[Test] Add ut test for torchair (#4287)

### What this PR does / why we need it?
The current community lacks unit tests (UT) for files such as
torchair_worker, mtp_proposer, and model_runner. Therefore, UT coverage
for these files needs to be added.

### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

---------

Signed-off-by: CodeNine-CJ <chenjian343@huawei.com>

tests/ut/torchair/test_torchair_model_runner.py

@@ -0,0 +1,45 @@
+from unittest.mock import MagicMock, Mock
+
+import pytest
+import torch
+from pytest_mock import MockerFixture
+from vllm.config import VllmConfig
+
+from tests.ut.base import PytestBase
+from vllm_ascend.torchair.torchair_model_runner import NPUTorchairModelRunner
+
+
+class TestNPUTorchairModelRunner(PytestBase):
+
+    @pytest.fixture
+    def setup_npu_torchair_model_runner(self, mocker: MockerFixture):
+        mocker.patch.object(NPUTorchairModelRunner, "__init__",
+                            lambda self, *args, **kwargs: None)
+        runner = NPUTorchairModelRunner(Mock(), Mock())
+
+        runner.device = torch.device("cpu")
+        runner.vllm_config = MagicMock(spec=VllmConfig)
+
+        runner.speculative_config = MagicMock(
+            method="deepseek_mtp",
+            num_speculative_tokens=4,
+            disable_padded_drafter_batch=False)
+
+        runner.ascend_config = MagicMock(enable_shared_expert_dp=False,
+                                         torchair_graph_config=MagicMock(
+                                             use_cached_graph=True,
+                                             graph_batch_sizes=[1, 2, 4]))
+
+        runner.decode_token_per_req = 2
+        runner.is_kv_consumer = True
+        runner.max_num_reqs = 100
+
+        runner.model_config = MagicMock(hf_config=MagicMock(index_topk=2))
+        runner.attn_backend = MagicMock(get_builder_cls=lambda: Mock())
+
+        return runner
+
+    def test_init(self, mocker: MockerFixture,
+                  setup_npu_torchair_model_runner):
+        runner = setup_npu_torchair_model_runner
+        assert isinstance(runner, NPUTorchairModelRunner)

tests/ut/torchair/test_torchair_mtp_proposer.py

@@ -0,0 +1,85 @@
+from unittest.mock import MagicMock, Mock
+
+import pytest
+import torch
+from pytest_mock import MockerFixture
+from vllm.config import CacheConfig, VllmConfig
+
+from tests.ut.base import PytestBase
+from vllm_ascend.torchair.torchair_mtp_proposer import TorchairMtpProposer
+from vllm_ascend.utils import vllm_version_is
+
+
+class TestTorchairMtpProposer(PytestBase):
+
+    @pytest.fixture
+    def setup_torchair_mtp_proposer(self, mocker: MockerFixture):
+        vllm_config = MagicMock(spec=VllmConfig)
+        vllm_config.device_config = MagicMock()
+        vllm_config.device_config.device = torch.device("cpu")
+        vllm_config.speculative_config = MagicMock()
+        vllm_config.speculative_config.draft_model_config = MagicMock()
+        vllm_config.speculative_config.draft_model_config.dtype = torch.float16
+        vllm_config.speculative_config.method = "deepseek_mtp"
+        vllm_config.speculative_config.num_speculative_tokens = 5
+        vllm_config.load_config = MagicMock()
+        cache_config = CacheConfig(block_size=16)
+        vllm_config.cache_config = cache_config
+        vllm_config.scheduler_config = MagicMock(max_num_batched_tokens=1024,
+                                                 max_num_seqs=64)
+
+        device = torch.device("cpu")
+        runner = MagicMock()
+        runner.pcp_size = 1
+        runner.dcp_size = 1
+        runner.pcp_rank = 0
+        runner.max_num_tokens = 1024
+        runner.max_num_reqs = 10
+        runner._use_aclgraph.return_value = True
+
+        mocker.patch(
+            "vllm_ascend.torchair.torchair_mtp_proposer.MtpProposer.__init__",
+            return_value=None)
+
+        if vllm_version_is("0.11.0"):
+            mock_set_default_dtype = mocker.patch(
+                'vllm.model_executor.model_loader.utils.set_default_torch_dtype'
+            )
+        else:
+            mock_set_default_dtype = mocker.patch(
+                'vllm.utils.torch_utils.set_default_torch_dtype')
+        mock_set_default_dtype.return_value.__enter__.return_value = None
+
+        mock_model_loader = MagicMock()
+        mocker.patch("vllm.model_executor.model_loader.get_model_loader",
+                     return_value=mock_model_loader)
+        mock_layers = {
+            "target_attn_layer_1": Mock(),
+            "draft_attn_layer_2": Mock()
+        }
+        mocker.patch("vllm.config.get_layers_from_vllm_config",
+                     return_value=mock_layers)
+        mock_set_current = mocker.patch("vllm.config.set_current_vllm_config")
+        mock_set_current.return_value.__enter__.return_value = None
+        mock_torchair_deepseek_mtp = MagicMock()
+        mock_torchair_deepseek_mtp.to.return_value = mock_torchair_deepseek_mtp
+        mocker.patch(
+            "vllm_ascend.torchair.models.torchair_deepseek_mtp.TorchairDeepSeekMTP",
+            return_value=mock_torchair_deepseek_mtp)
+        mocker.patch(
+            "vllm.model_executor.model_loader.utils.process_weights_after_loading"
+        )
+
+        proposer = TorchairMtpProposer(vllm_config, device, runner)
+        proposer.vllm_config = vllm_config
+        proposer.device = device
+        proposer.runner = runner
+        proposer.speculative_config = vllm_config.speculative_config
+        proposer.draft_model_config = vllm_config.speculative_config.draft_model_config
+        proposer.method = vllm_config.speculative_config.method
+
+        return proposer, mock_model_loader, mock_torchair_deepseek_mtp
+
+    def test_init(self, setup_torchair_mtp_proposer):
+        proposer, _, _, = setup_torchair_mtp_proposer
+        assert isinstance(proposer, TorchairMtpProposer)

tests/ut/torchair/test_torchair_worker.py

@@ -0,0 +1,107 @@
+from unittest.mock import MagicMock, patch
+
+import torch
+from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig
+
+from tests.ut.base import TestBase
+from vllm_ascend.utils import vllm_version_is
+
+init_cache_hf_modules_path = "vllm.utils.init_cached_hf_modules" if vllm_version_is(
+    "0.11.0") else "vllm.utils.import_utils.init_cached_hf_modules"
+
+
+class TestNPUTorchairWorker(TestBase):
+
+    def setUp(self):
+        self.cache_config_mock = MagicMock(spec=CacheConfig)
+        self.cache_config_mock.cache_type = "auto"
+
+        self.model_config_mock = MagicMock(spec=ModelConfig)
+        self.model_config_mock.dtype = torch.float16
+        self.model_config_mock.trust_remote_code = False
+
+        self.hf_config_mock = MagicMock()
+        self.hf_config_mock.model_type = "test_model"
+        if hasattr(self.hf_config_mock, 'index_topk'):
+            delattr(self.hf_config_mock, 'index_topk')
+
+        self.model_config_mock.hf_config = self.hf_config_mock
+
+        self.parallel_config_mock = MagicMock(spec=ParallelConfig)
+
+        self.vllm_config_mock = MagicMock(spec=VllmConfig)
+        self.vllm_config_mock.cache_config = self.cache_config_mock
+        self.vllm_config_mock.model_config = self.model_config_mock
+        self.vllm_config_mock.parallel_config = self.parallel_config_mock
+        self.vllm_config_mock.additional_config = None
+        self.vllm_config_mock.load_config = None
+        self.vllm_config_mock.scheduler_config = None
+        self.vllm_config_mock.device_config = None
+        self.vllm_config_mock.compilation_config = None
+
+        self.local_rank = 0
+        self.rank = 0
+        self.distributed_init_method = "tcp://localhost:12345"
+        self.is_driver_worker = False
+
+    @patch(
+        "vllm_ascend.worker.worker_v1.NPUWorker._init_worker_distributed_environment"
+    )
+    @patch("vllm_ascend.worker.worker_v1.NPUPlatform")
+    def test_init_device(self, mock_platform, mock_init_dist_env):
+        from vllm_ascend.worker.worker_v1 import NPUWorker
+
+        mock_platform.mem_get_info.return_value = (1000, 2000)
+
+        with patch.object(NPUWorker, "__init__", lambda x, **kwargs: None):
+            worker = NPUWorker()
+            worker.local_rank = 1
+            worker.model_config = MagicMock()
+            worker.model_config.seed = 42
+            worker.vllm_config = MagicMock()
+
+            result = worker._init_device()
+
+            mock_platform.set_device.assert_called_once()
+            call_args = mock_platform.set_device.call_args[0][0]
+            self.assertEqual(str(call_args), "npu:1")
+
+            mock_platform.empty_cache.assert_called_once()
+            mock_platform.seed_everything.assert_called_once_with(42)
+            mock_platform.mem_get_info.assert_called_once()
+            mock_init_dist_env.assert_called_once()
+
+            self.assertEqual(str(result), "npu:1")
+            self.assertEqual(worker.init_npu_memory, 1000)
+
+    @patch(
+        "vllm_ascend.worker.worker_v1.NPUWorker._init_worker_distributed_environment"
+    )
+    @patch("vllm_ascend.worker.worker_v1.NPUPlatform")
+    def test_init_device_torchair_worker(self, mock_platform,
+                                         mock_init_dist_env):
+        from vllm_ascend.torchair.torchair_worker import NPUTorchairWorker
+
+        mock_platform.mem_get_info.return_value = (1000, 2000)
+
+        with patch.object(NPUTorchairWorker, "__init__",
+                          lambda x, **kwargs: None):
+            worker = NPUTorchairWorker()
+            worker.local_rank = 1
+            worker.model_config = MagicMock()
+            worker.model_config.seed = 42
+            worker.vllm_config = MagicMock()
+
+            result = worker._init_device()
+
+            mock_platform.set_device.assert_called_once()
+            call_args = mock_platform.set_device.call_args[0][0]
+            self.assertEqual(str(call_args), "npu:1")
+
+            mock_platform.empty_cache.assert_called_once()
+            mock_platform.seed_everything.assert_called_once_with(42)
+            mock_platform.mem_get_info.assert_called_once()
+            mock_init_dist_env.assert_called_once()
+
+            self.assertEqual(str(result), "npu:1")
+            self.assertEqual(worker.init_npu_memory, 1000)