Apply fixture-kit mode to MMMUVLMMixin (#15615)

test/srt/models/test_ministral3_models.py

@@ -1,9 +1,9 @@
 import unittest
-from types import SimpleNamespace
 
 from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
-from sglang.test.kits.mmmu_vlm_kit import MMMUVLMTestBase
+from sglang.test.kits.mmmu_vlm_kit import MMMUMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
+from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
 
 MODEL = "mistralai/Ministral-3-3B-Instruct-2512"
 
@@ -14,18 +14,13 @@ class TestMinistral3TextOnly(GSM8KMixin, DefaultServerBase):
     other_args = ["--trust-remote-code"]
 
 
-class TestMinistral3MMMU(MMMUVLMTestBase):
+class TestMinistral3MMMU(MMMUMixin, MMMUServerBase):
     accuracy = 0.3
     model = MODEL
     other_args = ["--trust-remote-code"]
     mmmu_args = ["--limit=0.1"]
     """`--limit=0.1`: 10 percent of each task - this is fine for testing since the nominal result isn't interesting - this run is just to prevent relative regressions."""
 
-    def test_vlm_mmmu_benchmark(self):
-        self._run_vlm_mmmu_test(
-            SimpleNamespace(model=self.model, mmmu_accuracy=self.accuracy), "./logs"
-        )
-
 
 if __name__ == "__main__":
     unittest.main()

test/srt/models/test_nvidia_nemotron_nano_v2_vl.py

@@ -1,9 +1,9 @@
 import unittest
-from types import SimpleNamespace
 
 from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
-from sglang.test.kits.mmmu_vlm_kit import MMMUVLMTestBase
+from sglang.test.kits.mmmu_vlm_kit import MMMUMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
+from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
 
 MODEL = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
 
@@ -14,18 +14,13 @@ class TestNvidiaNemotronNanoV2VLTextOnly(GSM8KMixin, DefaultServerBase):
     other_args = ["--max-mamba-cache-size", "256", "--trust-remote-code"]
 
 
-class TestNvidiaNemotronNanoV2VLMMMU(MMMUVLMTestBase):
+class TestNvidiaNemotronNanoV2VLMMMU(MMMUMixin, MMMUServerBase):
     accuracy = 0.444
     model = MODEL
     other_args = ["--max-mamba-cache-size", "128", "--trust-remote-code"]
     mmmu_args = ["--limit=0.1"]
     """`--limit=0.1`: 10 percent of each task - this is fine for testing since the nominal result isn't interesting - this run is just to prevent relative regressions."""
 
-    def test_vlm_mmmu_benchmark(self):
-        self._run_vlm_mmmu_test(
-            SimpleNamespace(model=self.model, mmmu_accuracy=self.accuracy), "./logs"
-        )
-
 
 if __name__ == "__main__":
     unittest.main()

test/srt/models/test_vlm_models.py

@@ -5,7 +5,10 @@ import unittest
 from types import SimpleNamespace
 
 from sglang.srt.utils import is_hip
-from sglang.test.kits.mmmu_vlm_kit import DEFAULT_MEM_FRACTION_STATIC, MMMUVLMTestBase
+from sglang.test.kits.mmmu_vlm_kit import (
+    DEFAULT_MEM_FRACTION_STATIC,
+    MMMUMultiModelTestBase,
+)
 from sglang.test.test_utils import is_in_ci
 
 _is_hip = is_hip()
@@ -20,7 +23,7 @@ else:
     ]
 
 
-class TestVLMModels(MMMUVLMTestBase):
+class TestVLMModels(MMMUMultiModelTestBase):
     def test_vlm_mmmu_benchmark(self):
         """Test VLM models against MMMU benchmark."""
         models_to_test = MODELS