chore: vendor sglang v0.5.10 snapshot

third_party/sglang/test/manual/nightly/test_deepseek_v31_perf.py

@@ -0,0 +1,86 @@
+import unittest
+
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+DEEPSEEK_V31_MODEL_PATH = "deepseek-ai/DeepSeek-V3.1"
+PROFILE_DIR = "performance_profiles_deepseek_v31"
+
+
+class TestNightlyDeepseekV31Performance(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V31_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # Define variant configurations
+        cls.variants = [
+            {
+                "name": "basic",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+            {
+                "name": "mtp",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--speculative-algorithm",
+                    "EAGLE",
+                    "--speculative-num-steps",
+                    "3",
+                    "--speculative-eagle-topk",
+                    "1",
+                    "--speculative-num-draft-tokens",
+                    "4",
+                    "--mem-frac",
+                    "0.7",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+
+    def test_bench_one_batch(self):
+        failed_variants = []
+
+        try:
+            for variant_config in self.variants:
+                with self.subTest(variant=variant_config["name"]):
+                    results, success = self.runner.run_benchmark_for_model(
+                        model_path=self.model,
+                        batch_sizes=self.batch_sizes,
+                        input_lens=self.input_lens,
+                        output_lens=self.output_lens,
+                        other_args=variant_config["other_args"],
+                        variant=variant_config["name"],
+                    )
+
+                    if not success:
+                        failed_variants.append(variant_config["name"])
+
+                    self.runner.add_report(results, variant=variant_config["name"])
+        finally:
+            self.runner.write_final_report()
+
+        if failed_variants:
+            raise AssertionError(
+                f"Benchmark failed for {self.model} with the following variants: "
+                f"{', '.join(failed_variants)}"
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()

third_party/sglang/test/manual/nightly/test_deepseek_v32_perf.py

@@ -0,0 +1,127 @@
+import unittest
+
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
+
+DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
+PROFILE_DIR = "performance_profiles_deepseek_v32"
+
+
+class TestNightlyDeepseekV32Performance(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V32_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # Define variant configurations
+        cls.variants = [
+            {
+                "name": "basic",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--dp",
+                    "8",
+                    "--enable-dp-attention",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+            {
+                "name": "mtp",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--dp",
+                    "8",
+                    "--enable-dp-attention",
+                    "--speculative-algorithm",
+                    "EAGLE",
+                    "--speculative-num-steps",
+                    "3",
+                    "--speculative-eagle-topk",
+                    "1",
+                    "--speculative-num-draft-tokens",
+                    "4",
+                    "--mem-frac",
+                    "0.7",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+            {
+                "name": "nsa",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--dp",
+                    "8",
+                    "--enable-dp-attention",
+                    "--attention-backend",
+                    "nsa",
+                    "--nsa-prefill-backend",
+                    "flashmla_sparse",
+                    "--nsa-decode-backend",
+                    "flashmla_kv",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+            {
+                "name": "pure_tp",
+                "other_args": [
+                    "--trust-remote-code",
+                    "--tp",
+                    "8",
+                    "--attention-backend",
+                    "nsa",
+                    "--nsa-prefill-backend",
+                    "flashmla_sparse",
+                    "--nsa-decode-backend",
+                    "flashmla_kv",
+                    "--model-loader-extra-config",
+                    '{"enable_multithread_load": true}',
+                ],
+            },
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+
+    def test_bench_one_batch(self):
+        failed_variants = []
+
+        try:
+            for variant_config in self.variants:
+                with self.subTest(variant=variant_config["name"]):
+                    results, success = self.runner.run_benchmark_for_model(
+                        model_path=self.model,
+                        batch_sizes=self.batch_sizes,
+                        input_lens=self.input_lens,
+                        output_lens=self.output_lens,
+                        other_args=variant_config["other_args"],
+                        variant=variant_config["name"],
+                    )
+
+                    if not success:
+                        failed_variants.append(variant_config["name"])
+
+                    self.runner.add_report(results, variant=variant_config["name"])
+        finally:
+            self.runner.write_final_report()
+
+        if failed_variants:
+            raise AssertionError(
+                f"Benchmark failed for {self.model} with the following variants: "
+                f"{', '.join(failed_variants)}"
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()

third_party/sglang/test/manual/nightly/test_text_models_gsm8k_eval.py

@@ -0,0 +1,124 @@
+import json
+import unittest
+import warnings
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP1,
+    DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2,
+    DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1,
+    DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP2,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    ModelLaunchSettings,
+    check_evaluation_test_results,
+    parse_models,
+    popen_launch_server,
+    write_results_to_json,
+)
+
+MODEL_SCORE_THRESHOLDS = {
+    "meta-llama/Llama-3.1-8B-Instruct": 0.82,
+    "mistralai/Mistral-7B-Instruct-v0.3": 0.58,
+    "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": 0.85,
+    "google/gemma-2-27b-it": 0.91,
+    "meta-llama/Llama-3.1-70B-Instruct": 0.95,
+    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.616,
+    "Qwen/Qwen2-57B-A14B-Instruct": 0.86,
+    "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.83,
+    "neuralmagic/Mistral-7B-Instruct-v0.3-FP8": 0.54,
+    "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8": 0.835,
+    "zai-org/GLM-4.5-Air-FP8": 0.75,
+    # The threshold of neuralmagic/gemma-2-2b-it-FP8 should be 0.6, but this model has some accuracy regression.
+    # The fix is tracked at https://github.com/sgl-project/sglang/issues/4324, we set it to 0.50, for now, to make CI green.
+    "neuralmagic/gemma-2-2b-it-FP8": 0.50,
+    "neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8": 0.94,
+    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8": 0.65,
+    "neuralmagic/Qwen2-72B-Instruct-FP8": 0.94,
+    "neuralmagic/Qwen2-57B-A14B-Instruct-FP8": 0.82,
+}
+
+
+# Do not use `CustomTestCase` since `test_mgsm_en_all_models` does not want retry
+class TestNightlyGsm8KEval(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.models = []
+        models_tp1 = parse_models(
+            DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1
+        ) + parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP1)
+        for model_path in models_tp1:
+            cls.models.append(ModelLaunchSettings(model_path, tp_size=1))
+
+        models_tp2 = parse_models(
+            DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP2
+        ) + parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2)
+        for model_path in models_tp2:
+            cls.models.append(ModelLaunchSettings(model_path, tp_size=2))
+
+        cls.base_url = DEFAULT_URL_FOR_TEST
+
+    def test_mgsm_en_all_models(self):
+        warnings.filterwarnings(
+            "ignore", category=ResourceWarning, message="unclosed.*socket"
+        )
+        is_first = True
+        all_results = []
+        for model_setup in self.models:
+            with self.subTest(model=model_setup.model_path):
+                other_args = list(model_setup.extra_args)
+
+                if model_setup.model_path == "meta-llama/Llama-3.1-70B-Instruct":
+                    other_args.extend(["--mem-fraction-static", "0.9"])
+
+                process = popen_launch_server(
+                    model=model_setup.model_path,
+                    other_args=other_args,
+                    base_url=self.base_url,
+                    timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+                )
+
+                try:
+                    args = SimpleNamespace(
+                        base_url=self.base_url,
+                        model=model_setup.model_path,
+                        eval_name="mgsm_en",
+                        num_examples=None,
+                        num_threads=1024,
+                    )
+
+                    metrics = run_eval(args)
+                    print(
+                        f"{'=' * 42}\n{model_setup.model_path} - metrics={metrics} score={metrics['score']}\n{'=' * 42}\n"
+                    )
+
+                    write_results_to_json(
+                        model_setup.model_path, metrics, "w" if is_first else "a"
+                    )
+                    is_first = False
+
+                    # 0.0 for empty latency
+                    all_results.append((model_setup.model_path, metrics["score"], 0.0))
+                finally:
+                    kill_process_tree(process.pid)
+
+        try:
+            with open("results.json", "r") as f:
+                print("\nFinal Results from results.json:")
+                print(json.dumps(json.load(f), indent=2))
+        except Exception as e:
+            print(f"Error reading results.json: {e}")
+
+        # Check all scores after collecting all results
+        check_evaluation_test_results(
+            all_results,
+            self.__class__.__name__,
+            model_accuracy_thresholds=MODEL_SCORE_THRESHOLDS,
+            model_count=len(self.models),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

third_party/sglang/test/manual/nightly/test_text_models_perf.py

@@ -0,0 +1,59 @@
+import unittest
+
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    ModelLaunchSettings,
+    _parse_int_list_env,
+    parse_models,
+)
+
+PROFILE_DIR = "performance_profiles_text_models"
+
+
+class TestNightlyTextModelsPerformance(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.models = []
+        # TODO: replace with DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1 or other model lists
+        for model_path in parse_models("meta-llama/Llama-3.1-8B-Instruct"):
+            cls.models.append(ModelLaunchSettings(model_path, tp_size=1))
+        for model_path in parse_models("Qwen/Qwen2-57B-A14B-Instruct"):
+            cls.models.append(ModelLaunchSettings(model_path, tp_size=2))
+        # (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1), False, False),
+        # (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP2), False, True),
+        # (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP1), True, False),
+        # (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2), True, True),
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+
+    def test_bench_one_batch(self):
+        all_model_succeed = True
+
+        for model_setup in self.models:
+            with self.subTest(model=model_setup.model_path):
+                results, success = self.runner.run_benchmark_for_model(
+                    model_path=model_setup.model_path,
+                    batch_sizes=self.batch_sizes,
+                    input_lens=self.input_lens,
+                    output_lens=self.output_lens,
+                    other_args=model_setup.extra_args,
+                )
+
+                if not success:
+                    all_model_succeed = False
+
+                self.runner.add_report(results)
+
+        self.runner.write_final_report()
+
+        if not all_model_succeed:
+            raise AssertionError("Some models failed the perf tests.")
+
+
+if __name__ == "__main__":
+    unittest.main()

third_party/sglang/test/manual/nightly/test_vlms_mmmu_eval.py

@@ -0,0 +1,127 @@
+import json
+import unittest
+import warnings
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    ModelEvalMetrics,
+    ModelLaunchSettings,
+    check_evaluation_test_results,
+    popen_launch_server,
+    write_results_to_json,
+)
+
+MODEL_THRESHOLDS = {
+    # Conservative thresholds on 100 MMMU samples, especially for latency thresholds
+    ModelLaunchSettings("deepseek-ai/deepseek-vl2-small"): ModelEvalMetrics(
+        0.330, 56.1
+    ),
+    ModelLaunchSettings("deepseek-ai/Janus-Pro-7B"): ModelEvalMetrics(0.285, 40.3),
+    ModelLaunchSettings("Efficient-Large-Model/NVILA-8B-hf"): ModelEvalMetrics(
+        0.270, 56.7
+    ),
+    ModelLaunchSettings("Efficient-Large-Model/NVILA-Lite-2B-hf"): ModelEvalMetrics(
+        0.270, 23.8
+    ),
+    ModelLaunchSettings("google/gemma-3-4b-it"): ModelEvalMetrics(0.360, 10.9),
+    ModelLaunchSettings("google/gemma-3n-E4B-it"): ModelEvalMetrics(0.360, 17.7),
+    ModelLaunchSettings("mistral-community/pixtral-12b"): ModelEvalMetrics(0.360, 16.6),
+    ModelLaunchSettings("moonshotai/Kimi-VL-A3B-Instruct"): ModelEvalMetrics(
+        0.330, 22.3
+    ),
+    ModelLaunchSettings("openbmb/MiniCPM-o-2_6"): ModelEvalMetrics(0.330, 29.3),
+    ModelLaunchSettings("openbmb/MiniCPM-v-2_6"): ModelEvalMetrics(0.259, 36.3),
+    ModelLaunchSettings("OpenGVLab/InternVL2_5-2B"): ModelEvalMetrics(0.300, 17.0),
+    ModelLaunchSettings("Qwen/Qwen2-VL-7B-Instruct"): ModelEvalMetrics(0.310, 83.3),
+    ModelLaunchSettings("Qwen/Qwen2.5-VL-7B-Instruct"): ModelEvalMetrics(0.340, 31.9),
+    ModelLaunchSettings(
+        "Qwen/Qwen3-VL-30B-A3B-Instruct", extra_args=["--tp=2"]
+    ): ModelEvalMetrics(0.29, 37.0),
+    ModelLaunchSettings(
+        "unsloth/Mistral-Small-3.1-24B-Instruct-2503"
+    ): ModelEvalMetrics(0.310, 16.7),
+    ModelLaunchSettings("XiaomiMiMo/MiMo-VL-7B-RL"): ModelEvalMetrics(0.28, 32.0),
+    ModelLaunchSettings("zai-org/GLM-4.1V-9B-Thinking"): ModelEvalMetrics(0.280, 30.4),
+}
+
+
+class TestNightlyVLMMmmuEval(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.models = list(MODEL_THRESHOLDS.keys())
+        cls.base_url = DEFAULT_URL_FOR_TEST
+
+    def test_mmmu_vlm_models(self):
+        warnings.filterwarnings(
+            "ignore", category=ResourceWarning, message="unclosed.*socket"
+        )
+        is_first = True
+        all_results = []
+
+        for model in self.models:
+            model_path = model.model_path
+            with self.subTest(model=model_path):
+                process = popen_launch_server(
+                    model=model_path,
+                    base_url=self.base_url,
+                    other_args=model.extra_args,
+                    timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+                )
+                try:
+                    args = SimpleNamespace(
+                        base_url=self.base_url,
+                        model=model_path,
+                        eval_name="mmmu",
+                        num_examples=100,
+                        num_threads=64,
+                        max_tokens=30,
+                    )
+
+                    args.return_latency = True
+
+                    metrics, latency = run_eval(args)
+
+                    metrics["score"] = round(metrics["score"], 4)
+                    metrics["latency"] = round(latency, 4)
+                    print(
+                        f"{'=' * 42}\n{model_path} - metrics={metrics} score={metrics['score']}\n{'=' * 42}\n"
+                    )
+
+                    write_results_to_json(model_path, metrics, "w" if is_first else "a")
+                    is_first = False
+
+                    all_results.append(
+                        (model_path, metrics["score"], metrics["latency"])
+                    )
+                finally:
+                    kill_process_tree(process.pid)
+
+        try:
+            with open("results.json", "r") as f:
+                print("\nFinal Results from results.json:")
+                print(json.dumps(json.load(f), indent=2))
+        except Exception as e:
+            print(f"Error reading results: {e}")
+
+        model_accuracy_thresholds = {
+            model.model_path: threshold.accuracy
+            for model, threshold in MODEL_THRESHOLDS.items()
+        }
+        model_latency_thresholds = {
+            model.model_path: threshold.eval_time
+            for model, threshold in MODEL_THRESHOLDS.items()
+        }
+        check_evaluation_test_results(
+            all_results,
+            self.__class__.__name__,
+            model_accuracy_thresholds=model_accuracy_thresholds,
+            model_latency_thresholds=model_latency_thresholds,
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

third_party/sglang/test/manual/nightly/test_vlms_perf.py

@@ -0,0 +1,87 @@
+import os
+import unittest
+import warnings
+
+from sglang.test.nightly_utils import NightlyBenchmarkRunner
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    ModelLaunchSettings,
+    _parse_int_list_env,
+    parse_models,
+)
+
+PROFILE_DIR = "performance_profiles_vlms"
+
+MODEL_DEFAULTS = [
+    # Keep conservative defaults. Can be overridden by env NIGHTLY_VLM_MODELS
+    ModelLaunchSettings(
+        "Qwen/Qwen2.5-VL-7B-Instruct",
+        extra_args=["--mem-fraction-static=0.7"],
+    ),
+    ModelLaunchSettings(
+        "google/gemma-3-27b-it",
+    ),
+    ModelLaunchSettings("Qwen/Qwen3-VL-30B-A3B-Instruct", extra_args=["--tp=2"]),
+    # "OpenGVLab/InternVL2_5-2B",
+    # buggy in official transformers impl
+    # "openbmb/MiniCPM-V-2_6",
+]
+
+
+class TestNightlyVLMModelsPerformance(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        warnings.filterwarnings(
+            "ignore", category=ResourceWarning, message="unclosed.*socket"
+        )
+
+        nightly_vlm_models_str = os.environ.get("NIGHTLY_VLM_MODELS")
+        if nightly_vlm_models_str:
+            cls.models = []
+            model_paths = parse_models(nightly_vlm_models_str)
+            for model_path in model_paths:
+                cls.models.append(ModelLaunchSettings(model_path))
+        else:
+            cls.models = MODEL_DEFAULTS
+
+        cls.base_url = DEFAULT_URL_FOR_TEST
+
+        cls.batch_sizes = _parse_int_list_env("NIGHTLY_VLM_BATCH_SIZES", "1,1,2,8,16")
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_OUTPUT_LENS", "512"))
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+
+    def test_bench_one_batch(self):
+        all_model_succeed = True
+
+        for model_setup in self.models:
+            with self.subTest(model=model_setup.model_path):
+                # VLMs need additional benchmark args for dataset and trust-remote-code
+                extra_bench_args = [
+                    "--trust-remote-code",
+                    "--dataset-name=mmmu",
+                ]
+
+                results, success = self.runner.run_benchmark_for_model(
+                    model_path=model_setup.model_path,
+                    batch_sizes=self.batch_sizes,
+                    input_lens=self.input_lens,
+                    output_lens=self.output_lens,
+                    other_args=model_setup.extra_args,
+                    extra_bench_args=extra_bench_args,
+                )
+
+                if not success:
+                    all_model_succeed = False
+
+                self.runner.add_report(results)
+
+        self.runner.write_final_report()
+
+        if not all_model_succeed:
+            raise AssertionError("Some models failed the perf tests.")
+
+
+if __name__ == "__main__":
+    unittest.main()

third_party/sglang/test/manual/nightly/test_vlms_piecewise_cuda_graph.py

@@ -0,0 +1,262 @@
+import argparse
+import glob
+import json
+import os
+import random
+import sys
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.kits.mmmu_vlm_kit import _run_lmms_eval_with_retry
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+)
+
+MODELS = [
+    SimpleNamespace(model="Qwen/Qwen2.5-VL-7B-Instruct", mmmu_accuracy=0.60),
+]
+
+
+# Set default mem_fraction_static to 0.8
+DEFAULT_MEM_FRACTION_STATIC = 0.8
+
+
+class TestVLMPiecewiseCudaGraph(CustomTestCase):
+    parsed_args = None  # Class variable to store args
+
+    @classmethod
+    def setUpClass(cls):
+        # Removed argument parsing from here
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.time_out = DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+
+        if cls.parsed_args is None:
+            cls.parsed_args = SimpleNamespace(
+                mem_fraction_static=DEFAULT_MEM_FRACTION_STATIC
+            )
+
+        # Set OpenAI API key and base URL environment variables. Needed for lmm-evals to work.
+        os.environ["OPENAI_API_KEY"] = cls.api_key
+        os.environ["OPENAI_API_BASE"] = f"{cls.base_url}/v1"
+
+    def run_mmmu_eval(
+        self,
+        model_version: str,
+        output_path: str,
+        *,
+        env: dict | None = None,
+    ):
+        """
+        Evaluate a VLM on the MMMU validation set with lmms‑eval.
+        Only `model_version` (checkpoint) and `chat_template` vary;
+        We are focusing only on the validation set due to resource constraints.
+        """
+        # -------- fixed settings --------
+        model = "openai_compatible"
+        tp = 1
+        tasks = "mmmu_val"
+        batch_size = 32
+        log_suffix = "openai_compatible"
+        os.makedirs(output_path, exist_ok=True)
+
+        # -------- compose --model_args --------
+        model_args = f'model_version="{model_version}",' f"tp={tp}"
+
+        # -------- build command list --------
+        cmd = [
+            "python3",
+            "-m",
+            "lmms_eval",
+            "--model",
+            model,
+            "--model_args",
+            model_args,
+            "--tasks",
+            tasks,
+            "--batch_size",
+            str(batch_size),
+            "--output_path",
+            str(output_path),
+        ]
+
+        _run_lmms_eval_with_retry(cmd, timeout=3600)
+
+    def _run_vlm_mmmu_test(
+        self,
+        model,
+        output_path,
+        test_name="",
+        custom_env=None,
+        log_level="info",
+        capture_output=False,
+    ):
+        """
+        Common method to run VLM MMMU benchmark test.
+        Args:
+            model: Model to test
+            output_path: Path for output logs
+            test_name: Optional test name for logging
+            custom_env: Optional custom environment variables
+            log_level: Log level for server (default: "info")
+            capture_output: Whether to capture server stdout/stderr
+        """
+        print(f"\nTesting model: {model.model}{test_name}")
+
+        process = None
+        mmmu_accuracy = 0  # Initialize to handle potential exceptions
+        server_output = ""
+
+        try:
+            # Prepare environment variables
+            process_env = os.environ.copy()
+            if custom_env:
+                process_env.update(custom_env)
+            # if test vlm with cuda_ipc feature, open this env_var
+            process_env["SGLANG_USE_CUDA_IPC_TRANSPORT"] = "1"
+
+            # Prepare stdout/stderr redirection if needed
+            stdout_file = None
+            stderr_file = None
+            if capture_output:
+                stdout_file = open("/tmp/server_stdout.log", "w")
+                stderr_file = open("/tmp/server_stderr.log", "w")
+
+            # Launch server for testing
+            process = popen_launch_server(
+                model.model,
+                base_url=self.base_url,
+                timeout=self.time_out,
+                api_key=self.api_key,
+                other_args=[
+                    "--trust-remote-code",
+                    "--piecewise-cuda-graph-max-tokens",
+                    "8192",
+                    "--enforce-piecewise-cuda-graph",
+                    "--tp=8",
+                    "--piecewise-cuda-graph-compiler=eager",
+                    "--disable-radix-cache",
+                    "--log-level",
+                    log_level,
+                ],
+                env=process_env,
+                return_stdout_stderr=(
+                    (stdout_file, stderr_file) if capture_output else None
+                ),
+            )
+
+            # Run evaluation
+            self.run_mmmu_eval(model.model, output_path)
+
+            # Get the result file
+            # Search recursively for JSON result files (lmms-eval v0.4.1+ creates subdirectories)
+            result_files = glob.glob(f"{output_path}/**/*.json", recursive=True)
+            if not result_files:
+                result_files = glob.glob(f"{output_path}/*.json")
+
+            if not result_files:
+                raise FileNotFoundError(f"No JSON result files found in {output_path}")
+
+            result_file_path = result_files[0]
+
+            with open(result_file_path, "r") as f:
+                result = json.load(f)
+                print(f"Result{test_name}\n: {result}")
+
+            # Process the result
+            mmmu_accuracy = result["results"]["mmmu_val"]["mmmu_acc,none"]
+            print(
+                f"Model {model.model} achieved accuracy{test_name}: {mmmu_accuracy:.4f}"
+            )
+
+            # Capture server output if requested
+            if capture_output and process:
+                server_output = self._read_output_from_files()
+
+            # Assert performance meets expected threshold
+            self.assertGreaterEqual(
+                mmmu_accuracy,
+                model.mmmu_accuracy,
+                f"Model {model.model} accuracy ({mmmu_accuracy:.4f}) below expected threshold ({model.mmmu_accuracy:.4f}){test_name}",
+            )
+
+            return server_output
+
+        except Exception as e:
+            print(f"Error testing {model.model}{test_name}: {e}")
+            self.fail(f"Test failed for {model.model}{test_name}: {e}")
+
+        finally:
+            # Ensure process cleanup happens regardless of success/failure
+            if process is not None and process.poll() is None:
+                print(f"Cleaning up process {process.pid}")
+                try:
+                    kill_process_tree(process.pid)
+                except Exception as e:
+                    print(f"Error killing process: {e}")
+
+            # clean up temporary files
+            if capture_output:
+                if stdout_file:
+                    stdout_file.close()
+                if stderr_file:
+                    stderr_file.close()
+                for filename in ["/tmp/server_stdout.log", "/tmp/server_stderr.log"]:
+                    try:
+                        if os.path.exists(filename):
+                            os.remove(filename)
+                    except Exception as e:
+                        print(f"Error removing {filename}: {e}")
+
+    def _read_output_from_files(self):
+        output_lines = []
+
+        log_files = [
+            ("/tmp/server_stdout.log", "[STDOUT]"),
+            ("/tmp/server_stderr.log", "[STDERR]"),
+        ]
+        for filename, tag in log_files:
+            try:
+                if os.path.exists(filename):
+                    with open(filename, "r") as f:
+                        for line in f:
+                            output_lines.append(f"{tag} {line.rstrip()}")
+            except Exception as e:
+                print(f"Error reading {tag.lower()} file: {e}")
+
+        return "\n".join(output_lines)
+
+    def test_vlm_mmmu_benchmark(self):
+        """Test VLM models against MMMU benchmark."""
+        models_to_test = MODELS
+
+        if is_in_ci():
+            models_to_test = [random.choice(MODELS)]
+
+        for model in models_to_test:
+            self._run_vlm_mmmu_test(model, "./logs")
+
+
+if __name__ == "__main__":
+    # Define and parse arguments here, before unittest.main
+    parser = argparse.ArgumentParser(description="Test VLM models")
+    parser.add_argument(
+        "--mem-fraction-static",
+        type=float,
+        help="Static memory fraction for the model",
+        default=DEFAULT_MEM_FRACTION_STATIC,
+    )
+
+    # Parse args intended for unittest
+    args = parser.parse_args()
+
+    # Store the parsed args object on the class
+    TestVLMPiecewiseCudaGraph.parsed_args = args
+
+    # Pass args to unittest
+    unittest.main(argv=[sys.argv[0]])

third_party/sglang/test/manual/nightly/test_vlms_vit_cuda_graph.py

@@ -0,0 +1,268 @@
+import argparse
+import glob
+import json
+import os
+import random
+import sys
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.kits.mmmu_vlm_kit import _run_lmms_eval_with_retry
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+)
+
+MODELS = [
+    SimpleNamespace(model="Qwen/Qwen2.5-VL-7B-Instruct", mmmu_accuracy=0.60),
+    SimpleNamespace(model="Qwen/Qwen3-VL-8B-Instruct", mmmu_accuracy=0.60),
+]
+
+
+# Set default mem_fraction_static to 0.8
+DEFAULT_MEM_FRACTION_STATIC = 0.8
+
+
+class TestVLMViTCudaGraph(CustomTestCase):
+    parsed_args = None  # Class variable to store args
+
+    @classmethod
+    def setUpClass(cls):
+        # Removed argument parsing from here
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.time_out = DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        cls.enable_vit_cuda_graph = "1"
+
+        if cls.parsed_args is None:
+            cls.parsed_args = SimpleNamespace(
+                mem_fraction_static=DEFAULT_MEM_FRACTION_STATIC
+            )
+
+        # Set OpenAI API key and base URL environment variables. Needed for lmm-evals to work.
+        os.environ["OPENAI_API_KEY"] = cls.api_key
+        os.environ["OPENAI_API_BASE"] = f"{cls.base_url}/v1"
+        os.environ["SGLANG_VIT_ENABLE_CUDA_GRAPH"] = cls.enable_vit_cuda_graph
+
+    def run_mmmu_eval(
+        self,
+        model_version: str,
+        output_path: str,
+        *,
+        env: dict | None = None,
+    ):
+        """
+        Evaluate a VLM on the MMMU validation set with lmms‑eval.
+        Only `model_version` (checkpoint) and `chat_template` vary;
+        We are focusing only on the validation set due to resource constraints.
+        """
+        # -------- fixed settings --------
+        model = "openai_compatible"
+        tp = 1
+        tasks = "mmmu_val"
+        batch_size = 32
+        log_suffix = "openai_compatible"
+        os.makedirs(output_path, exist_ok=True)
+
+        # -------- compose --model_args --------
+        model_args = f'model_version="{model_version}",' f"tp={tp}"
+
+        # -------- build command list --------
+        cmd = [
+            "python3",
+            "-m",
+            "lmms_eval",
+            "--model",
+            model,
+            "--model_args",
+            model_args,
+            "--tasks",
+            tasks,
+            "--batch_size",
+            str(batch_size),
+            "--output_path",
+            str(output_path),
+        ]
+
+        _run_lmms_eval_with_retry(cmd, timeout=3600)
+
+    def _run_vlm_mmmu_test(
+        self,
+        model,
+        output_path,
+        test_name="",
+        custom_env=None,
+        log_level="info",
+        capture_output=False,
+    ):
+        """
+        Common method to run VLM MMMU benchmark test.
+        Args:
+            model: Model to test
+            output_path: Path for output logs
+            test_name: Optional test name for logging
+            custom_env: Optional custom environment variables
+            log_level: Log level for server (default: "info")
+            capture_output: Whether to capture server stdout/stderr
+        """
+        print(f"\nTesting model: {model.model}{test_name}")
+
+        process = None
+        mmmu_accuracy = 0  # Initialize to handle potential exceptions
+        server_output = ""
+
+        try:
+            # Prepare environment variables
+            process_env = os.environ.copy()
+            if custom_env:
+                process_env.update(custom_env)
+            # if test vlm with cuda_ipc feature, open this env_var
+            process_env["SGLANG_USE_CUDA_IPC_TRANSPORT"] = "1"
+            process_env["SGLANG_VIT_ENABLE_CUDA_GRAPH"] = "1"
+
+            # Prepare stdout/stderr redirection if needed
+            stdout_file = None
+            stderr_file = None
+            if capture_output:
+                stdout_file = open("/tmp/server_stdout.log", "w")
+                stderr_file = open("/tmp/server_stderr.log", "w")
+
+            # Launch server for testing
+            process = popen_launch_server(
+                model.model,
+                base_url=self.base_url,
+                timeout=self.time_out,
+                api_key=self.api_key,
+                other_args=[
+                    "--mm-attention-backend",
+                    "fa3",
+                    "--enforce-piecewise-cuda-graph",
+                    "--piecewise-cuda-graph-max-tokens",
+                    "8192",
+                    "--chunked-prefill-size",
+                    "8192",
+                    "--disable-radix-cache",
+                    "--disable-overlap-schedule",
+                    "--piecewise-cuda-graph-compiler",
+                    "eager",
+                ],
+                env=process_env,
+                return_stdout_stderr=(
+                    (stdout_file, stderr_file) if capture_output else None
+                ),
+            )
+
+            # Run evaluation
+            self.run_mmmu_eval(model.model, output_path)
+
+            # Get the result file
+            # Search recursively for JSON result files (lmms-eval v0.4.1+ creates subdirectories)
+            result_files = glob.glob(f"{output_path}/**/*.json", recursive=True)
+            if not result_files:
+                result_files = glob.glob(f"{output_path}/*.json")
+
+            if not result_files:
+                raise FileNotFoundError(f"No JSON result files found in {output_path}")
+
+            result_file_path = result_files[0]
+
+            with open(result_file_path, "r") as f:
+                result = json.load(f)
+                print(f"Result{test_name}\n: {result}")
+
+            # Process the result
+            mmmu_accuracy = result["results"]["mmmu_val"]["mmmu_acc,none"]
+            print(
+                f"Model {model.model} achieved accuracy{test_name}: {mmmu_accuracy:.4f}"
+            )
+
+            # Capture server output if requested
+            if capture_output and process:
+                server_output = self._read_output_from_files()
+
+            # Assert performance meets expected threshold
+            self.assertGreaterEqual(
+                mmmu_accuracy,
+                model.mmmu_accuracy,
+                f"Model {model.model} accuracy ({mmmu_accuracy:.4f}) below expected threshold ({model.mmmu_accuracy:.4f}){test_name}",
+            )
+
+            return server_output
+
+        except Exception as e:
+            print(f"Error testing {model.model}{test_name}: {e}")
+            self.fail(f"Test failed for {model.model}{test_name}: {e}")
+
+        finally:
+            # Ensure process cleanup happens regardless of success/failure
+            if process is not None and process.poll() is None:
+                print(f"Cleaning up process {process.pid}")
+                try:
+                    kill_process_tree(process.pid)
+                except Exception as e:
+                    print(f"Error killing process: {e}")
+
+            # clean up temporary files
+            if capture_output:
+                if stdout_file:
+                    stdout_file.close()
+                if stderr_file:
+                    stderr_file.close()
+                for filename in ["/tmp/server_stdout.log", "/tmp/server_stderr.log"]:
+                    try:
+                        if os.path.exists(filename):
+                            os.remove(filename)
+                    except Exception as e:
+                        print(f"Error removing {filename}: {e}")
+
+    def _read_output_from_files(self):
+        output_lines = []
+
+        log_files = [
+            ("/tmp/server_stdout.log", "[STDOUT]"),
+            ("/tmp/server_stderr.log", "[STDERR]"),
+        ]
+        for filename, tag in log_files:
+            try:
+                if os.path.exists(filename):
+                    with open(filename, "r") as f:
+                        for line in f:
+                            output_lines.append(f"{tag} {line.rstrip()}")
+            except Exception as e:
+                print(f"Error reading {tag.lower()} file: {e}")
+
+        return "\n".join(output_lines)
+
+    def test_vlm_mmmu_benchmark(self):
+        """Test VLM models against MMMU benchmark."""
+        models_to_test = MODELS
+
+        if is_in_ci():
+            models_to_test = [random.choice(MODELS)]
+
+        for model in models_to_test:
+            self._run_vlm_mmmu_test(model, "./logs")
+
+
+if __name__ == "__main__":
+    # Define and parse arguments here, before unittest.main
+    parser = argparse.ArgumentParser(description="Test VLM models")
+    parser.add_argument(
+        "--mem-fraction-static",
+        type=float,
+        help="Static memory fraction for the model",
+        default=DEFAULT_MEM_FRACTION_STATIC,
+    )
+
+    # Parse args intended for unittest
+    args = parser.parse_args()
+
+    # Store the parsed args object on the class
+    TestVLMViTCudaGraph.parsed_args = args
+
+    # Pass args to unittest
+    unittest.main(argv=[sys.argv[0]])

third_party/sglang/test/manual/nightly/test_vlms_vit_flashinfer_cudnn.py

@@ -0,0 +1,258 @@
+import argparse
+import glob
+import json
+import os
+import random
+import sys
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.kits.mmmu_vlm_kit import _run_lmms_eval_with_retry
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+)
+
+MODELS = [
+    SimpleNamespace(model="Qwen/Qwen3-VL-30B-A3B-Instruct", mmmu_accuracy=0.51),
+]
+
+
+# Set default mem_fraction_static to 0.8
+DEFAULT_MEM_FRACTION_STATIC = 0.8
+
+
+class TestVLMViTFlashinferCudnn(CustomTestCase):
+    parsed_args = None  # Class variable to store args
+
+    @classmethod
+    def setUpClass(cls):
+        # Removed argument parsing from here
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.time_out = DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+
+        if cls.parsed_args is None:
+            cls.parsed_args = SimpleNamespace(
+                mem_fraction_static=DEFAULT_MEM_FRACTION_STATIC
+            )
+
+        # Set OpenAI API key and base URL environment variables. Needed for lmm-evals to work.
+        os.environ["OPENAI_API_KEY"] = cls.api_key
+        os.environ["OPENAI_API_BASE"] = f"{cls.base_url}/v1"
+
+    def run_mmmu_eval(
+        self,
+        model_version: str,
+        output_path: str,
+        *,
+        env: dict | None = None,
+    ):
+        """
+        Evaluate a VLM on the MMMU validation set with lmms‑eval.
+        Only `model_version` (checkpoint) and `chat_template` vary;
+        We are focusing only on the validation set due to resource constraints.
+        """
+        # -------- fixed settings --------
+        model = "openai_compatible"
+        tp = 1
+        tasks = "mmmu_val"
+        batch_size = 32
+        log_suffix = "openai_compatible"
+        os.makedirs(output_path, exist_ok=True)
+
+        # -------- compose --model_args --------
+        model_args = f'model_version="{model_version}",' f"tp={tp}"
+
+        # -------- build command list --------
+        cmd = [
+            "python3",
+            "-m",
+            "lmms_eval",
+            "--model",
+            model,
+            "--model_args",
+            model_args,
+            "--tasks",
+            tasks,
+            "--batch_size",
+            str(batch_size),
+            "--output_path",
+            str(output_path),
+        ]
+
+        _run_lmms_eval_with_retry(cmd, timeout=3600)
+
+    def _run_vlm_mmmu_test(
+        self,
+        model,
+        output_path,
+        test_name="",
+        custom_env=None,
+        log_level="info",
+        capture_output=False,
+    ):
+        """
+        Common method to run VLM MMMU benchmark test.
+        Args:
+            model: Model to test
+            output_path: Path for output logs
+            test_name: Optional test name for logging
+            custom_env: Optional custom environment variables
+            log_level: Log level for server (default: "info")
+            capture_output: Whether to capture server stdout/stderr
+        """
+        print(f"\nTesting model: {model.model}{test_name}")
+
+        process = None
+        mmmu_accuracy = 0  # Initialize to handle potential exceptions
+        server_output = ""
+
+        try:
+            # Prepare environment variables
+            process_env = os.environ.copy()
+            if custom_env:
+                process_env.update(custom_env)
+            # if test vlm with cuda_ipc feature, open this env_var
+            process_env["SGLANG_USE_CUDA_IPC_TRANSPORT"] = "1"
+
+            # Prepare stdout/stderr redirection if needed
+            stdout_file = None
+            stderr_file = None
+            if capture_output:
+                stdout_file = open("/tmp/server_stdout.log", "w")
+                stderr_file = open("/tmp/server_stderr.log", "w")
+
+            # Launch server for testing
+            process = popen_launch_server(
+                model.model,
+                base_url=self.base_url,
+                timeout=self.time_out,
+                api_key=self.api_key,
+                other_args=[
+                    "--mm-attention-backend",
+                    "flashinfer_cudnn",
+                    "--chunked-prefill-size",
+                    "8192",
+                    "--disable-radix-cache",
+                ],
+                env=process_env,
+                return_stdout_stderr=(
+                    (stdout_file, stderr_file) if capture_output else None
+                ),
+            )
+
+            # Run evaluation
+            self.run_mmmu_eval(model.model, output_path)
+
+            # Get the result file
+            # Search recursively for JSON result files (lmms-eval v0.4.1+ creates subdirectories)
+            result_files = glob.glob(f"{output_path}/**/*.json", recursive=True)
+            if not result_files:
+                result_files = glob.glob(f"{output_path}/*.json")
+
+            if not result_files:
+                raise FileNotFoundError(f"No JSON result files found in {output_path}")
+
+            result_file_path = result_files[0]
+
+            with open(result_file_path, "r") as f:
+                result = json.load(f)
+                print(f"Result{test_name}\n: {result}")
+
+            # Process the result
+            mmmu_accuracy = result["results"]["mmmu_val"]["mmmu_acc,none"]
+            print(
+                f"Model {model.model} achieved accuracy{test_name}: {mmmu_accuracy:.4f}"
+            )
+
+            # Capture server output if requested
+            if capture_output and process:
+                server_output = self._read_output_from_files()
+
+            # Assert performance meets expected threshold
+            self.assertGreaterEqual(
+                mmmu_accuracy,
+                model.mmmu_accuracy,
+                f"Model {model.model} accuracy ({mmmu_accuracy:.4f}) below expected threshold ({model.mmmu_accuracy:.4f}){test_name}",
+            )
+
+            return server_output
+
+        except Exception as e:
+            print(f"Error testing {model.model}{test_name}: {e}")
+            self.fail(f"Test failed for {model.model}{test_name}: {e}")
+
+        finally:
+            # Ensure process cleanup happens regardless of success/failure
+            if process is not None and process.poll() is None:
+                print(f"Cleaning up process {process.pid}")
+                try:
+                    kill_process_tree(process.pid)
+                except Exception as e:
+                    print(f"Error killing process: {e}")
+
+            # clean up temporary files
+            if capture_output:
+                if stdout_file:
+                    stdout_file.close()
+                if stderr_file:
+                    stderr_file.close()
+                for filename in ["/tmp/server_stdout.log", "/tmp/server_stderr.log"]:
+                    try:
+                        if os.path.exists(filename):
+                            os.remove(filename)
+                    except Exception as e:
+                        print(f"Error removing {filename}: {e}")
+
+    def _read_output_from_files(self):
+        output_lines = []
+
+        log_files = [
+            ("/tmp/server_stdout.log", "[STDOUT]"),
+            ("/tmp/server_stderr.log", "[STDERR]"),
+        ]
+        for filename, tag in log_files:
+            try:
+                if os.path.exists(filename):
+                    with open(filename, "r") as f:
+                        for line in f:
+                            output_lines.append(f"{tag} {line.rstrip()}")
+            except Exception as e:
+                print(f"Error reading {tag.lower()} file: {e}")
+
+        return "\n".join(output_lines)
+
+    def test_vlm_mmmu_benchmark(self):
+        """Test VLM models against MMMU benchmark."""
+        models_to_test = MODELS
+
+        if is_in_ci():
+            models_to_test = [random.choice(MODELS)]
+
+        for model in models_to_test:
+            self._run_vlm_mmmu_test(model, "./logs")
+
+
+if __name__ == "__main__":
+    # Define and parse arguments here, before unittest.main
+    parser = argparse.ArgumentParser(description="Test VLM models")
+    parser.add_argument(
+        "--mem-fraction-static",
+        type=float,
+        help="Static memory fraction for the model",
+        default=DEFAULT_MEM_FRACTION_STATIC,
+    )
+
+    # Parse args intended for unittest
+    args = parser.parse_args()
+
+    # Store the parsed args object on the class
+    TestVLMViTFlashinferCudnn.parsed_args = args
+
+    # Pass args to unittest
+    unittest.main(argv=[sys.argv[0]])