Add dump_metric to MMMU, lm-eval, and NeMo Skills eval paths (#22147)

python/sglang/test/accuracy_test_runner.py

@@ -8,6 +8,7 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     ModelLaunchSettings,
+    dump_metric,
     popen_launch_server,
     write_github_step_summary,
 )
@@ -421,6 +422,12 @@ def _run_nemo_skills_eval(
         if score is None:
             return False, "Could not parse accuracy from ns eval output", None
 
+        dump_metric(
+            f"{dataset}_score",
+            score,
+            labels={"model": model.model_path, "eval": dataset, "api": "nemo-skills"},
+        )
+
         return True, None, {"score": score}
 
     except subprocess.TimeoutExpired:

python/sglang/test/kits/lm_eval_kit.py

@@ -12,6 +12,8 @@ import numpy as np
 import requests
 import yaml
 
+from sglang.test.test_utils import dump_metric
+
 
 @contextmanager
 def scoped_env_vars(new_env: dict[str, str] | None):
@@ -69,6 +71,15 @@ class LMEvalMixin:
                     f"ground_truth={ground_truth:.3f} | "
                     f"measured={measured_value:.3f} | rtol={rtol}"
                 )
+                dump_metric(
+                    f"{task['name']}_{metric['name']}",
+                    measured_value,
+                    labels={
+                        "model": eval_config.get("model_name", ""),
+                        "eval": "lm-eval",
+                        "task": task["name"],
+                    },
+                )
                 success = success and np.isclose(
                     ground_truth, measured_value, rtol=rtol
                 )

python/sglang/test/kits/mmmu_vlm_kit.py

@@ -13,6 +13,7 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
+    dump_metric,
     popen_launch_server,
 )
 
@@ -216,6 +217,12 @@ class MMMUMixin:
             mmmu_accuracy = result["results"]["mmmu_val"]["mmmu_acc,none"]
             print(f"Model {self.model} achieved accuracy: {mmmu_accuracy:.4f}")
 
+            dump_metric(
+                "mmmu_score",
+                mmmu_accuracy,
+                labels={"model": self.model, "eval": "mmmu", "api": "lmms-eval"},
+            )
+
             # Assert performance meets expected threshold
             self.assertGreaterEqual(
                 mmmu_accuracy,
@@ -403,6 +410,12 @@ class MMMUMultiModelTestBase(CustomTestCase):
                 f"Model {model.model} achieved accuracy{test_name}: {mmmu_accuracy:.4f}"
             )
 
+            dump_metric(
+                "mmmu_score",
+                mmmu_accuracy,
+                labels={"model": model.model, "eval": "mmmu", "api": "lmms-eval"},
+            )
+
             # Capture server output if requested
             if capture_output and process:
                 server_output = self._read_output_from_files()