[VLM] Add VLM TP=4 per-commit CI test and improve MMMU eval prompt/parser (#21841)

python/sglang/test/simple_eval_mmmu_vlm.py

@@ -148,11 +148,19 @@ class MMMUVLMEval(Eval):
                     options = None
 
             # Build final textual prompt; include choices if MC
-            prompt_text = f"Question: {question}\n\n"
+            prompt_text = f"{question}\n"
             if options:
                 letters = [chr(ord("A") + i) for i in range(len(options))]
                 for letter, opt in zip(letters, options):
-                    prompt_text += f"{letter}) {opt}\n"
+                    prompt_text += f"{letter}. {opt}\n"
+                prompt_text += (
+                    "\nAnswer the following multiple-choice question. "
+                    "The last line of your response should be of the "
+                    "following format: 'Answer: $LETTER' (without quotes) "
+                    "where LETTER is one of the options. "
+                    "Think step by step before answering."
+                )
+            else:
                 prompt_text += "\nAnswer: "
 
             samples.append(
@@ -330,6 +338,14 @@ def _parse_multi_choice_response(
     response: str, all_choices: List[str], index2ans: dict
 ) -> str:
     # loosely adapted from benchmark mmmu eval
+
+    # First, look for explicit "Answer: X" pattern (last occurrence)
+    answer_matches = re.findall(r"[Aa]nswer\s*:\s*\*?\*?\s*\(?([A-Z])\)?", response)
+    if answer_matches:
+        candidate = answer_matches[-1]
+        if candidate in all_choices:
+            return candidate
+
     for char in [",", ".", "!", "?", ";", ":", "'"]:
         response = response.strip(char)
     response = " " + response + " "

test/registered/vlm/test_vlm_tp4.py

@@ -0,0 +1,82 @@
+"""
+VLM TP=4 per-commit test using Qwen3.5-27B with MMMU evaluation.
+"""
+
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=200, suite="stage-c-test-4-gpu-h100")
+
+QWEN35_27B_MODEL = "Qwen/Qwen3.5-27B"
+MMMU_ACCURACY_THRESHOLD = 0.65
+MMMU_NUM_EXAMPLES = 32
+
+
+class TestVLMTP4(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = QWEN35_27B_MODEL
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--tp-size",
+                "4",
+                "--cuda-graph-max-bs",
+                "32",
+                "--mem-fraction-static",
+                "0.8",
+                "--trust-remote-code",
+                "--mamba-scheduler-strategy",
+                "extra_buffer",
+                "--mamba-track-interval",
+                "128",
+                "--mamba-ssm-dtype",
+                "bfloat16",
+                "--chunked-prefill-size",
+                "2048",
+                "--max-running-requests",
+                "128",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        if hasattr(cls, "process") and cls.process:
+            kill_process_tree(cls.process.pid)
+
+    def test_mmmu_accuracy(self):
+        args = SimpleNamespace(
+            model=self.model,
+            eval_name="mmmu",
+            num_examples=MMMU_NUM_EXAMPLES,
+            num_threads=16,
+            max_tokens=2048,
+            chat_template_kwargs={"enable_thinking": False},
+            base_url=self.base_url,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval(args)
+        print(f"MMMU score: {metrics['score']}")
+        self.assertGreaterEqual(
+            metrics["score"],
+            MMMU_ACCURACY_THRESHOLD,
+            f"MMMU accuracy {metrics['score']:.4f} below threshold {MMMU_ACCURACY_THRESHOLD}",
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()