance(method, AscendUnquantizedFusedMoEMethod310)
+
+        # Test quantized layer
+        mock_scheme = MagicMock()
+        with (
+            patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=False),
+            patch("vllm.config.vllm.get_current_vllm_config", return_value=mock_config),
+            patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
+            patch("vllm_ascend.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
+            patch("vllm_ascend._310p.quantization.modelslim_config.create_scheme_for_layer", return_value=mock_scheme),
+            patch(
+                "vllm_ascend._310p.quantization.modelslim_config.AscendFusedMoEMethod", return_value=MagicMock()
+            ) as fused_moe_method,
+        ):
+            method = self.ascend_config.get_quant_method(fused_moe_layer, ".moe")
+            self.assertIs(method, fused_moe_method.return_value)
+            fused_moe_method.assert_called_once_with(mock_scheme, fused_moe_layer.moe_config)

tests/ut/_310p/quantization/test_w8a8_dynamic_310.py

@@ -0,0 +1,145 @@
+#
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from unittest.mock import MagicMock, Mock, patch
+
+import torch
+
+from tests.ut.base import TestBase
+from vllm_ascend._310p.quantization.methods.w8a8_dynamic import (
+    AscendW8A8DynamicFusedMoEMethod310,
+    AscendW8A8DynamicLinearMethod310,
+)
+
+
+class TestAscendW8A8FusedMoEMethod310(TestBase):
+    num_experts = 8
+    hidden_size = 128
+    intermediate_size = 128
+
+    @patch("vllm_ascend._310p.quantization.methods.w8a8_dynamic.get_ep_group")
+    def setUp(self, mock_get_ep_group):
+        with patch(
+            "vllm_ascend._310p.quantization.methods.w8a8_dynamic.get_current_vllm_config"
+        ) as mock_get_current_vllm_config:
+            mock_vllm_config = Mock()
+            mock_vllm_config.quant_config = Mock(quant_description={"group_size": 0})
+            mock_vllm_config.scheduler_config = Mock(
+                max_num_batched_tokens=2048, max_model_len=2048, enable_chunked_prefill=False
+            )
+            mock_get_current_vllm_config.return_value = mock_vllm_config
+            mock_ep_group = Mock()
+            mock_get_ep_group.return_value = mock_ep_group
+            mock_ascend_config = Mock()
+
+            mock_ascend_config.enable_chunked_prefill = False
+
+            self.quant_method = AscendW8A8DynamicFusedMoEMethod310()
+
+    def test_get_weight_310(self):
+        param_dict = self.quant_method.get_weight(
+            self.num_experts, self.intermediate_size, self.hidden_size, torch.float16
+        )
+        self.assertEqual(param_dict["w13_weight"].dtype, torch.int8)
+        self.assertEqual(
+            param_dict["w13_weight"].shape, (self.num_experts, 2 * self.intermediate_size, self.hidden_size)
+        )
+        self.assertEqual(param_dict["w2_weight"].dtype, torch.int8)
+        self.assertEqual(param_dict["w2_weight"].shape, (self.num_experts, self.hidden_size, self.intermediate_size))
+
+    def test_get_dynamic_quant_param_310(self):
+        param_dict = self.quant_method.get_dynamic_quant_param(
+            self.num_experts, self.intermediate_size, self.hidden_size, torch.float16
+        )
+        self.assertEqual(param_dict["w13_weight_scale"].dtype, torch.float32)
+        self.assertEqual(param_dict["w13_weight_scale"].shape, (self.num_experts, 2 * self.intermediate_size, 1))
+        self.assertEqual(param_dict["w2_weight_scale"].dtype, torch.float32)
+        self.assertEqual(param_dict["w2_weight_scale"].shape, (self.num_experts, self.hidden_size, 1))
+
+
+class TestAscendW8A8DynamicLinearMethod310(TestBase):
+    def setUp(self):
+        self.method = AscendW8A8DynamicLinearMethod310()
+
+    def test_get_weight_310(self):
+        weight = self.method.get_weight(10, 20)
+        self.assertEqual(weight["weight"].dtype, torch.int8)
+        self.assertEqual(weight["weight"].shape, (20, 10))
+
+    def test_get_perchannel_param_310(self):
+        params = self.method.get_perchannel_param(10, torch.float32)
+
+        self.assertEqual(params["weight_scale"].dtype, torch.float32)
+        self.assertEqual(params["weight_offset"].dtype, torch.float32)
+
+        self.assertEqual(params["weight_scale"].shape, (10, 1))
+        self.assertEqual(params["weight_offset"].shape, (10, 1))
+
+    @patch("torch_npu.npu_dynamic_quant", create=True)
+    @patch("torch_npu.npu_quant_matmul")
+    def test_apply_310(self, mock_npu_quant_matmul, mock_npu_dynamic_quantize):
+        layer = MagicMock()
+        layer.weight = torch.randn(128, 256, dtype=torch.float16)
+        layer.weight_scale = torch.randn(128, dtype=torch.float32)
+        layer.params_dtype = torch.float16
+
+        x = torch.randn(32, 128, dtype=torch.float16)
+        expect_x_output = torch.randint(-128, 127, x.shape, dtype=torch.int8)
+        expect_pertoken_scale_output = torch.randn(x.shape[0], dtype=torch.float32)
+        mock_npu_dynamic_quantize.return_value = expect_x_output, expect_pertoken_scale_output
+
+        expected_y_output = torch.randn(32, 256)
+        mock_npu_quant_matmul.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, tp_rank=0)
+
+        mock_npu_dynamic_quantize.assert_called_with(x)
+        mock_npu_quant_matmul.assert_called_once()
+        (args, kwargs) = mock_npu_quant_matmul.call_args
+
+        # positional args
+        self.assertTrue(torch.equal(args[0], expect_x_output))
+        self.assertTrue(torch.equal(args[1], layer.weight.data))
+        self.assertTrue(torch.equal(args[2], layer.weight_scale))
+
+        # kwargs
+        self.assertTrue(torch.equal(kwargs["pertoken_scale"], expect_pertoken_scale_output))
+        self.assertTrue(kwargs["bias"] is None)
+        self.assertEqual(kwargs["output_dtype"], layer.params_dtype)
+
+        self.assertTrue(torch.equal(output, expected_y_output))
+
+    @patch("vllm_ascend.utils.is_310p", return_value=True)
+    @patch("torch_npu.npu_format_cast")
+    def test_process_weights_after_loading_calls_nz_format_cast_310p(self, mock_npu_format_cast, _mock_is_310p):
+        mock_npu_format_cast.side_effect = lambda x, fmt: x
+
+        layer = MagicMock()
+
+        # Attributes used by process_weights_after_loading()
+        layer.weight = MagicMock()
+        layer.weight_scale = MagicMock()
+        layer.weight_offset = MagicMock()
+
+        layer.weight.data = torch.randint(-127, 128, (128, 256), dtype=torch.int8)
+
+        layer.weight_scale.data = torch.randn(128, 1, dtype=torch.bfloat16)
+        layer.weight_offset.data = torch.randn(128, 1, dtype=torch.bfloat16)
+        # w2_weight_offset is reshaped to (N, -1); any (N, 1) is fine
+        layer.w2_weight_offset.data = torch.randn(128, 1, dtype=torch.bfloat16)
+
+        self.method.process_weights_after_loading(layer)
+
+        mock_npu_format_cast.assert_called_once()

tests/ut/_310p/quantization/test_w8a8_static_310.py

@@ -0,0 +1,151 @@
+#
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from unittest.mock import MagicMock, patch
+
+import torch
+
+from tests.ut.base import TestBase
+from vllm_ascend._310p.quantization.methods.w8a8_static import AscendW8A8LinearMethod310
+
+
+class TestAscendW8A8LinearMethod310(TestBase):
+    def setUp(self):
+        self.method = AscendW8A8LinearMethod310()
+
+    def test_get_weight_310(self):
+        weight = self.method.get_weight(10, 20)
+        self.assertEqual(weight["weight"].dtype, torch.int8)
+        self.assertEqual(weight["weight"].shape, (20, 10))
+
+    def test_get_pertensor_param_310(self):
+        params = self.method.get_pertensor_param(torch.float16)
+        self.assertEqual(params["input_scale"].dtype, torch.float16)
+        self.assertEqual(params["input_offset"].dtype, torch.int8)
+        self.assertEqual(params["input_scale"].shape, (1,))
+        self.assertEqual(params["input_offset"].shape, (1,))
+
+    def test_get_perchannel_param_310(self):
+        params = self.method.get_perchannel_param(10, torch.float16)
+
+        self.assertEqual(params["quant_bias"].dtype, torch.int32)
+        self.assertEqual(params["deq_scale"].dtype, torch.int64)
+        self.assertEqual(params["weight_scale"].dtype, torch.float16)
+        self.assertEqual(params["weight_offset"].dtype, torch.float16)
+
+        self.assertEqual(params["quant_bias"].shape, (10,))
+        self.assertEqual(params["deq_scale"].shape, (10,))
+        self.assertEqual(params["weight_scale"].shape, (10, 1))
+        self.assertEqual(params["weight_offset"].shape, (10, 1))
+
+    @patch("torch.ops.vllm.quantize")
+    @patch("torch_npu.npu_quant_matmul")
+    def test_apply_with_x_not_int8_310(self, mock_npu_quant_matmul, mock_quantize):
+        layer = MagicMock()
+        layer.aclnn_input_scale = torch.randn(256)
+        layer.aclnn_input_scale_reciprocal = 1.0 / layer.aclnn_input_scale
+        layer.aclnn_input_offset = torch.randint(-128, 127, (256,), dtype=torch.int8)
+        layer.weight = torch.randn(128, 256)
+        layer.deq_scale = torch.randn(128)
+        layer.quant_bias = torch.randint(-128, 127, (256,))
+        layer.params_dtype = torch.float16
+
+        x = torch.randn(32, 128)
+        expect_x_output = torch.randint(-128, 127, x.shape, dtype=torch.int8)
+        mock_quantize.return_value = expect_x_output
+
+        expected_y_output = torch.randn(32, 256)
+        mock_npu_quant_matmul.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, tp_rank=0)
+
+        mock_quantize.assert_called_with(
+            x,
+            layer.aclnn_input_scale,
+            layer.aclnn_input_scale_reciprocal,
+            layer.aclnn_input_offset,
+        )
+        mock_npu_quant_matmul.assert_called_once()
+        (args, kwargs) = mock_npu_quant_matmul.call_args
+
+        # positional args
+        self.assertTrue(torch.equal(args[0], expect_x_output))
+        self.assertTrue(torch.equal(args[1], layer.weight.data))
+        self.assertTrue(torch.equal(args[2], layer.deq_scale))
+
+        # kwargs
+        self.assertTrue(torch.equal(kwargs["bias"], layer.quant_bias))
+        self.assertEqual(kwargs["output_dtype"], layer.params_dtype)
+
+        self.assertTrue(torch.equal(output, expected_y_output))
+
+    @patch("torch.ops.vllm.quantize")
+    @patch("torch_npu.npu_quant_matmul")
+    def test_apply_with_x_is_int8_310(self, mock_npu_quant_matmul, mock_quantize):
+        layer = MagicMock()
+        layer.aclnn_input_scale = torch.randn(256)
+        layer.aclnn_input_offset = torch.randint(-128, 127, (256,), dtype=torch.int8)
+        layer.weight = torch.randn(128, 256)
+        layer.deq_scale = torch.randn(128)
+        layer.quant_bias = torch.randint(-128, 127, (256,))
+        layer.params_dtype = torch.float16
+
+        x = torch.randint(-128, 127, (32, 128), dtype=torch.int8)
+
+        expected_y_output = torch.randn(32, 256)
+        mock_npu_quant_matmul.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, tp_rank=0)
+
+        mock_quantize.assert_not_called()
+        mock_npu_quant_matmul.assert_called_once()
+        (args, kwargs) = mock_npu_quant_matmul.call_args
+
+        self.assertTrue(torch.equal(args[0], x))
+        self.assertTrue(torch.equal(args[1], layer.weight.data))
+        self.assertTrue(torch.equal(args[2], layer.deq_scale))
+
+        self.assertTrue(torch.equal(kwargs["bias"], layer.quant_bias))
+        self.assertEqual(kwargs["output_dtype"], layer.params_dtype)
+
+        self.assertTrue(torch.equal(output, expected_y_output))
+
+    @patch("vllm_ascend.utils.is_310p", return_value=True)
+    @patch("torch_npu.npu_format_cast")
+    def test_process_weights_after_loading_calls_nz_format_cast_310p(self, mock_npu_format_cast, _mock_is_310p):
+        mock_npu_format_cast.side_effect = lambda x, fmt: x
+
+        layer = MagicMock()
+
+        # Attributes used by process_weights_after_loading()
+        layer.weight = MagicMock()
+        layer.input_scale = MagicMock()
+        layer.input_offset = MagicMock()
+        layer.weight_scale = MagicMock()
+        layer.weight_offset = MagicMock()
+        layer.w2_weight_offset = MagicMock()
+
+        layer.weight.data = torch.randint(-127, 128, (128, 256), dtype=torch.int8)
+        layer.input_scale.data = torch.tensor([0.1], dtype=torch.float16)
+        layer.input_offset.data = torch.tensor([0], dtype=torch.int8)
+
+        layer.weight_scale.data = torch.randn(128, 1, dtype=torch.bfloat16)
+        layer.weight_offset.data = torch.randn(128, 1, dtype=torch.bfloat16)
+        # w2_weight_offset is reshaped to (N, -1); any (N, 1) is fine
+        layer.w2_weight_offset.data = torch.randn(128, 1, dtype=torch.bfloat16)
+
+        self.method.process_weights_after_loading(layer)
+
+        mock_npu_format_cast.assert_called_once()

tests/ut/_310p/quantization/test_w8a8s_310.py

@@ -0,0 +1,93 @@
+#
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from unittest.mock import MagicMock, patch
+
+import torch
+
+from tests.ut.base import TestBase
+from vllm_ascend._310p.quantization.methods.w8a8s import AscendW8A8SLinearMethod310
+
+
+class TestAscendW8A8SLinearMethod310(TestBase):
+    def setUp(self):
+        self.method = AscendW8A8SLinearMethod310()
+
+    def test_get_weight_310(self):
+        weight = self.method.get_weight(10, 20)
+        self.assertEqual(weight["weight"].dtype, torch.int8)
+        self.assertEqual(weight["weight"].shape, (20, 10))
+
+    def test_get_pertensor_param_310(self):
+        params = self.method.get_pertensor_param(torch.float16)
+        self.assertEqual(params["input_scale"].dtype, torch.float16)
+        self.assertEqual(params["input_offset"].dtype, torch.int8)
+        self.assertEqual(params["input_scale"].shape, (1,))
+        self.assertEqual(params["input_offset"].shape, (1,))
+
+    def test_get_perchannel_param_310(self):
+        params = self.method.get_perchannel_param(10, torch.float16)
+
+        self.assertEqual(params["quant_bias"].dtype, torch.int32)
+        self.assertEqual(params["deq_scale"].dtype, torch.int64)
+        self.assertEqual(params["quant_bias"].shape, (10,))
+        self.assertEqual(params["deq_scale"].shape, (10,))
+
+    @patch("torch.ops.vllm.quantize")
+    @patch("torch_npu.npu_quant_matmul")
+    def test_apply_with_x_not_int8_310(self, mock_npu_quant_matmul, mock_quantize):
+        layer = MagicMock()
+        layer.aclnn_input_scale = torch.randn(256)
+        layer.aclnn_input_scale_reciprocal = 1.0 / layer.aclnn_input_scale
+        layer.aclnn_input_offset = torch.randint(-128, 127, (256,), dtype=torch.int8)
+        layer.weight = torch.randn(128, 256)
+        layer.deq_scale = torch.randn(128)
+        layer.quant_bias = torch.randint(-128, 127, (256,))
+        layer.params_dtype = torch.float16
+
+        x = torch.randn(32, 128)
+        expect_x_output = torch.randint(-128, 127, x.shape, dtype=torch.int8)
+        mock_quantize.return_value = expect_x_output
+
+        expected_y_output = torch.randn(32, 256)
+        mock_npu_quant_matmul.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, tp_rank=0)
+
+        mock_quantize.assert_called_with(
+            x, layer.aclnn_input_scale, layer.aclnn_input_scale_reciprocal, layer.aclnn_input_offset
+        )
+        self.assertTrue(torch.equal(output, expected_y_output))
+
+    @patch("torch.ops.vllm.quantize")
+    @patch("torch_npu.npu_quant_matmul")
+    def test_apply_with_x_is_int8_310(self, mock_npu_quant_matmul, mock_quantize):
+        layer = MagicMock()
+        layer.aclnn_input_scale = torch.randn(256)
+        layer.aclnn_input_offset = torch.randint(-128, 127, (256,), dtype=torch.int8)
+        layer.weight = torch.randn(128, 256)
+        layer.deq_scale = torch.randn(128)
+        layer.quant_bias = torch.randint(-128, 127, (256,))
+        layer.params_dtype = torch.float16
+
+        x = torch.randint(-128, 127, (32, 128), dtype=torch.int8)
+
+        expected_y_output = torch.randn(32, 256)
+        mock_npu_quant_matmul.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, tp_rank=0)
+
+        mock_quantize.assert_not_called()
+        self.assertTrue(torch.equal(output, expected_y_output))

tests/ut/_310p/quantization/test_w8a8sc_310.py

@@ -0,0 +1,108 @@
+#
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from unittest.mock import MagicMock, patch
+
+import pytest
+import torch
+
+from tests.ut.base import TestBase
+from vllm_ascend._310p.quantization.methods.w8a8sc import AscendW8A8SCLinearMethod310
+
+
+class TestAscendW8A8SCLinearMethod310(TestBase):
+    def setUp(self):
+        self.method = AscendW8A8SCLinearMethod310()
+
+    def test_get_weight_310(self):
+        weight = self.method.get_weight(10, 20)
+        self.assertEqual(weight["weight"].dtype, torch.int8)
+        self.assertEqual(weight["weight"].shape, (10 * 20,))
+        self.assertEqual(weight["index"].dtype, torch.int8)
+        index_len = math.ceil(10 / 256) * math.ceil(20 / 128) * 8
+        self.assertEqual(weight["index"].shape, (index_len,))
+        self.assertEqual(weight["info"].dtype, torch.int64)
+        self.assertEqual(weight["info"].shape, (5,))
+
+    def test_get_pertensor_param_310(self):
+        params = self.method.get_pertensor_param(torch.float16)
+        self.assertEqual(params["input_scale"].dtype, torch.float16)
+        self.assertEqual(params["input_offset"].dtype, torch.int8)
+        self.assertEqual(params["input_scale"].shape, (1,))
+        self.assertEqual(params["input_offset"].shape, (1,))
+
+    def test_get_perchannel_param_310(self):
+        params = self.method.get_perchannel_param(10, torch.float16)
+
+        self.assertEqual(params["quant_bias"].dtype, torch.int32)
+        self.assertEqual(params["deq_scale"].dtype, torch.int64)
+        self.assertEqual(params["quant_bias"].shape, (10,))
+        self.assertEqual(params["deq_scale"].shape, (10,))
+
+    @pytest.mark.skip("Skip as npu_matmul_compress_dequant will be supported in PTA 26.0.0.")
+    @patch("torch.ops.vllm.quantize")
+    @patch("torch_npu.npu_matmul_compress_dequant")
+    def test_apply_with_x_not_int8_310(self, mock_matmul_compress_dequant, mock_quantize):
+        layer = MagicMock()
+        layer.aclnn_input_scale = torch.randn(256)
+        layer.aclnn_input_scale_reciprocal = 1.0 / layer.aclnn_input_scale
+        layer.aclnn_input_offset = torch.randint(-128, 127, (256,), dtype=torch.int8)
+        layer.weight = torch.randint(-128, 127, (256 * 128,), dtype=torch.int8)
+        layer.index = torch.randint(-128, 127, (8,), dtype=torch.int8)
+        layer.deq_scale = torch.randn(128)
+        layer.quant_bias = torch.randint(-128, 127, (256,))
+        layer.params_dtype = torch.float16
+
+        x = torch.randn(32, 128)
+        expect_x_output = torch.randint(-128, 127, x.shape, dtype=torch.int8)
+        mock_quantize.return_value = expect_x_output
+
+        expected_y_output = torch.randn(32, 256)
+        mock_matmul_compress_dequant.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, tp_rank=0)
+
+        mock_quantize.assert_called_with(
+            x, layer.aclnn_input_scale, layer.aclnn_input_scale_reciprocal, layer.aclnn_input_offset
+        )
+        mock_matmul_compress_dequant.assert_called_with(
+            expect_x_output, layer.weight, layer.index, layer.quant_bias, layer.deq_scale
+        )
+        self.assertTrue(torch.equal(output, expected_y_output))
+
+    @pytest.mark.skip("Skip as npu_matmul_compress_dequant will be supported in PTA 26.0.0.")
+    @patch("torch.ops.vllm.quantize")
+    @patch("torch_npu.npu_matmul_compress_dequant")
+    def test_apply_with_x_is_int8_310(self, mock_matmul_compress_dequant, mock_quantize):
+        layer = MagicMock()
+        layer.aclnn_input_scale = torch.randn(256)
+        layer.aclnn_input_offset = torch.randint(-128, 127, (256,), dtype=torch.int8)
+        layer.weight = torch.randint(-128, 127, (256 * 128,), dtype=torch.int8)
+        layer.index = torch.randint(-128, 127, (8,), dtype=torch.int8)
+        layer.deq_scale = torch.randn(128)
+        layer.quant_bias = torch.randint(-128, 127, (256,))
+        layer.params_dtype = torch.float16
+
+        x = torch.randint(-128, 127, (32, 128), dtype=torch.int8)
+
+        expected_y_output = torch.randn(32, 256)
+        mock_matmul_compress_dequant.return_value = expected_y_output
+
+        output = self.method.apply(layer, x, tp_rank=0)
+
+        mock_quantize.assert_not_called()
+        mock_matmul_compress_dequant.assert_called_with(x, layer.weight, layer.index, layer.quant_bias, layer.deq_scale)
+        self.assertTrue(torch.equal(output, expected_y_output))