[feat]: Add math eval to CI nightly run (#2663)

Co-authored-by: Chayenne <zhaochen20@outlook.com>

test/srt/run_suite.py

@@ -49,6 +49,7 @@ suites = {
     "nightly": [
         "test_nightly_gsm8k_eval.py",
         "test_nightly_human_eval.py",
+        "test_nightly_math_eval.py",
     ],
     "sampling/penaltylib": glob.glob(
         "sampling/penaltylib/**/test_*.py", recursive=True

test/srt/test_nightly_gsm8k_eval.py

@@ -25,7 +25,7 @@ MODEL_SCORE_THRESHOLDS = {
     "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": 0.84,
     "google/gemma-2-27b-it": 0.92,
     "meta-llama/Llama-3.1-70B-Instruct": 0.96,
-    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.64,
+    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.63,
     "Qwen/Qwen2-57B-A14B-Instruct": 0.87,
     "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.84,
     "neuralmagic/Mistral-7B-Instruct-v0.3-FP8": 0.54,
@@ -36,7 +36,7 @@ MODEL_SCORE_THRESHOLDS = {
     "neuralmagic/Qwen2-72B-Instruct-FP8": 0.95,
     "neuralmagic/Qwen2-57B-A14B-Instruct-FP8": 0.82,
     "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4": 0.84,
-    "hugging-quants/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4": 0.84,
+    "hugging-quants/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4": 0.83,
 }
 
 

test/srt/test_nightly_human_eval.py

@@ -12,19 +12,28 @@ from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2,
     DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1,
     DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP2,
+    DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_URL_FOR_TEST,
+    is_in_ci,
 )
 
 
 class TestEvalAccuracyLarge(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
-        cls.model_groups = [
+        if is_in_ci():
-            (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1), False, False),
+            cls.model_groups = [([DEFAULT_MODEL_NAME_FOR_TEST], False, False)]
-            (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP2), False, True),
+        else:
-            (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP1), True, False),
+            cls.model_groups = [
-            (parse_models(DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2), True, True),
+                (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.process = None
         cls.eval_process = None

test/srt/test_nightly_math_eval.py

@@ -0,0 +1,46 @@
+import unittest
+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_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestEvalAccuracyLarge(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST
+        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=["--log-level-http", "warning"],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_math(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="math",
+            num_examples=5000,
+            num_threads=1024,
+        )
+
+        metrics = run_eval(args)
+        self.assertGreaterEqual(
+            metrics["score"], 0.519 - 0.02
+        )  # -2% to account for sampling variance
+
+
+if __name__ == "__main__":
+    unittest.main()