Support nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 (and nvidia/C-RADIOv2-H) (#12277)

test/srt/models/test_nvidia_nemotron_nano_v2_vl.py

@@ -0,0 +1,31 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.test.gsm8k_mixin import GSM8KMixin
+from sglang.test.mmmu_vlm_mixin import MMMUVLMMixin
+from sglang.test.test_utils import CustomTestCase
+
+MODEL = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
+
+
+class TestNvidiaNemotronNanoV2VLTextOnly(GSM8KMixin, CustomTestCase):
+    accuracy = 0.87
+    model = MODEL
+    other_args = ["--max-mamba-cache-size", "256", "--trust-remote-code"]
+
+
+class TestNvidiaNemotronNanoV2VLMMMU(MMMUVLMMixin, CustomTestCase):
+    accuracy = 0.454
+    model = MODEL
+    other_args = ["--max-mamba-cache-size", "128", "--trust-remote-code"]
+    mmmu_args = ["--limit=0.1"]
+    """`--limit=0.1`: 10 percent of each task - this is fine for testing since the nominal result isn't interesting - this run is just to prevent relative regressions."""
+
+    def test_vlm_mmmu_benchmark(self):
+        self._run_vlm_mmmu_test(
+            SimpleNamespace(model=self.model, mmmu_accuracy=self.accuracy), "./logs"
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/run_suite.py

@@ -24,6 +24,7 @@ suites = {
         TestFile("models/test_encoder_embedding_models.py", 460),
         TestFile("models/test_generation_models.py", 103),
         TestFile("models/test_nvidia_nemotron_nano_v2.py", 160),
+        TestFile("models/test_nvidia_nemotron_nano_v2_vl.py", 350),  # GSM8k + MMMU
         TestFile("models/test_qwen_models.py", 150),
         TestFile("models/test_reward_models.py", 132),
         TestFile("models/test_transformers_models.py", 320),
@@ -125,6 +126,7 @@ suites = {
         TestFile("test_triton_moe_channel_fp8_kernel.py", 25),
         TestFile("test_triton_sliding_window.py", 100),
         TestFile("test_utils_update_weights.py", 48),
+        TestFile("test_video_utils.py", 5),
         TestFile("test_vision_chunked_prefill.py", 170),
         TestFile("test_vision_openai_server_a.py", 900),
         TestFile("test_vlm_input_format.py", 300),

test/srt/test_video_utils.py

@@ -0,0 +1,59 @@
+from dataclasses import dataclass
+
+import pytest
+
+from sglang.srt.utils import sample_video_frames
+
+
+class DummyVideo:
+    def __init__(self, total_frames: int, avg_fps: float):
+        self._frames = total_frames
+        self._fps = avg_fps
+
+    def __len__(self):
+        return self._frames
+
+    def get_avg_fps(self):
+        return self._fps
+
+
+@dataclass(kw_only=True)
+class Case:
+    frames: int
+    avg_fps: float
+    desired_fps: int
+    max_frames: int
+    expected_frames: list[int]
+    description: str
+
+
+# fmt: off
+@pytest.mark.parametrize("case", [
+    Case(
+        frames=100, avg_fps=25.0, desired_fps=5, max_frames=200,
+        expected_frames=[0, 5, 10, 15, 20, 26, 31, 36, 41, 46, 52, 57, 62, 67, 72, 78, 83, 88, 93, 99],
+        description="capped by desired_fps"
+    ),
+    Case(
+        frames=10, avg_fps=10.0, desired_fps=100, max_frames=5,
+        expected_frames=[0, 2, 4, 6, 9],
+        description="capped by max_frames"
+    ),
+    Case(
+        frames=50, avg_fps=25.0, desired_fps=50, max_frames=200,
+        expected_frames=list(range(50)),
+        description="capped by total_frames"
+    ),
+    Case(
+        frames=1, avg_fps=30.0, desired_fps=0, max_frames=0,
+        expected_frames=[0],
+        description="always sample at least 1 frame"
+    )
+],     ids=lambda c: c.description)
+def test_sample_video_frames_lengths(case: Case):
+    video = DummyVideo(case.frames, case.avg_fps)
+    result = sample_video_frames(video, desired_fps=case.desired_fps, max_frames=case.max_frames)
+    assert result == case.expected_frames
+
+if __name__ == "__main__":
+    pytest.main([__file__])