From 46ccd7ce3e70455a971e6a7f7765cd78bc246322 Mon Sep 17 00:00:00 2001
From: Joe Rowell <joe@poolside.ai>
Date: Sat, 29 Aug 2026 12:34:46 +0200
Subject: [PATCH] Add Laguna-XS-2.1 / S-2.1 NVFP4 nightly gsm8k tests (#35547)

Signed-off-by: Joe Rowell <joe@poolside.ai>
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
---
 python/sglang/test/accuracy_test_runner.py    |  6 ++
 .../4-gpu-models/test_laguna_nvfp4_nightly.py | 78 +++++++++++++++++++
 2 files changed, 84 insertions(+)
 create mode 100644 test/registered/4-gpu-models/test_laguna_nvfp4_nightly.py

diff --git a/python/sglang/test/accuracy_test_runner.py b/python/sglang/test/accuracy_test_runner.py
index 9322e9b92..7da5d4289 100644
--- a/python/sglang/test/accuracy_test_runner.py
+++ b/python/sglang/test/accuracy_test_runner.py
@@ -34,6 +34,7 @@ class AccuracyTestParams:
     # sgl-eval-backed datasets only: force chat_template_kwargs.thinking instead
     # of letting _run_sgl_eval infer it from the model name.
     sgl_eval_thinking: Optional[bool] = None
+    num_shots: Optional[int] = None  # few-shot count; None = run_eval's default
 
 
 @dataclass
@@ -84,6 +85,7 @@ def _run_simple_eval(
     dataset: str,
     num_examples: Optional[int] = None,
     num_threads: Optional[int] = None,
+    num_shots: Optional[int] = None,
     max_tokens: Optional[int] = None,
     return_latency: bool = False,
     thinking_mode: Optional[str] = None,
@@ -118,6 +120,9 @@ def _run_simple_eval(
             num_threads=num_threads or 1024,
         )
 
+        if num_shots is not None:
+            args.num_shots = num_shots
+
         if api is not None:
             args.api = api
 
@@ -196,6 +201,7 @@ def run_accuracy_test(
         dataset=params.dataset,
         num_examples=params.num_examples,
         num_threads=params.num_threads,
+        num_shots=params.num_shots,
         max_tokens=params.max_tokens,
         return_latency=params.return_latency,
         thinking_mode=params.thinking_mode,
diff --git a/test/registered/4-gpu-models/test_laguna_nvfp4_nightly.py b/test/registered/4-gpu-models/test_laguna_nvfp4_nightly.py
new file mode 100644
index 000000000..8f0c2ffde
--- /dev/null
+++ b/test/registered/4-gpu-models/test_laguna_nvfp4_nightly.py
@@ -0,0 +1,78 @@
+import unittest
+
+from sglang.test.accuracy_test_runner import AccuracyTestParams
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.performance_test_runner import PerformanceTestParams
+from sglang.test.run_combined_tests import run_combined_tests
+from sglang.test.test_utils import ModelLaunchSettings, is_blackwell_system
+
+register_cuda_ci(est_time=1800, stage="nightly", runner_config="4-gpu-b200")
+
+LAGUNA_XS_NVFP4_MODEL = "poolside/Laguna-XS-2.1-NVFP4"
+LAGUNA_S_NVFP4_MODEL = "poolside/Laguna-S-2.1-NVFP4"
+
+# Measured 0.935 (temp=1.0, 200 examples); floored with margin for
+# FP4-kernel variance across Blackwell parts.
+LAGUNA_XS_GSM8K_BASELINE = 0.87
+# Measured 0.95 (temp=1.0, 200 examples); floored with margin.
+LAGUNA_S_GSM8K_BASELINE = 0.89
+
+
+class TestLagunaNVFP4Nightly(unittest.TestCase):
+    """Nightly test for Laguna-XS-2.1 / S-2.1 NVFP4, TP=1, Blackwell only.
+
+    Each model runs BOTH:
+    - Performance test (using NightlyBenchmarkRunner)
+    - Accuracy test (using run_eval with gsm8k)
+
+    The models carry different gsm8k baselines, so each gets its own
+    run_combined_tests call (the runner takes one baseline per call).
+    """
+
+    def _run_model(self, model_path: str, test_name: str, baseline: float) -> None:
+        run_combined_tests(
+            models=[
+                ModelLaunchSettings(
+                    model_path,
+                    tp_size=1,
+                    variant="TP1",
+                )
+            ],
+            test_name=test_name,
+            accuracy_params=AccuracyTestParams(
+                dataset="gsm8k",
+                baseline_accuracy=baseline,
+                num_examples=200,
+                num_shots=5,
+                num_threads=128,
+                max_tokens=4096,
+                temperature=1.0,
+                top_p=0.95,
+                repeat=1,
+            ),
+            performance_params=PerformanceTestParams(
+                result_dir="performance_results_laguna_nvfp4",
+            ),
+        )
+
+    @unittest.skipIf(not is_blackwell_system(), "NVFP4 requires Blackwell")
+    def test_laguna_xs_nvfp4(self):
+        """Run performance and accuracy for Laguna-XS-2.1-NVFP4 (TP1)."""
+        self._run_model(
+            model_path=LAGUNA_XS_NVFP4_MODEL,
+            test_name="Laguna-XS-2.1-NVFP4",
+            baseline=LAGUNA_XS_GSM8K_BASELINE,
+        )
+
+    @unittest.skipIf(not is_blackwell_system(), "NVFP4 requires Blackwell")
+    def test_laguna_s_nvfp4(self):
+        """Run performance and accuracy for Laguna-S-2.1-NVFP4 (TP1)."""
+        self._run_model(
+            model_path=LAGUNA_S_NVFP4_MODEL,
+            test_name="Laguna-S-2.1-NVFP4",
+            baseline=LAGUNA_S_GSM8K_BASELINE,
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()