[Refactor] Create torchair module

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

Author: wangxiyuan

tests/ut/__init__.py

@@ -523,7 +523,7 @@ def test_get_attn_backend_cls_use_v1_and_torchair(self,
         )
         self.assertEqual(
             result,
-            "vllm_ascend.attention.attention_v1_torchair.AscendAttentionTorchairBackend"
+            "vllm_ascend.torchair.attention_torchair.AscendAttentionTorchairBackend"
         )
 
     @patch('vllm_ascend.platform.get_ascend_config')

tests/ut/test_platform.py

@@ -280,27 +280,6 @@ def test_update_aclgraph_sizes(self):
             3,
             len(test_vllm_config.compilation_config.cudagraph_capture_sizes))
 
-    def test_get_torchair_current_work_dir(self):
-        cache_dir = utils.TORCHAIR_CACHE_DIR
-        work_dir = utils.get_torchair_current_work_dir()
-        self.assertEqual(cache_dir, work_dir)
-        work_dir = utils.get_torchair_current_work_dir("test")
-        self.assertEqual(os.path.join(cache_dir, "test"), work_dir)
-
-    def test_torchair_cache_dir(self):
-        utils.write_kv_cache_bytes_to_file(0, 100)
-        self.assertTrue(utils.check_torchair_cache_exist(),
-                        "Create torchair cache dir failed")
-        self.assertTrue(utils.check_kv_cache_bytes_cache_exist(),
-                        "Create kv cache bytes cache dir failed")
-        kv_cache_bytes = utils.read_kv_cache_bytes_from_file(0)
-        self.assertEqual(100, kv_cache_bytes)
-        utils.delete_torchair_cache_file()
-        self.assertFalse(utils.check_torchair_cache_exist(),
-                         "Delete torchair cache dir failed")
-        self.assertFalse(utils.check_kv_cache_bytes_cache_exist(),
-                         "Delete kv cache bytes cache dir failed")
-
 
 class TestProfileExecuteDuration(unittest.TestCase):
 

tests/ut/test_utils.py

@@ -0,0 +1,28 @@
+import os
+
+from tests.ut.base import TestBase
+from vllm_ascend.torchair import utils
+
+
+class TestTorchairUtils(TestBase):
+
+    def test_get_torchair_current_work_dir(self):
+        cache_dir = utils.TORCHAIR_CACHE_DIR
+        work_dir = utils.get_torchair_current_work_dir()
+        self.assertEqual(cache_dir, work_dir)
+        work_dir = utils.get_torchair_current_work_dir("test")
+        self.assertEqual(os.path.join(cache_dir, "test"), work_dir)
+
+    def test_torchair_cache_dir(self):
+        utils.write_kv_cache_bytes_to_file(0, 100)
+        self.assertTrue(utils.check_torchair_cache_exist(),
+                        "Create torchair cache dir failed")
+        self.assertTrue(utils.check_kv_cache_bytes_cache_exist(),
+                        "Create kv cache bytes cache dir failed")
+        kv_cache_bytes = utils.read_kv_cache_bytes_from_file(0)
+        self.assertEqual(100, kv_cache_bytes)
+        utils.delete_torchair_cache_file()
+        self.assertFalse(utils.check_torchair_cache_exist(),
+                         "Delete torchair cache dir failed")
+        self.assertFalse(utils.check_kv_cache_bytes_cache_exist(),
+                         "Delete kv cache bytes cache dir failed")

tests/ut/torchair/__init__.py

@@ -1,3 +1,4 @@
+import os
 import unittest
 from unittest.mock import MagicMock, patch
 
@@ -24,9 +25,10 @@ def _create_model_runner(self, model: str, *args,
     def setUp(self):
         """Initialize test fixtures and common mocks"""
         self.attn_backend = "npu"
-
+        model_path = os.path.join(os.path.dirname(__file__), "..",
+                                  "fake_weight")
         model_runner = self._create_model_runner(
-            "tests/ut/fake_weight",
+            model_path,
             trust_remote_code=True,
             enable_chunked_prefill=False,
         )

tests/ut/torchair/test_utils.py

@@ -181,7 +181,10 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
 
         if parallel_config and parallel_config.worker_cls == "auto":
             if envs.VLLM_USE_V1:
-                parallel_config.worker_cls = "vllm_ascend.worker.worker_v1.NPUWorker"
+                if ascend_config.torchair_graph_config.enabled:
+                    parallel_config.worker_cls = "vllm_ascend.torchair.worker_v1_torchair.NPUTorchairWorker"
+                else:
+                    parallel_config.worker_cls = "vllm_ascend.worker.worker_v1.NPUWorker"
             elif vllm_config.speculative_config:
                 # NOTE: We set this var to `1` in vllm-ascend to avoid segment
                 # fault when using spec decode with V0 engine.
@@ -224,7 +227,7 @@ def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
             return "vllm_ascend.attention.mla_v1.AscendMLABackend"
         use_torchair = get_ascend_config().torchair_graph_config.enabled
         if use_v1 and use_torchair:
-            return "vllm_ascend.attention.attention_v1_torchair.AscendAttentionTorchairBackend"
+            return "vllm_ascend.torchair.attention_torchair.AscendAttentionTorchairBackend"
         if use_v1:
             return "vllm_ascend.attention.attention_v1.AscendAttentionBackend"
         if use_mla: