[CI/UT] Fix UTs on register customop and warm up model (#2862)

### What this PR does / why we need it?
Fix UTs on register customop and warm up model

### How was this patch tested?
CI passed with existing test.

Co-authored-by: Icey <1790571317@qq.com>

- vLLM version: main
- vLLM main:
cc99baf14d

---------

Signed-off-by: MengqingCao <cmq0113@163.com>

tests/ut/test_utils.py

@@ -24,6 +24,7 @@ from vllm.config import (CompilationConfig, ModelConfig, ParallelConfig,
 
 from tests.ut.base import TestBase
 from vllm_ascend import utils
+from vllm_ascend.utils import REGISTERED_ASCEND_OPS
 
 
 class TestUtils(TestBase):
@@ -302,14 +303,14 @@ class TestUtils(TestBase):
 
         # ascend custom op is not registered
         utils.register_ascend_customop()
-        # should call register_oot three
+        self.assertEqual(mock_customop.register_oot.call_count,
-        self.assertEqual(mock_customop.register_oot.call_count, 13)
+                         len(REGISTERED_ASCEND_OPS))
         self.assertTrue(utils._ASCEND_CUSTOMOP_IS_REIGISTERED)
 
         # ascend custom op is already registered
         utils.register_ascend_customop()
-        # should not register_oot again, thus only called three in this ut
+        self.assertEqual(mock_customop.register_oot.call_count,
-        self.assertEqual(mock_customop.register_oot.call_count, 13)
+                         len(REGISTERED_ASCEND_OPS))
 
 
 class TestProfileExecuteDuration(TestBase):

tests/ut/worker/test_worker_v1.py

@@ -1009,9 +1009,8 @@ class TestNPUWorker(TestBase):
 
     @patch("vllm_ascend.worker.worker_v1.NPUPlatform.seed_everything")
     @patch("vllm_ascend.worker.worker_v1.logger")
-    @patch("torch_npu._npu_matmul_add_fp32")
+    @patch("vllm_ascend.worker.worker_v1.NPUWorker._warm_up_atb")
-    def test_compile_or_warm_up_model_with_eager_mode(self,
+    def test_compile_or_warm_up_model_with_eager_mode(self, mock_warm_up_atb,
-                                                      mock_npu_matmul_add,
                                                       mock_logger,
                                                       mock_seed_everything):
         """Test compile_or_warm_up_model method - eager mode"""
@@ -1054,14 +1053,14 @@ class TestNPUWorker(TestBase):
             # Verify seed setting
             mock_seed_everything.assert_called_once_with(12345)
 
-            # Verify calls
+            # Verify atb warm up
-            mock_npu_matmul_add.assert_called_once()
+            mock_warm_up_atb.assert_called_once()
 
     @patch("vllm_ascend.worker.worker_v1.NPUPlatform.seed_everything")
     @patch("vllm_ascend.worker.worker_v1.logger")
-    @patch("torch_npu._npu_matmul_add_fp32")
+    @patch("vllm_ascend.worker.worker_v1.NPUWorker._warm_up_atb")
     def test_compile_or_warm_up_model_with_graph_capture(
-            self, mock_npu_matmul_add, mock_logger, mock_seed_everything):
+            self, mock_warm_up_atb, mock_logger, mock_seed_everything):
         """Test compile_or_warm_up_model method - with graph capture enabled"""
         from vllm_ascend.worker.worker_v1 import NPUWorker
 
@@ -1094,8 +1093,8 @@ class TestNPUWorker(TestBase):
             # Verify seed setting
             mock_seed_everything.assert_called_once_with(67890)
 
-            # Verify calls
+            # Verify atb warm up
-            mock_npu_matmul_add.assert_called_once()
+            mock_warm_up_atb.assert_called_once()
 
     @patch("vllm_ascend.worker.worker_v1.CaMemAllocator")
     def test_initialize_from_config_with_sleep_mode(self,

vllm_ascend/utils.py

@@ -50,6 +50,7 @@ MAX_CAPTURE_SIZE = 1800
 
 ASCEND_QUANTIZATION_METHOD = "ascend"
 SOC_VERSION_INFERENCE_SERIES = ["Ascend310P3"]
+REGISTERED_ASCEND_OPS = {}
 
 ACL_FORMAT_FRACTAL_ND = 2
 ACL_FORMAT_FRACTAL_NZ = 29
@@ -493,7 +494,10 @@ def register_ascend_customop():
         return
     from vllm.model_executor.custom_op import CustomOp
 
+    from vllm_ascend.models.layers.mla import AscendMultiHeadLatentAttention
     from vllm_ascend.ops.activation import AscendQuickGELU, AscendSiluAndMul
+    from vllm_ascend.ops.common_fused_moe import AscendFusedMoE
+    from vllm_ascend.ops.layernorm import AscendRMSNorm
     from vllm_ascend.ops.linear import (AscendColumnParallelLinear,
                                         AscendMergedColumnParallelLinear,
                                         AscendQKVParallelLinear,
@@ -503,38 +507,27 @@ def register_ascend_customop():
     from vllm_ascend.ops.vocab_parallel_embedding import (
         AscendLogitsProcessor, AscendParallelLMHead,
         AscendVocabParallelEmbedding)
-    CustomOp.register_oot(_decorated_op_cls=AscendQuickGELU, name="QuickGELU")
-    CustomOp.register_oot(_decorated_op_cls=AscendSiluAndMul,
-                          name="SiluAndMul")
-    CustomOp.register_oot(_decorated_op_cls=AscendRotaryEmbedding,
-                          name="RotaryEmbedding")
-    CustomOp.register_oot(_decorated_op_cls=AscendColumnParallelLinear,
-                          name="ColumnParallelLinear")
-    CustomOp.register_oot(_decorated_op_cls=AscendRowParallelLinear,
-                          name="RowParallelLinear")
-    CustomOp.register_oot(_decorated_op_cls=AscendMergedColumnParallelLinear,
-                          name="MergedColumnParallelLinear")
-    CustomOp.register_oot(_decorated_op_cls=AscendQKVParallelLinear,
-                          name="QKVParallelLinear")
-    CustomOp.register_oot(
-        _decorated_op_cls=AscendDeepseekScalingRotaryEmbedding,
-        name="DeepseekScalingRotaryEmbedding")
-    CustomOp.register_oot(_decorated_op_cls=AscendVocabParallelEmbedding,
-                          name="VocabParallelEmbedding")
-    CustomOp.register_oot(_decorated_op_cls=AscendParallelLMHead,
-                          name="ParallelLMHead")
-    CustomOp.register_oot(_decorated_op_cls=AscendLogitsProcessor,
-                          name="LogitsProcessor")
 
-    from vllm_ascend.ops.layernorm import AscendRMSNorm
+    global REGISTERED_ASCEND_OPS
-    CustomOp.register_oot(_decorated_op_cls=AscendRMSNorm, name="RMSNorm")
+    REGISTERED_ASCEND_OPS = {
+        "QuickGELU": AscendQuickGELU,
+        "SiluAndMul": AscendSiluAndMul,
+        "RotaryEmbedding": AscendRotaryEmbedding,
+        "ColumnParallelLinear": AscendColumnParallelLinear,
+        "RowParallelLinear": AscendRowParallelLinear,
+        "MergedColumnParallelLinear": AscendMergedColumnParallelLinear,
+        "QKVParallelLinear": AscendQKVParallelLinear,
+        "DeepseekScalingRotaryEmbedding": AscendDeepseekScalingRotaryEmbedding,
+        "VocabParallelEmbedding": AscendVocabParallelEmbedding,
+        "ParallelLMHead": AscendParallelLMHead,
+        "LogitsProcessor": AscendLogitsProcessor,
+        "RMSNorm": AscendRMSNorm,
+        "FusedMoE": AscendFusedMoE,
+        "MultiHeadLatentAttention": AscendMultiHeadLatentAttention,
+    }
 
-    from vllm_ascend.ops.common_fused_moe import AscendFusedMoE
+    for name, op_cls in REGISTERED_ASCEND_OPS.items():
-    CustomOp.register_oot(_decorated_op_cls=AscendFusedMoE, name="FusedMoE")
+        CustomOp.register_oot(_decorated_op_cls=op_cls, name=name)
-
-    from vllm_ascend.models.layers.mla import AscendMultiHeadLatentAttention
-    CustomOp.register_oot(_decorated_op_cls=AscendMultiHeadLatentAttention,
-                          name="MultiHeadLatentAttention")
 
     # NOTE: Keep this at last to ensure all custom actions are registered
     _ASCEND_CUSTOMOP_IS_REIGISTERED = True