[NPU] update nightly tests (#17952)

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: cy <chenyang08056032@163.com>

test/registered/ascend/vlm_models/test_ascend_deepseek_vl2.py

@@ -1,13 +1,20 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import DEEPSEEK_VL2_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/deepseek-ai/deepseek-vl2"
+class TestDeepseekVl2(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the deepseek-ai/deepseek-vl2 model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] deepseek-ai/deepseek-vl2
+    """
+
+    model = DEEPSEEK_VL2_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_gemma_3_4b_it.py

@@ -1,13 +1,20 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import GEMMA_3_4B_IT_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/google/gemma-3-4b-it"
+class TestGemma34bModels(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the google/gemma-3-4b-it model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] google/gemma-3-4b-it
+    """
+
+    model = GEMMA_3_4B_IT_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_janus_pro_1b.py

@@ -1,13 +1,20 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import JANUS_PRO_1B_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/deepseek-ai/Janus-Pro-1B"
+class TestJanusPro1B(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the deepseek-ai/Janus-Pro-1B model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] deepseek-ai/Janus-Pro-1B
+    """
+
+    model = JANUS_PRO_1B_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_janus_pro_7b.py

@@ -1,5 +1,6 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import JANUS_PRO_7B_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -7,7 +8,13 @@ register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
 class TestJanusPro7B(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/deepseek-ai/Janus-Pro-7B"
+    """Testcase: Verify that the inference accuracy of the deepseek-ai/Janus-Pro-7B model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] deepseek-ai/Janus-Pro-7B
+    """
+
+    model = JANUS_PRO_7B_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_llama_3_2_11b_vision_instruct.py

@@ -3,7 +3,12 @@ import unittest
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(
+    est_time=400,
+    suite="nightly-1-npu-a3",
+    nightly=True,
+    disabled="run failed",
+)
 
 
 class TestLlama3211BVisionInstruct(TestVLMModels):

test/registered/ascend/vlm_models/test_ascend_mimo_vl_7b_rl.py

@@ -1,13 +1,20 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import MIMO_VL_7B_RL_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/XiaomiMiMo/MiMo-VL-7B-RL"
+class TestMiMoModels(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the XiaomiMiMo/MiMo-VL-7B-RL model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] XiaomiMiMo/MiMo-VL-7B-RL
+    """
+
+    model = MIMO_VL_7B_RL_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_minicpm_o_2_6.py

@@ -1,13 +1,25 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import MINICPM_O_2_6_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
-register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
+register_npu_ci(
+    est_time=400,
+    suite="nightly-4-npu-a3",
+    nightly=True,
+    disabled="run failed",
+)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/openbmb/MiniCPM-o-2_6"
+class TestMiniCPMModelsO(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the openbmb/MiniCPM-o-2_6 model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] openbmb/MiniCPM-o-2_6
+    """
+
+    model = MINICPM_O_2_6_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_minicpm_v_2_6.py

@@ -1,13 +1,20 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import MINICPM_V_2_6_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/openbmb/MiniCPM-V-2_6"
+class TestMiniCPMModelsV(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the openbmb/MiniCPM-V-2_6 model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] openbmb/MiniCPM-V-2_6
+    """
+
+    model = MINICPM_V_2_6_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_mistral_small_3_1_24b_instruct_2503.py

@@ -0,0 +1,27 @@
+import unittest
+
+from sglang.test.ascend.test_ascend_utils import (
+    MISTRAL_SMALL_3_1_24B_INSTRUCT_2503_WEIGHTS_PATH,
+)
+from sglang.test.ascend.vlm_utils import TestVLMModels
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
+
+
+class TestMistralModels(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the mistralai/Mistral-Small-3.1-24B-Instruct-2503 model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] mistralai/Mistral-Small-3.1-24B-Instruct-2503
+    """
+
+    model = MISTRAL_SMALL_3_1_24B_INSTRUCT_2503_WEIGHTS_PATH
+    mmmu_accuracy = 0.2
+
+    def test_vlm_mmmu_benchmark(self):
+        self._run_vlm_mmmu_test()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/vlm_models/test_ascend_phi4_multimodal_instruct.py

@@ -1,13 +1,20 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import PHI_4_MULTIMODAL_INSTRUCT_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/microsoft/Phi-4-multimodal-instruct"
+class TestPhi4Multimodal(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the microsoft/Phi-4-multimodal-instruct model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] microsoft/Phi-4-multimodal-instruct
+    """
+
+    model = PHI_4_MULTIMODAL_INSTRUCT_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_qwen2_5_vl_3b_instruct.py

@@ -1,13 +1,20 @@
 import unittest
 
+from sglang.test.ascend.test_ascend_utils import QWEN2_5_VL_3B_INSTRUCT_WEIGHTS_PATH
 from sglang.test.ascend.vlm_utils import TestVLMModels
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
 
 
-class TestGemmaModels(TestVLMModels):
-    model = "/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-VL-3B-Instruct"
+class TestQwen25VL3B(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen2.5-VL-3B-Instruct model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen2.5-VL-3B-Instruct
+    """
+
+    model = QWEN2_5_VL_3B_INSTRUCT_WEIGHTS_PATH
     mmmu_accuracy = 0.2
 
     def test_vlm_mmmu_benchmark(self):

test/registered/ascend/vlm_models/test_ascend_qwen2_5_vl_72b_instruct.py

@@ -0,0 +1,40 @@
+import unittest
+
+from sglang.test.ascend.test_ascend_utils import QWEN2_5_VL_72B_INSTRUCT_WEIGHTS_PATH
+from sglang.test.ascend.vlm_utils import TestVLMModels
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(est_time=400, suite="nightly-8-npu-a3", nightly=True)
+
+
+class TestQwen25VL72B(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen2.5-VL-72B-Instruct model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen2.5-VL-72B-Instruct
+    """
+
+    model = QWEN2_5_VL_72B_INSTRUCT_WEIGHTS_PATH
+    mmmu_accuracy = 0.2
+    other_args = [
+        "--trust-remote-code",
+        "--cuda-graph-max-bs",
+        "32",
+        "--enable-multimodal",
+        "--mem-fraction-static",
+        0.6,
+        "--log-level",
+        "info",
+        "--attention-backend",
+        "ascend",
+        "--disable-cuda-graph",
+        "--tp-size",
+        8,
+    ]
+
+    def test_vlm_mmmu_benchmark(self):
+        self._run_vlm_mmmu_test()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/vlm_models/test_ascend_qwen3_vl_235b_a22b_instruct.py

@@ -0,0 +1,41 @@
+import unittest
+
+from sglang.test.ascend.test_ascend_utils import (
+    QWEN3_VL_235B_A22B_INSTRUCT_WEIGHTS_PATH,
+)
+from sglang.test.ascend.vlm_utils import TestVLMModels
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
+
+
+class TestQwen3VL235BA22B(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen3-VL-235B-A22B-Instruct model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen3-VL-235B-A22B-Instruct
+    """
+
+    model = QWEN3_VL_235B_A22B_INSTRUCT_WEIGHTS_PATH
+    mmmu_accuracy = 0.2
+    other_args = [
+        "--trust-remote-code",
+        "--cuda-graph-max-bs",
+        "32",
+        "--enable-multimodal",
+        "--mem-fraction-static",
+        0.8,
+        "--attention-backend",
+        "ascend",
+        "--disable-cuda-graph",
+        "--tp-size",
+        16,
+    ]
+    timeout_for_server_launch = 3000
+
+    def test_vlm_mmmu_benchmark(self):
+        self._run_vlm_mmmu_test()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/vlm_models/test_ascend_qwen3_vl_30b_a3b_instruct.py

@@ -0,0 +1,25 @@
+import unittest
+
+from sglang.test.ascend.test_ascend_utils import QWEN3_VL_30B_A3B_INSTRUCT_WEIGHTS_PATH
+from sglang.test.ascend.vlm_utils import TestVLMModels
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
+
+
+class TestQwen3VL30BA3B(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen3-VL-30B-A3B-Instruct model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen3-VL-30B-A3B-Instruct
+    """
+
+    model = QWEN3_VL_30B_A3B_INSTRUCT_WEIGHTS_PATH
+    mmmu_accuracy = 0.2
+
+    def test_vlm_mmmu_benchmark(self):
+        self._run_vlm_mmmu_test()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/vlm_models/test_ascend_qwen3_vl_4b_instruct.py

@@ -0,0 +1,25 @@
+import unittest
+
+from sglang.test.ascend.test_ascend_utils import QWEN3_VL_4B_INSTRUCT_WEIGHTS_PATH
+from sglang.test.ascend.vlm_utils import TestVLMModels
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
+
+
+class TestQwen3VL4B(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen3-VL-4B-Instruct model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen3-VL-4B-Instruct
+    """
+
+    model = QWEN3_VL_4B_INSTRUCT_WEIGHTS_PATH
+    mmmu_accuracy = 0.2
+
+    def test_vlm_mmmu_benchmark(self):
+        self._run_vlm_mmmu_test()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/vlm_models/test_ascend_qwen3_vl_8b_instruct.py

@@ -0,0 +1,25 @@
+import unittest
+
+from sglang.test.ascend.test_ascend_utils import QWEN3_VL_8B_INSTRUCT_WEIGHTS_PATH
+from sglang.test.ascend.vlm_utils import TestVLMModels
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
+
+
+class TestQwen3VL8B(TestVLMModels):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen3-VL-8B-Instruct model on the MMMU dataset is no less than 0.2.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen3-VL-8B-Instruct
+    """
+
+    model = QWEN3_VL_8B_INSTRUCT_WEIGHTS_PATH
+    mmmu_accuracy = 0.2
+
+    def test_vlm_mmmu_benchmark(self):
+        self._run_vlm_mmmu_test()
+
+
+if __name__ == "__main__":
+    unittest.main()