[CI] speed up ut (#4901)

avoid model download to speed up ut test. 

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

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

tests/e2e/singlecard/test_aclgraph_accuracy.py

@@ -192,3 +192,22 @@ def test_output_between_eager_and_full_decode_only(
         name_0="vllm_eager_outputs",
         name_1="vllm_aclgraph_outputs",
     )
+
+
+def test_aclgraph_enable():
+    # Generally, this test is not belong to e2e, but it is a good way to check if
+    # aclgraph is enabled in real environment
+    from vllm.config.compilation import CompilationMode, CUDAGraphMode
+    from vllm.engine.arg_utils import EngineArgs
+
+    from vllm_ascend.platform import NPUPlatform
+
+    # vLLM default mode is piecewise cudagraph
+    config = EngineArgs()
+    VllmConfig = config.create_engine_config()
+    assert VllmConfig.compilation_config.cudagraph_mode == CUDAGraphMode.PIECEWISE
+
+    # after check_and_update_config, mode should be VLLM_COMPILE and piecewise cudagraph
+    NPUPlatform.check_and_update_config(VllmConfig)
+    assert VllmConfig.compilation_config.mode == CompilationMode.VLLM_COMPILE
+    assert VllmConfig.compilation_config.cudagraph_mode == CUDAGraphMode.PIECEWISE

tests/ut/attention/test_mla_v1.py

@@ -1,3 +1,4 @@
+import os
 from unittest.mock import MagicMock, patch
 
 import torch
@@ -507,8 +508,6 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
 
     def setUp(self):
         self.mock_vllm_config = MagicMock(spec=VllmConfig)
-        self.mock_vllm_config.model_config = ModelConfig(max_model_len=2048)
-        self.mock_vllm_config.model_config.hf_text_config.qk_rope_head_dim = 32
         self.mock_vllm_config.cache_config = CacheConfig(block_size=32)
         mock_scheduler_config = MagicMock(spec=SchedulerConfig)
         mock_scheduler_config.max_num_seqs = 8
@@ -516,7 +515,15 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
         self.mock_vllm_config.scheduler_config = mock_scheduler_config
         self.mock_vllm_config.speculative_config = None
         self.mock_device = torch.device("cpu")
-
+        fake_weight_path = os.path.join(os.path.dirname(__file__), "..",
+                                        "fake_weight")
+        model_config = ModelConfig(
+            model=fake_weight_path,
+            skip_tokenizer_init=True,
+        )
+        model_config.hf_text_config.head_dim = 128
+        model_config.hf_text_config.qk_rope_head_dim = 32
+        self.mock_vllm_config.model_config = model_config
         self.kv_cache_spec = MagicMock()
         self.kv_cache_spec.num_layers = 32
         self.kv_cache_spec.head_size = 128

tests/ut/test_platform.py

@@ -4,7 +4,6 @@ from unittest.mock import MagicMock, patch
 import pytest
 import torch
 from vllm.config.compilation import CompilationMode, CUDAGraphMode
-from vllm.engine.arg_utils import EngineArgs
 from vllm.platforms import PlatformEnum
 
 from tests.ut.base import TestBase
@@ -579,30 +578,3 @@ class TestNPUPlatform(TestBase):
             self.platform.get_static_graph_wrapper_cls(),
             "vllm_ascend.compilation.acl_graph.ACLGraphWrapper",
         )
-
-    def test_aclgraph_enable(self):
-        config = EngineArgs()
-        VllmConfig = config.create_engine_config()
-        self.assertEqual(VllmConfig.compilation_config.cudagraph_mode,
-                         CUDAGraphMode.PIECEWISE)
-
-        with self.assertLogs(logger="vllm", level="INFO") as cm:
-            from vllm_ascend import platform
-
-            importlib.reload(platform)
-            self.platform.check_and_update_config(VllmConfig)
-            target_msg = "PIECEWISE compilation enabled on NPU. use_inductor not supported - using only ACL Graph mode"
-            found = any(target_msg in log for log in cm.output)
-
-            self.assertTrue(
-                found,
-                f"Expected log message not found. Captured logs: {cm.output}")
-
-            self.assertEqual(
-                VllmConfig.compilation_config.mode,
-                CompilationMode.VLLM_COMPILE,
-            )
-            self.assertEqual(
-                VllmConfig.compilation_config.cudagraph_mode,
-                CUDAGraphMode.PIECEWISE,
-            )

vllm_ascend/eplb/core/policy/policy_dynamic_ep_v2.py

@@ -4,6 +4,7 @@ from abc import abstractmethod
 from collections import defaultdict
 
 import numpy as np
+from vllm.logger import logger
 
 
 class DynamicConfig:
@@ -69,21 +70,21 @@ class DynamicEplbV2(EplbPolicy):
     @staticmethod
     def safe_divide(a, b):
         if b == 0:
-            print("Division by zero is not allowed")
+            logger.info("Division by zero is not allowed")
             return 0
         return a / b
 
     @staticmethod
     def safe_exact_divide(a, b):
         if b == 0:
-            print("Division by zero is not allowed")
+            logger.info("Division by zero is not allowed")
             return 0
         return a // b
 
     @staticmethod
     def safe_mod(a, b):
         if b == 0:
-            print("Division by zero is not allowed")
+            logger.info("Division by zero is not allowed")
             return 0
         return a % b
 

vllm_ascend/model_loader/netloader/interaction/elastic.py

@@ -60,7 +60,7 @@ class ElasticClient:
                 ip, port_str = source.split(':')
                 port = int(port_str)
             except Exception as e:
-                logger.error(f"IP format error: {source}, detail: {e}")
+                logger.info(f"IP format error: {source}, detail: {e}")
                 continue
 
             self.server_addr = ip

vllm_ascend/model_loader/netloader/load.py

@@ -80,5 +80,5 @@ def elastic_load(
                 time.perf_counter() - t0))
             return model_loaded
     except Exception as e:
-        logger.error(f"elastic_load error: {e}")
+        logger.info(f"elastic_load error: {e}")
         return None